[{"data":1,"prerenderedAt":2108},["ShallowReactive",2],{"blog-post-en-\u002Fblog\u002Fmi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach":3,"blog-posts-sidebar-en":1549},{"id":4,"title":5,"body":6,"categories":1525,"date":1535,"description":1536,"extension":1537,"heading":1538,"image":1539,"meta":1540,"navigation":1541,"originalUrl":1542,"path":1543,"seo":1544,"slug":1545,"socialImage":1546,"stem":1547,"updated":1546,"video":1546,"__hash__":1548},"blog\u002Fblog\u002Fmi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach.md","MI300X FP8 Benchmarks: GPU Partitioning with Paiton",{"type":7,"value":8,"toc":1513},"minimark",[9,13,22,25,28,34,80,83,88,99,106,109,249,252,258,261,269,277,280,291,299,302,308,311,319,322,328,334,337,346,349,356,362,369,375,378,381,387,390,396,402,404,410,416,419,425,432,435,438,443,446,451,454,460,465,775,781,785,975,978,992,997,1000,1003,1085,1130,1135,1171,1177,1180,1292,1297,1302,1324,1330,1333,1336,1339,1346,1363,1369,1395,1401,1404,1407,1410,1413,1419,1422,1431,1434,1442,1448],[10,11,12],"p",{},"At ElioVP, we’re all about pushing AI inference past the limits, and packaging every squeeze of performance into a plug‑and‑play runtime.",[10,14,15,16,21],{},"Remember ",[17,18,20],"a",{"href":19},"\u002Fblog\u002Fpaiton-fp8-beats-nvidias-h200-on-amds-mi300x","our last blog",", where Paiton’s FP8 pipeline on AMD’s MI300X completely outclassed NVIDIA’s H200? Well, buckle up, because we’ve gone back to the drawing board.",[10,23,24],{},"This time, we’re loading Llama-3.1-8B-Instruct-FP8-KV, the leaner, meaner FP8‑quantized Llama variant, into not 8 GPUs but 64 virtual GPUs carved out of a single MI300X server.",[10,26,27],{},"Powered by vLLM and Paiton’s kernel magic, we expected modest gains in multi‑tenant scaling…what we got instead was an unexpected, yet amazing surprise and a near dead‑heat with Nvidia’s B200’s.",[10,29,30],{},[31,32,33],"strong",{},"Why did we do this?",[35,36,37,44,50,56,62,68,74],"ul",{},[38,39,40,43],"li",{},[31,41,42],{},"Maximize utilization",": Slice the silicon so every tenant only pays for, and uses, exactly the VRAM and compute they need.",[38,45,46,49],{},[31,47,48],{},"Elastic multi‑tenancy",": Spin up isolated vGPUs in seconds, eliminating noisy‑neighbor slowdowns and siloed resource contention.",[38,51,52,55],{},[31,53,54],{},"Granular SLAs",": Tailor QoS per slice, ultra‑low latency for chatbots, bulk throughput for batch jobs, without juggling hardware.",[38,57,58,61],{},[31,59,60],{},"Cost‑efficient scaling",": Right‑size your compute footprint (and your budget) by renting mini‑GPUs instead of the whole chip.",[38,63,64,67],{},[31,65,66],{},"Rapid CI\u002FCD provisioning",": Integrate GPU slices into your pipeline for instant A\u002FB tests, blue\u002Fgreen rollouts, and regression benchmarks.",[38,69,70,73],{},[31,71,72],{},"Fault isolation",": Contain OOMs and driver hiccups at the slice level, so one bad job doesn’t take down the entire server.",[38,75,76,79],{},[31,77,78],{},"Future‑proof flexibility",": Re‑slice on the fly to match new model footprints or quant formats, no forklift upgrades required.",[10,81,82],{},"With these building blocks in place, we set out to see how far Paiton could stretch inference on a partitioned MI300X, and the numbers? Let’s just say they’ll make you sit up and take notice.",[10,84,85],{},[31,86,87],{},"Goals",[35,89,90,93,96],{},[38,91,92],{},"Evaluate the inference scalability of Paiton on MI300X when using GPU partitioning.",[38,94,95],{},"Measure latency and throughput of Llama 3.1 8B in FP8 format using vLLM.",[38,97,98],{},"Validate memory efficiency and kernel fusion benefits of plug-and-play Paiton models.",[100,101,103],"h3",{"id":102},"benchmarking-testbed-methodology",[31,104,105],{},"Benchmarking Testbed & Methodology",[10,107,108],{},"Our benchmarking method follows a clear set of rules and steps. This makes sure our tests are open and reproducible.",[35,110,111,128,168,179,195,211,217],{},[38,112,113,116,117],{},[31,114,115],{},"Hardware Configuration",":\n",[35,118,119,122,125],{},[38,120,121],{},"8 x AMD MI300x",[38,123,124],{},"8 x Nvidia H200",[38,126,127],{},"8 x Nvidia B200",[38,129,130,116,133],{},[31,131,132],{},"Inference Library",[35,134,135,144,151,160],{},[38,136,137,138],{},"AMD MI300x (Paiton): ",[17,139,143],{"href":140,"rel":141},"https:\u002F\u002Fgithub.com\u002FROCm\u002Fvllm",[142],"nofollow","vLLM v0.9.0",[38,145,146,147],{},"AMD MI300x (AITER): ",[17,148,150],{"href":140,"rel":149},[142],"v0.9.2",[38,152,153,154,159],{},"NVIDIA H200: ",[17,155,158],{"href":156,"rel":157},"https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm",[142],"v0.10.0","  (V1 mode)",[38,161,162,163,167],{},"NVIDIA B200: ",[17,164,166],{"href":156,"rel":165},[142],"v0.10.1","  (V1 mode) (Had to build from source to support the B200 arch)",[38,169,170,173,174],{},[31,171,172],{},"Language Model:"," ",[17,175,178],{"href":176,"rel":177},"https:\u002F\u002Fhuggingface.co\u002Famd\u002FLlama-3.1-8B-Instruct-FP8-KV",[142],"Llama-3.1-8B-Instruct-FP8-KV",[38,180,181,116,184],{},[31,182,183],{},"Driver Stack",[35,185,186,189,192],{},[38,187,188],{},"AMD MI300x: ROCm 6.4.2",[38,190,191],{},"NVIDIA H200: CUDA 12.8.1",[38,193,194],{},"NVIDIA B200: CUDA 12.8.1",[38,196,197,200],{},[31,198,199],{},"Framework:",[35,201,202,205,208],{},[38,203,204],{},"AMD MI300x: Torch 2.7.1+rocm6.3",[38,206,207],{},"NVIDIA H200: Torch 2.7.1+cu128",[38,209,210],{},"NVIDIA B200: Torch 2.9.0.dev+cu128",[38,212,213,216],{},[31,214,215],{},"Batch Size",": 1024",[38,218,219,222,223],{},[31,220,221],{},"Measurement Protocol",": Each benchmark was run 10 times, and the numbers we report are overall averages. This helps reduce the effect of temporary system changes. Our careful measurement steps include:\n",[35,224,225,231,237,243],{},[38,226,227,230],{},[31,228,229],{},"Startup Times",": Important for checking how long it takes to load the model and get the system ready.",[38,232,233,236],{},[31,234,235],{},"Cold-Start TTFT (Time to First Token)",": Measures how long it takes from a new request until the first generated token appears. This is key for how quickly interactive applications respond.",[38,238,239,242],{},[31,240,241],{},"Steady-State TTFT",": Checks the TTFT after the system has been running steadily, showing typical performance under constant use.",[38,244,245,248],{},[31,246,247],{},"End-to-End Latency Metrics",": Gives a full picture of the time it takes for a complete inference request, from sending input to getting the final output.",[10,250,251],{},"This detailed method provides a strong way to check the specific performance details of Paiton in busy, partitioned GPU environments.",[100,253,255],{"id":254},"data-parallelism-without-partitioning",[31,256,257],{},"Data Parallelism Without Partitioning",[10,259,260],{},"Our first approach was to try to utilize vllm’s built in “–data-parallel-size” option, we quickly realized that this was not going to work out of the box and would require some serious modification. So instead, we took a different approach.",[10,262,263,264,268],{},"To run the benchmarks across 8 containers using vLLM, we first followed the official NGINX load balancing guide (",[17,265,266],{"href":266,"rel":267},"https:\u002F\u002Fdocs.vllm.ai\u002Fen\u002Fstable\u002Fdeployment\u002Fnginx.html",[142],")",[270,271,272],"ol",{},[38,273,274],{},[31,275,276],{},"NGINX Configuration",[10,278,279],{},"Here is the load balancing configuration we used in \u002Fetc\u002Fnginx\u002Fnginx.conf:",[281,282,288],"pre",{"className":283,"code":285,"language":286,"meta":287},[284],"language-text","upstream backend {\n    least_conn;\n    server vllm0:8000 max_fails=3 fail_timeout=10000s;\n    server vllm1:8000 max_fails=3 fail_timeout=10000s;\n    server vllm2:8000 max_fails=3 fail_timeout=10000s;\n    server vllm3:8000 max_fails=3 fail_timeout=10000s;\n    server vllm4:8000 max_fails=3 fail_timeout=10000s;\n    server vllm5:8000 max_fails=3 fail_timeout=10000s;\n    server vllm6:8000 max_fails=3 fail_timeout=10000s;\n    server vllm7:8000 max_fails=3 fail_timeout=10000s;\n}\n\nserver {\n    listen 80;\n    location \u002F {\n        proxy_pass http:\u002F\u002Fbackend;\n        proxy_set_header Host $host;\n        proxy_set_header X-Real-IP $remote_addr;\n        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;\n        proxy_set_header X-Forwarded-Proto $scheme;\n    }\n}\n","text","",[289,290,285],"code",{"__ignoreMap":287},[270,292,294],{"start":293},2,[38,295,296],{},[31,297,298],{},"Launching the Docker Containers",[10,300,301],{},"We used the following script to launch 8 containers using incremental device and port numbers:",[281,303,306],{"className":304,"code":305,"language":286,"meta":287},[284],"#!\u002Fbin\u002Fbash\n\necho \"Starting vLLM containers with incremental configuration…\"\n\nfor i in {0..7}; do\n    device_num=$((128 + (i * 8)))\n    device_path=\"\u002Fdev\u002Fdri\u002FrenderD${device_num}\"\n    port=$((8080 + i))\n    container_name=\"vllm${i}\"\n\n    echo \"Starting container ${container_name} on port ${port} with device ${device_path}…\"\n\n    docker run -itd \\\n        –ipc host \\\n        -v \u002Fdata:\u002Fdata \\\n        –network vllm_nginx \\\n        -e VLLM_ROCM_USE_AITER=True \\\n        -e HF_HOME=root\u002F.cache\u002Fhuggingface \\\n        -e HF_HUB_CACHE=\u002Froot\u002F.cache\u002Fhuggingface\u002Fhub \\\n        –device=\u002Fdev\u002Fkfd \\\n        –device=${device_path} \\\n        –group-add video \\\n        -p ${port}:8000 \\\n        –name ${container_name} \\\n        rocm\u002Fvllm:latest \\\n        vllm serve \\\n        amd\u002FLlama-3.1-8B-Instruct-FP8-KV \\\n        –num-scheduler-steps 10 \\\n        –kv-cache-dtype fp8 \\\n        –max-model-len 4096\n\n    if [ $? -eq 0 ]; then\n        echo \"✓ Container ${container_name} started successfully\"\n    else\n        echo \"✗ Failed to start container ${container_name}\"\n    fi\n\n    echo \"—\"\ndone\n\necho \"All containers started. Summary:\"\necho \"Containers: vllm0 through vllm7\"\necho \"Ports: 8081 through 8088\"\necho \"Devices: renderD128 through renderD184 (in steps of 8)\"\n",[289,307,305],{"__ignoreMap":287},[10,309,310],{},"This line, device_num=$((128 + (i * 8))), was necessary because of leftover render device entries in \u002Fdev\u002Fdri\u002F from previous GPU partitioning. Even after resetting the partitions, the device numbers did not reset to their original state. As a result, we had to offset each device path to correctly reference the available render nodes.",[270,312,314],{"start":313},3,[38,315,316],{},[31,317,318],{},"Benchmarking",[10,320,321],{},"Finally, we ran the following command to benchmark across all containers:",[281,323,326],{"className":324,"code":325,"language":286,"meta":287},[284],"for i in {1..10}; do\n    echo \"=== Running benchmark iteration $i\u002F10 ===\"\n    python3 ~\u002Fvllm\u002Fbenchmarks\u002Fbenchmark_serving.py \\\n      –backend vllm \\\n      –model amd\u002FLlama-3.1-8B-Instruct-FP8-KV \\\n      –dataset-name sharegpt \\\n      –dataset-path ~\u002Fvllm\u002FShareGPT_V3_unfiltered_cleaned_split.json \\\n      –num-prompts 1024 \\\n      –random-range-ratio 1.0 \\\n      –percentile-metrics ttft,tpot,itl,e2el \\\n      –sharegpt-output-len 256\n    echo \"=== Completed iteration $i\u002F10 ===\"\n    echo\ndone\n",[289,327,325],{"__ignoreMap":287},[100,329,331],{"id":330},"data-parallelism-with-partitioning",[31,332,333],{},"Data Parallelism With Partitioning",[10,335,336],{},"The most important first step was to partitionize our GPUs.",[10,338,339,340,345],{},"This was very straightforward and easy to do following ",[17,341,344],{"href":342,"rel":343},"https:\u002F\u002Finstinct.docs.amd.com\u002Fprojects\u002Famdgpu-docs\u002Fen\u002Flatest\u002Fgpu-partitioning\u002Findex.html",[142],"AMD’s official documentation",".",[10,347,348],{},"Steps:",[270,350,351],{},[38,352,353],{},[31,354,355],{},"Set the compute partitions.",[281,357,360],{"className":358,"code":359,"language":286,"meta":287},[284],"sudo amd-smi set –gpu all –compute-partition CPX\n",[289,361,359],{"__ignoreMap":287},[270,363,364],{"start":293},[38,365,366],{},[31,367,368],{},"Set the memory partitions.",[281,370,373],{"className":371,"code":372,"language":286,"meta":287},[284],"sudo amd-smi set –memory-partition NPS4\n",[289,374,372],{"__ignoreMap":287},[10,376,377],{},"Wait a few seconds and, done!",[10,379,380],{},"Result:",[10,382,383],{},[384,385],"img",{"alt":287,"src":386},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Frocmoverview.webp",[10,388,389],{},"Ready to go!",[10,391,392,393,268],{},"As mentioned in the previous section, to run the benchmarks across multiple containers using vLLM, we first followed the official NGINX load balancing guide (",[17,394,266],{"href":266,"rel":395},[142],[270,397,398],{},[38,399,400],{},[31,401,276],{},[10,403,279],{},[281,405,408],{"className":406,"code":407,"language":286,"meta":287},[284],"upstream backend {\n    least_conn;\n    server vllm0:8000 max_fails=3 fail_timeout=10000s;\n    .\n    .\n    .\n    server vllm63:8000 max_fails=3 fail_timeout=10000s;\n}\n\nserver {\n    listen 80;\n    location \u002F {\n        proxy_pass http:\u002F\u002Fbackend;\n        proxy_set_header Host $host;\n        proxy_set_header X-Real-IP $remote_addr;\n        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;\n        proxy_set_header X-Forwarded-Proto $scheme;\n    }\n}\n",[289,409,407],{"__ignoreMap":287},[270,411,412],{"start":313},[38,413,414],{},[31,415,298],{},[10,417,418],{},"We used the following script to launch 64 containers using incremental device and port numbers:",[281,420,423],{"className":421,"code":422,"language":286,"meta":287},[284],"#!\u002Fbin\u002Fbash\n\n# Script to run vLLM containers with incremental device, port, and name changes\n# Runs 64 containers with device=\u002Fdev\u002Fdri\u002FrenderD128 increasing in steps of 64\n# Port starting at 8081 and increasing by 1 each time\n# Container name starting at vllm0 and increasing by 1 each time\n\necho \"Starting vLLM containers with incremental configuration…\"\n\nfor i in {0..63}; do\n    # Calculate device number (renderD128, renderD192, renderD256, etc.)\n    device_num=$((128 + i))\n    device_path=\"\u002Fdev\u002Fdri\u002FrenderD${device_num}\"\n\n    # Calculate port (8081, 8082, 8083, etc.)\n    port=$((8081 + i))\n\n    # Calculate container name (vllm0, vllm1, vllm2, etc.)\n    container_name=\"vllm${i}\"\n\n    echo \"Starting container ${container_name} on port ${port} with device ${device_path}…\"\n    docker run -itd \\\n        –ipc host \\\n        -v \u002Fdata:\u002Fdata \\\n        –network vllm_nginx \\\n        -e VLLM_ROCM_USE_AITER=True \\\n        -e HF_HOME=root\u002F.cache\u002Fhuggingface \\\n        -e HF_HUB_CACHE=\u002Froot\u002F.cache\u002Fhuggingface\u002Fhub \\\n        –device=\u002Fdev\u002Fkfd \\\n        –device=${device_path} \\\n        –group-add video \\\n        -p ${port}:8000 \\\n        –name ${container_name} \\\n        rocm\u002Fvllm:latest \\\n        vllm serve \\\n        \u002Fdata\u002F.cache\u002Fhuggingface\u002Fhub\u002Fmodels–amd–Llama-3.1-8B-Instruct-FP8-KV\u002Fsnapshots\u002Ffa42f9a9105c545755fea25cf69f49ac8c8b40e1\u002F \\\n        –num-scheduler-steps 10 \\\n        –kv-cache-dtype fp8 \\\n        –max-model-len 4096\n\n    # Check if container started successfully\n    if [ $? -eq 0 ]; then\n        echo \"✓ Container ${container_name} started successfully\"\n    else\n        echo \"✗ Failed to start container ${container_name}\"\n    fi\n\n    echo \"—\"\ndone\n\necho \"All containers started. Summary:\"\necho \"Containers: vllm0 through vllm63\"\necho \"Ports: 8081 through 8144\"\necho \"Devices: renderD128 through renderD4160 (in steps of 64)\"\necho \"\"\necho \"To check container status: docker ps\"\necho \"To view logs: docker logs <container_name>\"\n",[289,424,422],{"__ignoreMap":287},[270,426,428],{"start":427},4,[38,429,430],{},[31,431,318],{},[10,433,434],{},"Same script as in the previous section.",[10,436,437],{},"Action view :)",[10,439,440],{},[384,441],{"alt":287,"src":442},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Flh7-rt.googleusercontent.com-dc3c57fbfc36.gif",[10,444,445],{},"Paiton MI300x",[10,447,448],{},[384,449],{"alt":287,"src":450},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Flh7-rt.googleusercontent.com-d1bd449add79.gif",[10,452,453],{},"Stock MI300x",[100,455,457],{"id":456},"benchmark-results",[31,458,459],{},"Benchmark Results",[461,462,464],"h4",{"id":463},"no-partitions-8-gpus","No Partitions – 8 GPUs",[466,467,468,515],"table",{},[469,470,471],"thead",{},[472,473,474,480,485,490,495,500,505,510],"tr",{},[475,476,477],"th",{},[31,478,479],{},"Metric",[475,481,482],{},[31,483,484],{},"Paiton ∆",[475,486,487],{},[31,488,489],{},"Stock",[475,491,492],{},[31,493,494],{},"∆ vs Stock",[475,496,497],{},[31,498,499],{},"H200",[475,501,502],{},[31,503,504],{},"∆ vs H200",[475,506,507],{},[31,508,509],{},"B200",[475,511,512],{},[31,513,514],{},"∆ vs B200",[516,517,518,551,583,615,647,679,711,743],"tbody",{},[472,519,520,526,529,532,537,540,545,548],{},[521,522,523],"td",{},[31,524,525],{},"Benchmark duration (s) ↓",[521,527,528],{},"4.812",[521,530,531],{},"11.029",[521,533,534],{},[31,535,536],{},"+129.20%",[521,538,539],{},"11.84",[521,541,542],{},[31,543,544],{},"+146.05%",[521,546,547],{},"4.59",[521,549,550],{},"-4.61%",[472,552,553,558,561,564,569,572,577,580],{},[521,554,555],{},[31,556,557],{},"Request throughput (req\u002Fs) ↑",[521,559,560],{},"213.55",[521,562,563],{},"94.308",[521,565,566],{},[31,567,568],{},"+126.44%",[521,570,571],{},"83.22",[521,573,574],{},[31,575,576],{},"+156.61%",[521,578,579],{},"225.99",[521,581,582],{},"–5.50%",[472,584,585,590,593,596,601,604,609,612],{},[521,586,587],{},[31,588,589],{},"Output token throughput (tok\u002Fs) ↑",[521,591,592],{},"53851.639",[521,594,595],{},"23809.63",[521,597,598],{},[31,599,600],{},"+126.18%",[521,602,603],{},"20940.86",[521,605,606],{},[31,607,608],{},"+157.16%",[521,610,611],{},"56989.26",[521,613,614],{},"-5.52%",[472,616,617,622,625,628,633,636,641,644],{},[521,618,619],{},[31,620,621],{},"Total Token throughput (tok\u002Fs) ↑",[521,623,624],{},"101941.667",[521,626,627],{},"45047.076",[521,629,630],{},[31,631,632],{},"+126.30%",[521,634,635],{},"39674.51",[521,637,638],{},[31,639,640],{},"+156.94%",[521,642,643],{},"107827.34",[521,645,646],{},"-5.46%",[472,648,649,654,657,660,665,668,673,676],{},[521,650,651],{},[31,652,653],{},"Mean TTFT (ms) ↓",[521,655,656],{},"543.799",[521,658,659],{},"4252.513",[521,661,662],{},[31,663,664],{},"+682.47%",[521,666,667],{},"3027.49",[521,669,670],{},[31,671,672],{},"+456.96%",[521,674,675],{},"1245.55",[521,677,678],{},"+129.05%",[472,680,681,686,689,692,697,700,705,708],{},[521,682,683],{},[31,684,685],{},"Mean TPOT (ms) ↓",[521,687,688],{},"15.075",[521,690,691],{},"16.872",[521,693,694],{},[31,695,696],{},"+11.92%",[521,698,699],{},"26.70",[521,701,702],{},[31,703,704],{},"+77.02%",[521,706,707],{},"10.27",[521,709,710],{},"-31.87%",[472,712,713,718,721,724,729,732,737,740],{},[521,714,715],{},[31,716,717],{},"Mean ITL (ms) ↓",[521,719,720],{},"15.025",[521,722,723],{},"16.509",[521,725,726],{},[31,727,728],{},"+9.88%",[521,730,731],{},"71.11",[521,733,734],{},[31,735,736],{},"+373.37%",[521,738,739],{},"32.62",[521,741,742],{},"+117.10%",[472,744,745,750,753,756,761,764,769,772],{},[521,746,747],{},[31,748,749],{},"Mean E2EL (ms) ↓",[521,751,752],{},"4317.43",[521,754,755],{},"8403.948",[521,757,758],{},[31,759,760],{},"+94.65%",[521,762,763],{},"9705.94",[521,765,766],{},[31,767,768],{},"+124.79%",[521,770,771],{},"3818.69",[521,773,774],{},"-11.51%",[10,776,777],{},[384,778],{"alt":287,"src":779,"title":780},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Faverage-throughput-2-2.webp","Chart",[461,782,784],{"id":783},"partitions-64-vgpus","Partitions – 64 vGPUs",[466,786,787,815],{},[469,788,789],{},[472,790,791,795,800,804,809,812],{},[475,792,793],{},[31,794,479],{},[475,796,797,799],{},[31,798,484],{},"*",[475,801,802],{},[31,803,489],{},[475,805,806],{},[31,807,808],{},"∆vs Stock(Ratio)",[475,810,811],{},"**H200",[475,813,814],{},"**∆vs H200",[516,816,817,837,857,877,896,916,936,955],{},[472,818,819,824,827,830,833,835],{},[521,820,821],{},[31,822,823],{},"Benchmark duration (s)",[521,825,826],{},"7.875",[521,828,829],{},"17.294",[521,831,832],{},"2.20",[521,834],{},[521,836],{},[472,838,839,844,847,850,853,855],{},[521,840,841],{},[31,842,843],{},"Request throughput (req\u002Fs)",[521,845,846],{},"130.234",[521,848,849],{},"59.727",[521,851,852],{},"2.18",[521,854],{},[521,856],{},[472,858,859,864,867,870,873,875],{},[521,860,861],{},[31,862,863],{},"Output token throughput (tok\u002Fs)",[521,865,866],{},"33339.931",[521,868,869],{},"15047.115",[521,871,872],{},"2.22",[521,874],{},[521,876],{},[472,878,879,884,887,890,892,894],{},[521,880,881],{},[31,882,883],{},"Total Token throughput (tok\u002Fs)",[521,885,886],{},"62667.914",[521,888,889],{},"28497.62",[521,891,832],{},[521,893],{},[521,895],{},[472,897,898,903,906,909,912,914],{},[521,899,900],{},[31,901,902],{},"Mean TTFT (ms)",[521,904,905],{},"1082.885",[521,907,908],{},"6255.879",[521,910,911],{},"5.78",[521,913],{},[521,915],{},[472,917,918,923,926,929,932,934],{},[521,919,920],{},[31,921,922],{},"Mean TPOT (ms)",[521,924,925],{},"20.99",[521,927,928],{},"31.289",[521,930,931],{},"1.49",[521,933],{},[521,935],{},[472,937,938,943,945,948,951,953],{},[521,939,940],{},[31,941,942],{},"Mean ITL (ms)",[521,944,925],{},[521,946,947],{},"31.13",[521,949,950],{},"1.48",[521,952],{},[521,954],{},[472,956,957,962,965,968,971,973],{},[521,958,959],{},[31,960,961],{},"Mean E2EL (ms)",[521,963,964],{},"6435.477",[521,966,967],{},"14067.724",[521,969,970],{},"2.19",[521,972],{},[521,974],{},[10,976,977],{},"**Note: We are working on improving these numbers even more.",[10,979,980,981,991],{},"**Note2: Not possible with Nvidia, or at least very difficult (",[982,983,984],"em",{},[17,985,988],{"href":986,"rel":987},"https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm\u002Fissues\u002F6551#issuecomment-2237624342",[142],[982,989,990],{},"complicated",")*",[10,993,994],{},[384,995],{"alt":287,"src":996,"title":780},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Faverage-throughput-2-2-1.webp",[10,998,999],{},"Now let’s look at this from an ROI-driven perspective.",[10,1001,1002],{},"If we haven’t impressed you so far, we’re pretty sure this will. We use the MI300X server as a reference\u002Fbaseline to compare its cost factor and throughput against the H200 and B200.",[466,1004,1005,1025],{},[469,1006,1007],{},[472,1008,1009,1014,1019,1022],{},[475,1010,1011],{},[31,1012,1013],{},"Architecture",[475,1015,1016],{},[31,1017,1018],{},"Cost Factor vs Paiton",[475,1020,1021],{},"**Throughput Cost-Eff ***",[475,1023,1024],{},"**Latency Cost-Eff",[516,1026,1027,1041,1056,1071],{},[472,1028,1029,1034,1037,1039],{},[521,1030,1031],{},[31,1032,1033],{},"MI300X+Paiton",[521,1035,1036],{},"Ref",[521,1038,1036],{},[521,1040,1036],{},[472,1042,1043,1047,1050,1053],{},[521,1044,1045],{},[31,1046,453],{},[521,1048,1049],{},"1x",[521,1051,1052],{},"+126.31%",[521,1054,1055],{},"+94.57%",[472,1057,1058,1062,1065,1068],{},[521,1059,1060],{},[31,1061,499],{},[521,1063,1064],{},"1.375x",[521,1066,1067],{},"+253.30%",[521,1069,1070],{},"+209.07%",[472,1072,1073,1077,1080,1083],{},[521,1074,1075],{},[31,1076,509],{},[521,1078,1079],{},"2x",[521,1081,1082],{},"+89.18%",[521,1084,704],{},[1086,1087,1088,1111],"blockquote",{},[10,1089,1090,173,1093,173,1098,1101,1102,173,1105,1110],{},[982,1091,1092],{},"*Throughput Cost-Efficiency: = %",[982,1094,1095],{},[31,1096,1097],{},"more",[982,1099,1100],{},"total‑token","** ",[982,1103,1104],{},"throughput",[982,1106,1107],{},[31,1108,1109],{},"per dollar"," *vs each platform.",[10,1112,1113,1114,173,1119,173,1122,1101,1127],{},"**Latency Cost-Efficiency: = %* ",[982,1115,1116],{},[31,1117,1118],{},"better",[982,1120,1121],{},"end‑to‑end",[982,1123,1124,1109],{},[31,1125,1126],{},"latency",[982,1128,1129],{},"vs each platform.",[10,1131,1132],{},[31,1133,1134],{},"What this tells you",[35,1136,1137,1147,1161],{},[38,1138,1139,1142,1143,1146],{},[31,1140,1141],{},"Paiton delivers 2.5 × the throughput per $"," over an H200 and ",[31,1144,1145],{},"+126 %"," over stock.",[38,1148,1149,1152,1153,1156,1157,1160],{},[31,1150,1151],{},"Latency per $"," is ",[31,1154,1155],{},"3.1 × better than the H200"," and ",[31,1158,1159],{},"+94 % over stock",", solid ROI on every millisecond shaved.",[38,1162,1163,1164,1166,1167,1170],{},"The ",[31,1165,509],{}," gap is real, but remember it costs ",[31,1168,1169],{},"twice"," as much, Paiton still wins on cost‑efficiency across the board.",[100,1172,1174],{"id":1173},"cost-per-million-tokens",[31,1175,1176],{},"Cost per Million tokens",[10,1178,1179],{},"If we use available renting prices for the different systems, we could calculate the relative cost per 1M tokens:",[466,1181,1182,1215],{},[469,1183,1184],{},[472,1185,1186,1190,1195,1200,1205,1210],{},[475,1187,1188],{},[31,1189,1013],{},[475,1191,1192],{},[31,1193,1194],{},"Throughput (tok\u002Fs)",[475,1196,1197],{},[31,1198,1199],{},"GPU Count",[475,1201,1202],{},[31,1203,1204],{},"Approx. hourly cost",[475,1206,1207],{},[31,1208,1209],{},"Inference Cost \u002F 1M Tokens",[475,1211,1212],{},[31,1213,1214],{},"Relative Cost",[516,1216,1217,1237,1255,1274],{},[472,1218,1219,1223,1225,1228,1231,1234],{},[521,1220,1221],{},[31,1222,1033],{},[521,1224,624],{},[521,1226,1227],{},"~8",[521,1229,1230],{},"$20.50",[521,1232,1233],{},"$0.06",[521,1235,1236],{},"REF",[472,1238,1239,1243,1245,1247,1249,1252],{},[521,1240,1241],{},[31,1242,453],{},[521,1244,627],{},[521,1246,1227],{},[521,1248,1230],{},[521,1250,1251],{},"$0.13",[521,1253,1254],{},"2.26× ↑",[472,1256,1257,1261,1263,1265,1268,1271],{},[521,1258,1259],{},[31,1260,499],{},[521,1262,635],{},[521,1264,1227],{},[521,1266,1267],{},"$28.20",[521,1269,1270],{},"$0.20",[521,1272,1273],{},"3.54× ↑",[472,1275,1276,1280,1282,1284,1287,1289],{},[521,1277,1278],{},[31,1279,509],{},[521,1281,643],{},[521,1283,1227],{},[521,1285,1286],{},"$48.60",[521,1288,1251],{},[521,1290,1291],{},"2.24× ↑",[10,1293,1294],{},[384,1295],{"alt":287,"src":1296},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fimage-2-1.jpg",[10,1298,1299],{},[31,1300,1301],{},"Insights:",[35,1303,1304,1310,1315],{},[38,1305,1306,1309],{},[31,1307,1308],{},"Paiton"," delivers 2.26× cost savings compared to unoptimized MI300X.",[38,1311,1312,1314],{},[31,1313,499],{}," costs 3.54× more than MI300X+Paiton per 1M tokens.",[38,1316,1317,1319,1320,1323],{},[31,1318,509],{}," is the most expensive, costing over ",[31,1321,1322],{},"2.24× more"," than the optimized MI300X setup.",[100,1325,1327],{"id":1326},"big-win-for-amd",[31,1328,1329],{},"Big win for AMD",[10,1331,1332],{},"Trying to figure out how MIG worked on Nvidia with vLLM was like trying to find a perfect gift for your spouse, it was exhausting. Eventually we ran into a NCCL error which seemed unsolvable and that was the last straw.",[10,1334,1335],{},"While MIG allows virtual partitioning on supported NVIDIA GPUs, as previously mentioned we encountered significant limitations when attempting to use it in conjunction with vLLM for data-parallel workloads. Specifically, vLLM was unable to properly leverage MIG slices for distributed inference.",[10,1337,1338],{},"In contrast, AMD’s architecture enabled straightforward partitioning and containerized deployment of vLLM instances without any issues. This streamlined setup, along with ROCm’s compatibility, made AMD far better suited for true multi-tenancy out of the box.",[10,1340,1341,1342,1345],{},"This represents a ",[31,1343,1344],{},"major win for AMD",", particularly for enterprises aiming to deploy isolated inference workloads across shared hardware without too much friction or compromise.",[1086,1347,1348],{},[1086,1349,1350],{},[1086,1351,1352,1357,1360],{},[10,1353,1354],{},[31,1355,1356],{},"Having methodically outpaced Intel in performance, AMD is now strategically poised to challenge NVIDIA’s leadership, an evolution we’re proud to drive.",[10,1358,1359],{},"***Kian Mohadjerin",[10,1361,1362],{},"Head of AI, Eliovp BV*",[100,1364,1366],{"id":1365},"key-results",[31,1367,1368],{},"Key Results",[35,1370,1371,1377,1383,1389],{},[38,1372,1373,1376],{},[31,1374,1375],{},"Throughput scaling"," was near-linear up to 64 partitions, thanks to Paiton’s minimized memory overhead and fast kernel dispatch.",[38,1378,1379,1382],{},[31,1380,1381],{},"Latency"," remained stable across parallel sessions, demonstrating the strength of Paiton’s per-GPU scheduling and shared memory optimizations.",[38,1384,1385,1388],{},[31,1386,1387],{},"Memory usage per partition"," was significantly lower compared to standard vLLM or runtimes, enabling high-density deployment.",[38,1390,1391,1394],{},[31,1392,1393],{},"Cost per Million Tokens"," was reduced by over 2× compared to high-end systems like the B200, showcasing Paiton’s ability to deliver industry-leading efficiency even on more affordable AMD hardware.",[100,1396,1398],{"id":1397},"conclusion",[31,1399,1400],{},"Conclusion",[10,1402,1403],{},"This experiment highlights Paiton’s ability to unlock the full potential of modern hardware like the MI300X through advanced packaging and optimization techniques. Running Llama 3.1 8B FP8 across 64 GPU partitions showcases how inference workloads can be massively parallelized without sacrificing too much performance or usability.",[10,1405,1406],{},"Imagine the potential of Paiton paired with AMD’s upcoming MI355X. With even more memory bandwidth, compute, and architectural improvements on the horizon, the synergy between next-gen hardware and the Paiton runtime could redefine the state of high-performance AI serving.",[10,1408,1409],{},"Stay tuned for future updates as we expand Paiton’s capabilities.",[10,1411,1412],{},"Don’t believe our results? Neither did we, so test Paiton for yourself and request an evaluation model.",[100,1414,1416],{"id":1415},"pricing",[31,1417,1418],{},"Pricing",[10,1420,1421],{},"If you’re curious about pricing with Paiton, our formula is quite simple:",[466,1423,1424],{},[469,1425,1426],{},[472,1427,1428],{},[475,1429,1430],{},"50% of x% costs saved per 1M tokens",[10,1432,1433],{},"The cost saved is measured by looking at the customer’s current throughput compared to the throughput using Paiton.",[10,1435,1436,1441],{},[17,1437,1440],{"href":1438,"rel":1439},"https:\u002F\u002Fai.eliovp.com\u002Fpaiton",[142],"Reach out and let’s talk"," :)",[100,1443,1445],{"id":1444},"references",[31,1446,1447],{},"References",[270,1449,1450,1457,1464,1469,1475,1481,1487,1493,1499,1506],{},[38,1451,1452],{},[17,1453,1456],{"href":1454,"rel":1455},"https:\u002F\u002Fwww.supermicro.com\u002Fen\u002Fproducts\u002Fsystem\u002Fgpu\u002F8u\u002Fas%20-8125gs-tnmr2",[142],"Supermicro GPU System AS-8125GS-TNMR2",[38,1458,1459],{},[17,1460,1463],{"href":1461,"rel":1462},"https:\u002F\u002Fwww.amd.com\u002Fen\u002Fproducts\u002Faccelerators\u002Finstinct\u002Fmi300\u002Fmi300x.html",[142],"AMD Instinct MI300X",[38,1465,1466],{},[17,1467,1468],{"href":19},"Paiton FP8 beats NVIDIA’s H200 on AMD’s MI300X",[38,1470,1471],{},[17,1472,1474],{"href":140,"rel":1473},[142],"ROCm\u002Fvllm GitHub",[38,1476,1477],{},[17,1478,1480],{"href":156,"rel":1479},[142],"vllm-project\u002Fvllm GitHub",[38,1482,1483],{},[17,1484,1486],{"href":176,"rel":1485},[142],"Hugging Face AMD Llama-3.1-8B-Instruct-FP8-KV",[38,1488,1489],{},[17,1490,1492],{"href":266,"rel":1491},[142],"vLLM Nginx Deployment",[38,1494,1495],{},[17,1496,1498],{"href":342,"rel":1497},[142],"AMD GPU Partitioning Documentation",[38,1500,1501],{},[17,1502,1505],{"href":1503,"rel":1504},"https:\u002F\u002Frocm.blogs.amd.com\u002Fsoftware-tools-optimization\u002Fcompute-memory-modes\u002FREADME.html",[142],"ROCm Compute Memory Modes",[38,1507,1508],{},[17,1509,1512],{"href":1510,"rel":1511},"https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm\u002Fissues\u002F6551",[142],"vLLM GitHub Issue #6551",{"title":287,"searchDepth":293,"depth":293,"links":1514},[1515,1516,1517,1518,1519,1520,1521,1522,1523,1524],{"id":102,"depth":313,"text":105},{"id":254,"depth":313,"text":257},{"id":330,"depth":313,"text":333},{"id":456,"depth":313,"text":459},{"id":1173,"depth":313,"text":1176},{"id":1326,"depth":313,"text":1329},{"id":1365,"depth":313,"text":1368},{"id":1397,"depth":313,"text":1400},{"id":1415,"depth":313,"text":1418},{"id":1444,"depth":313,"text":1447},[1526,1527,1308,1528,1529,1530,509,499,1531,1532,1533,1308,1534],"All","Artificial Intelligence","Uncategorized","AI","AMD","MI300X","MI355x","NVidia","vLLM","2025-07-31T13:32:57","Explore Llama 3.1 8B FP8 benchmarks on partitioned MI300X GPUs with Paiton, comparing throughput and latency against NVIDIA H200 and B200.","md","MI300X FP8 Data‑Parallel Benchmarks (8–64 GPUs): H200 Left Behind, B200 Within Reach","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fimage-2-1.webp",{},true,"https:\u002F\u002Feliovp.com\u002Fmi300x-fp8-data%e2%80%91parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach\u002F","\u002Fblog\u002Fmi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach",{"title":5,"description":1536},"mi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach",null,"blog\u002Fmi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach","SfhmP4Z5VqUzSad6jNACgjuJAgGqUZyE60GvesI9rBo",[1550,1562,1575,1585,1594,1609,1618,1627,1636,1646,1658,1668,1680,1689,1702,1711,1722,1750,1760,1782,1800,1819,1837,1854,1871,1882,1898,1913,1915,1924,1932,1947,1958,1969,1980,1990,2003,2013,2026,2037,2047,2058,2067,2079,2090,2099],{"path":1551,"title":1552,"description":1553,"date":1554,"slug":1555,"image":1556,"originalUrl":1557,"categories":1558},"\u002Fblog\u002Fpaiton-idol-radeon-ai-pro-r9700","The Radeon AI PRO R9700 does 3D too: Idol in 16.4 seconds","Idol turns an image into a 3D Gaussian splat in 16.44 seconds on one Radeon AI PRO R9700 with Paiton, versus 39.22 seconds for stock on the same GPU.","2026-11-06T09:00:00Z","paiton-idol-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-idol\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-idol-radeon-ai-pro-r9700",[1308,1559,1560,1561],"AMD Radeon","Local AI","3D Generation",{"path":1563,"title":1564,"description":1565,"date":1566,"slug":1567,"image":1568,"originalUrl":1569,"categories":1570},"\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p","Deepseek V4.1 Flash on a single MI350P","DeepSeek V4.1 Flash on one 144 GB MI350P: 77.9 tokens\u002Fs for one request and 1,107.6 in a 512-request batch, with Paiton and CPU expert offload.","2026-11-03T12:00:00Z","paiton-deepseek-v41-flash-amd-instinct-mi350p","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p",[1308,1571,1572,1573,1574],"AMD Instinct","AI Inference","DeepSeek","Quantization",{"path":1576,"title":1577,"description":1578,"date":1579,"slug":1580,"image":1581,"originalUrl":1582,"categories":1583},"\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p","Qwen3.8 Flash Next on MI350P: twice the speed | Paiton","Qwen3.8 Flash Next on one AMD MI350P: 2.24× the tested vLLM setup's single-request decode speed and about 2.2 million shared KV-cache tokens with Paiton.","2026-10-30T12:00:00Z","paiton-qwen38-flash-next-amd-instinct-mi350p","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p",[1308,1571,1572,1584,1574],"Qwen",{"path":1586,"title":1587,"description":1588,"date":1589,"slug":1590,"image":1591,"originalUrl":1592,"categories":1593},"\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp","Qwen3.8 Flash Next: 3-bit near 16-bit accuracy, with MTP prefix caching","Paiton's new Qwen3.8 3-bit weights approach 16-bit task scores on two R9700 cards. Native MTP prefix caching resumes a 200K prompt in 0.68 seconds. Results, limits and setup.","2026-10-24T18:45:00Z","paiton-qwen38-3bit-accuracy-prefix-caching-mtp","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-3bit-mtp\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp",[1308,1559,1560,1584,1574],{"path":1595,"title":1596,"description":1597,"date":1598,"slug":1599,"image":1600,"originalUrl":1601,"categories":1602},"\u002Fblog\u002Fnvidia-dsx-pod-vera-rubin-modular-data-centers","Nvidia confirms the modular DC direction","NVIDIA DSX POD brings Vera Rubin AI factories to modular data centers. ElioVP explains site readiness, cooling and the thinking behind ModFlex.","2026-10-21T13:00:00Z","nvidia-dsx-pod-vera-rubin-modular-data-centers","\u002Fasset\u002Fimages\u002Fblog\u002Fnvidia-dsx-pod-modflex\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fnvidia-dsx-pod-vera-rubin-modular-data-centers",[1603,1604,1605,1606,1607,1608],"Modular DC","AI Infrastructure","ModFlex","NVIDIA","Vera Rubin","Liquid Cooling",{"path":1610,"title":1611,"description":1612,"date":1613,"slug":1614,"image":1615,"originalUrl":1616,"categories":1617},"\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","Qwen3.8 Flash Next in 4-bit on 2 × Radeon R9700: tested against 16-bit","Paiton runs Qwen3.8 Flash Next in 4-bit on two R9700 cards: 96.8% GSM8K, 96.3% HumanEval, 189.8 tokens\u002Fs and an optional RAM prefix cache. Accuracy, speed and setup.","2026-10-12T15:00:00Z","paiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700",[1308,1559,1560,1584,1574],{"path":1619,"title":1620,"description":1621,"date":1622,"slug":1623,"image":1624,"originalUrl":1625,"categories":1626},"\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700","Qwen3.8 Flash Next on 2 × Radeon AI PRO R9700: 216 tokens\u002Fs, 200K context","Paiton runs Qwen3.8 Flash Next with 3-bit experts on two R9700 cards: 216.3 tokens\u002Fs, 565.1 aggregate tokens\u002Fs and a separate 200K mode. Benchmarks, quality and setup.","2026-10-10T10:30:00Z","paiton-qwen38-flash-next-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700",[1308,1559,1560,1584,1574],{"path":1628,"title":1629,"description":1630,"date":1631,"slug":1632,"image":1633,"originalUrl":1634,"categories":1635},"\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700","Qwen3.8 27B on 1 × Radeon AI PRO R9700: 3-bit weights, 20% faster decode","Qwen3.8 27B on one R9700: 19.9% faster weighted decode, a new 4-bit cache, 200K context and optional vision. Benchmarks and current Paiton setup.","2026-09-27T09:00:00Z","paiton-qwen38-w3a4-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-w3a4\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700",[1308,1559,1560,1584,1574],{"path":1637,"title":1638,"description":1639,"date":1640,"slug":1641,"image":1642,"originalUrl":1643,"categories":1644},"\u002Fblog\u002Fpaiton-qwen-image-21-radeon-ai-pro-r9700","Qwen-Image 2.1 on 1 × Radeon AI PRO R9700: 2048×2048 images in 103 seconds","Generate 2048×2048 Qwen-Image 2.1 images locally with 1 × Radeon AI PRO R9700. The released Paiton v1.0.2 container measured 103.29 seconds per warm request, through PNG delivery.","2026-09-23T09:00:00Z","paiton-qwen-image-21-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen-image-21\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen-image-21-radeon-ai-pro-r9700",[1308,1559,1560,1645,1584],"Image Generation",{"path":1647,"title":1648,"description":1649,"date":1650,"slug":1651,"image":1652,"originalUrl":1653,"categories":1654},"\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700","Qwen3.8: 400.7 tok\u002Fs on R9700 | Paiton","Qwen3.8 on one R9700: 400.7 aggregate tok\u002Fs with ROCm 10 and vLLM 0.29, plus public 200K\u002F220K chat profiles. Benchmarks, limits and launch commands.","2026-09-16T07:30:00Z","paiton-qwen38-mxfp4-dflash2-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-mxfp4\u002Fupdate-2026-09-19\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700",[1308,1559,1534,1655,1656,1657],"Qwen3.8","Inference Optimization","DFlash2",{"path":1659,"title":1660,"description":1661,"date":1662,"slug":1663,"image":1664,"originalUrl":1665,"categories":1666},"\u002Fblog\u002Fpaiton-qwen38-neo-gguf-vllm-r9700","Qwen3.8 GGUF in vLLM: Faster Responses on One Radeon","Run the original NEO CODER MAX GGUF in vLLM with Paiton on an R9700. Explore measured latency gains, image input and local deployment.","2026-09-14T07:30:00Z","paiton-qwen38-neo-gguf-vllm-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-neo-gguf\u002F00-hero-neo-gguf-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-neo-gguf-vllm-r9700",[1308,1559,1560,1667,1534],"GGUF",{"path":1669,"title":1670,"description":1671,"date":1672,"slug":1673,"image":1674,"originalUrl":1675,"categories":1676},"\u002Fblog\u002Fpaiton-minimax-h3-radeon-ai-pro-r9700","MiniMax H3 on Radeon: 15-Second Video With Native Sound","Paiton generates a 15-second MiniMax H3 video with stereo audio on one Radeon AI PRO R9700 in 5m 33s, with 16.7% lower latency than matched stock.","2026-09-09T07:30:00Z","paiton-minimax-h3-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-minimax-h3\u002F00-featured-minimax-h3-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-minimax-h3-radeon-ai-pro-r9700",[1308,1559,1560,1677,1678,1679],"Video Generation","MiniMax H3","ComfyUI",{"path":1681,"title":1682,"description":1683,"date":1684,"slug":1685,"image":1686,"originalUrl":1681,"categories":1687},"\u002Fblog\u002Fpaiton-flux2-klein-radeon-ai-pro-r9700","Local FLUX.2 klein on Radeon AI PRO R9700: Faster Image Generation with Less VRAM","Paiton generates 1024×1024 FLUX.2 klein images in 1.054 seconds on an R9700: 16.2% lower warm latency and 33.4% lower peak Torch allocation.","2026-09-07T09:00:00","paiton-flux2-klein-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-flux2-klein\u002Ffox-paiton.webp",[1308,1559,1560,1645,1688,1679],"FLUX",{"path":1690,"title":1691,"description":1692,"date":1693,"slug":1694,"image":1695,"originalUrl":1696,"categories":1697},"\u002Fblog\u002Fpaiton-ornith15-radeon-ai-pro-r9700","Ornith 1.5 at 44.6 tok\u002Fs on One Radeon AI PRO R9700","Paiton serves Ornith 1.5 35B A3B at 44.63 output tok\u002Fs on one Radeon AI PRO R9700, 27% faster than tuned stock vLLM, with 21.3% lower modeled cost.","2026-09-05T09:00:00","paiton-ornith15-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-ornith15\u002F00-featured-ornith15-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-ornith15-radeon-ai-pro-r9700",[1308,1527,1559,1572,1698,1699,1656,1700,1534,1701],"GPU Performance","Inference Latency","Large Language Models","Cost Efficiency",{"path":1703,"title":1704,"description":1705,"date":1706,"slug":1707,"image":1708,"originalUrl":1709,"categories":1710},"\u002Fblog\u002Fpaiton-qwen38-radeon-ai-pro-r9700","Paiton: 21% More Qwen3.8 Throughput on Radeon AI PRO R9700","Paiton serves AMD’s Qwen3.8 27B at 39.77 output tokens\u002Fs on one Radeon AI PRO R9700, delivering 21% more throughput and 17.4% lower modeled cost.","2026-09-04T09:00:00","paiton-qwen38-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-r9700\u002F00-featured-paiton-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-radeon-ai-pro-r9700",[1308,1527,1559,1572,1698,1699,1656,1700,1534,1701],{"path":1712,"title":1713,"description":1714,"date":1715,"slug":1716,"image":1717,"originalUrl":1546,"categories":1718},"\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt","AI Data Center Power Requirements: The GPU\u002FMW Illusion","Why do AI data-center proposals quote different GPU capacities? See how PUE, peak loads, storage, networking and cooling determine deployable compute.","2026-07-27T23:52:00","ai-data-center-power-requirements-gpu-per-megawatt","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fgpu-per-megawatt-illusion.webp",[1526,1604,1719,1605,1720,1721],"Data Centers","HPC","AMD Helios",{"path":1723,"title":1724,"description":1725,"date":1726,"slug":1727,"image":1728,"originalUrl":1729,"categories":1730},"\u002Fblog\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","Wan2.2 Video Generation: Paiton on AMD MI355X","Compare Wan2.2-T2V-A14B video generation on AMD MI355X with Paiton and NVIDIA B200 using Diffusers, and explore our diffusion optimization approach.","2026-06-10T14:04:04","paiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonwan2.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x\u002F",[1526,1527,1308,1731,1530,509,1732,1733,1734,1735,1736,1737,1738,1739,1740,1741,1532,1533,1742,1743,1308,1744,1745,1746,1747,1748,1749],"14B","Benchmarks","Blackwell","Compute","Diffusion","Eliovp","Generative AI","GPU","Hardware","Inference","Instinct","On-Premise","Optimization","Sovereign AI","T2V","Text-to-Video","Tuning","Video-Generation","Wan2.2",{"path":1751,"title":1752,"description":1753,"date":1754,"slug":1755,"image":1756,"originalUrl":1757,"categories":1758},"\u002Fblog\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","ElioVP in De Tijd: Chip Optimization and Data Centers","Read about De Tijd's coverage of ElioVP, from its origins in chip optimization to its work on modular data centers and high-density cooling.","2026-02-10T20:48:12","from-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fphysicalnewspaper.webp","https:\u002F\u002Feliovp.com\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure\u002F",[1526,1527,1603,1528,1530,1759,1603,1533],"De Tijd",{"path":1761,"title":1762,"description":1763,"date":1764,"slug":1765,"image":1766,"originalUrl":1767,"categories":1768},"\u002Fblog\u002Fprivacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit","AI Privacy: A Strategic Priority for Benelux Businesses","Explore generative AI privacy risks, trust, data retention and governance, and why Benelux businesses need a strategic approach to secure AI.","2026-01-29T13:51:11","privacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fheaderimage.webp","https:\u002F\u002Feliovp.com\u002Fprivacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit\u002F",[1526,1527,1769,1528,1770,1771,1772,1773,1774,1775,1776,1777,1736,1778,1737,1779,1780,1781],"Trending","AI Act","Anthropomorphism","AVG","Benelux","ChatGPT","Cybersecurity","Data Governance","Data Security","GDPR","Microsoft Copilot","Privacy","Shadow AI",{"path":1783,"title":1784,"description":1785,"date":1786,"slug":1787,"image":1788,"originalUrl":1789,"categories":1790},"\u002Fblog\u002Fitsme-bij-ons-is-het-its-not-me-en-dit-is-waarom","Why We Do Not Use itsme: Privacy and Data Sovereignty","Why ElioVP does not use itsme: our assessment of identity metadata, cloud dependence, data sovereignty and authentication risks.","2025-11-27T09:32:14","itsme-bij-ons-is-het-its-not-me-en-dit-is-waarom","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffrontimage.webp","https:\u002F\u002Feliovp.com\u002Fitsme-bij-ons-is-het-its-not-me-en-dit-is-waarom\u002F",[1526,1791,1792,1793,1775,1794,1795,1796,1778,1797,1798,1799,1780],"AWS","Belgian Mobile ID","Cloud Act","Data Sovereignty","Digital Identity","eIDAS","itsme","Liberty Global","MyGov.be",{"path":1801,"title":1802,"description":1803,"date":1804,"slug":1805,"image":1806,"originalUrl":1807,"categories":1808},"\u002Fblog\u002Ffield-report-the-reality-of-building-agentic-ai-in-2025","Field Report. The Reality of Building Agentic AI in 2025","Lessons from building on-premise AI agents in 2025 cover workflow design, observability, model training, hallucinations and GPU memory limits.","2025-11-25T14:03:39","field-report-the-reality-of-building-agentic-ai-in-2025","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffieldreport.webp","https:\u002F\u002Feliovp.com\u002Ffield-report-the-reality-of-building-agentic-ai-in-2025\u002F",[1526,1527,1809,1769,1810,1811,1812,1813,1814,1815,1816,1817,1744,1818],"Solutions","Agentic AI","AI Engineering","AI Strategy","Autonomous Agents","Enterprise AI","Local LLM","Model Fine-Tuning","On-Premise AI","VRAM Optimization",{"path":1820,"title":1821,"description":1822,"date":1823,"slug":1824,"image":1825,"originalUrl":1826,"categories":1827},"\u002Fblog\u002Fthe-synthetic-unicorn-bubble","The Synthetic Unicorn Bubble","An analysis of AI neocloud investment risks, examining circular financing, infrastructure claims, contract terms and due diligence.","2025-11-24T19:22:28","the-synthetic-unicorn-bubble","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fsyntheticunicorn.webp","https:\u002F\u002Feliovp.com\u002Fthe-synthetic-unicorn-bubble\u002F",[1526,1527,1769,1604,1828,1829,1830,1831,1832,1833,1834,1835,1836],"AI Neocloud","Circular Financing","GPU Cloud","Investment Risks","Startup Valuation","Synthetic Bubble","Tech Analysis","Vaporware","Venture Capital",{"path":1838,"title":1839,"description":1840,"date":1841,"slug":1842,"image":1843,"originalUrl":1844,"categories":1845},"\u002Fblog\u002Fbuilding-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power","NVIDIA GB300 NVL72: A Four-Month Modular Data Center Plan","Explore a modular data center design for NVIDIA GB300 NVL72, covering redundant power, hybrid cooling and a four-month deployment plan.","2025-11-20T14:10:19","building-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fsuperpodmodflexfrontimage.webp","https:\u002F\u002Feliovp.com\u002Fbuilding-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power\u002F",[1526,1603,1528,1846,1604,1847,1848,1608,1849,1850,1851,1852,1853],"150kW Rack","DLC","High Density","Modular Data Center","NVIDIA Blackwell Ultra","NVIDIA GB300","NVL72","Rapid Deployment",{"path":1855,"title":1856,"description":1857,"date":1858,"slug":1859,"image":1860,"originalUrl":1861,"categories":1862},"\u002Fblog\u002Fwhy-cuda-translation-wont-unlock-amds-real-potential","Why “CUDA” Translation Won’t Unlock AMD’s Real Potential","Why CUDA compatibility is not the same as AMD performance: explore ROCm, HIP, kernel tuning and the case for hardware-specific optimization.","2025-11-12T14:48:37","why-cuda-translation-wont-unlock-amds-real-potential","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fchatgpt-image-nov-11-2025-09_16_10-pm-1.webp","https:\u002F\u002Feliovp.com\u002Fwhy-cuda-translation-wont-unlock-amds-real-potential\u002F",[1526,1527,1308,1528,1863,1527,1864,1865,1866,1867,1868,1869,1308,1870],"AMD MI300X","CUDA Translation","FP8","GPU Optimization","High Performance Computing","HIP","Kernel Tuning","ROCm",{"path":1872,"title":1873,"description":1874,"date":1875,"slug":1876,"image":1877,"originalUrl":1878,"categories":1879},"\u002Fblog\u002Fpaiton-the-simplest-way-to-supercharge-ai-inference","Paiton: The Simplest Way to Supercharge AI Inference","Learn how Paiton integrates with existing inference stacks, with AMD MI300X benchmark results and performance-per-dollar comparisons.","2025-11-11T10:31:22","paiton-the-simplest-way-to-supercharge-ai-inference","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaiton-powaaah.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-the-simplest-way-to-supercharge-ai-inference\u002F",[1526,1527,1308,1572,1571,1863,1701,1880,1656,1869,1308,1881,1534],"High Throughput","SGLang",{"path":1883,"title":1884,"description":1885,"date":1886,"slug":1887,"image":1888,"originalUrl":1889,"categories":1890},"\u002Fblog\u002Fstop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens","Paiton MoE Benchmarks: MI300X vs H200 and B200","Compare Qwen3-30B-A3B MoE inference with Paiton on MI300X against H200 and B200, including throughput and cost per million tokens.","2025-09-26T13:36:18","stop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fhulkvshulkpaitonwins.webp","https:\u002F\u002Feliovp.com\u002Fstop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens\u002F",[1526,1527,1308,1891,1863,1892,1656,1893,1894,1895,1896,1308,1897],"AI Benchmarks","Cost per Token","Mixture of Experts","MoE","NVIDIA B200","NVIDIA H200","Qwen3",{"path":1899,"title":1900,"description":1901,"date":1902,"slug":1903,"image":1904,"originalUrl":1905,"categories":1906},"\u002Fblog\u002Fagentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en","Local Agentic AI: From Inbox to Action","Local-first AI agents turn email, documents and images into tickets, reports and actions, using models tailored to your data and systems.","2025-09-16T13:09:00","agentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffrontfotoblog.webp","https:\u002F\u002Feliovp.com\u002Fagentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en\u002F",[1526,1527,1809,1528,1810,1907,1908,1909,1910,1815,1817,1744,1911,1912],"Damage Detection","Document Processing","Email Automation","Invoice Extraction","Ticket Automation","Workflow Automation",{"path":1543,"title":5,"description":1536,"date":1535,"slug":1545,"image":1539,"originalUrl":1542,"categories":1914},[1526,1527,1308,1528,1529,1530,509,499,1531,1532,1533,1308,1534],{"path":1916,"title":1917,"description":1918,"date":1919,"slug":1920,"image":1921,"originalUrl":1922,"categories":1923},"\u002Fblog\u002Fapplicable-ai-for-businesses","Applicable AI for Businesses","Explore ElioVP's approach to local AI for business workflows, including custom model training and automated damage detection for logistics.","2025-07-09T21:35:30","applicable-ai-for-businesses","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fscherm_afbeelding-2025-07-09-om-23.30.17.webp","https:\u002F\u002Feliovp.com\u002Fapplicable-ai-for-businesses\u002F",[1526,1527,1809,1810,1907,1908,1909,1910,1815,1817,1744,1911,1912],{"path":1925,"title":1926,"description":1927,"date":1928,"slug":1929,"image":287,"originalUrl":1930,"categories":1931},"\u002Fblog\u002Fintroducing-paitons-free-evaluation-models","Introducing Paiton’s Free Evaluation Models","Test Paiton with free evaluation models for AMD GPUs. Compare text, vision and image generation performance using your own workloads.","2025-07-07T11:26:13","introducing-paitons-free-evaluation-models","https:\u002F\u002Feliovp.com\u002Fintroducing-paitons-free-evaluation-models\u002F",[1526,1527,1308],{"path":1933,"title":1934,"description":1935,"date":1936,"slug":1937,"image":1938,"originalUrl":1939,"categories":1940},"\u002Fblog\u002Fpaiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b","Llama 3.1 405B: Faster Startup with Paiton on MI300X","See Paiton benchmarks for Llama 3.1 405B on eight AMD MI300X GPUs, covering model startup, tensor parallelism, throughput and latency.","2025-06-12T20:15:23","paiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fservingscreenshot.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b\u002F",[1526,1527,1308,1528,1572,1863,1941,1865,1942,1943,1944,1308,1945,1946],"Cold Start","Graph Compilation","Llama 3.1 405B","LLM Optimization","Startup Latency","Tensor Parallelism",{"path":19,"title":1948,"description":1949,"date":1950,"slug":1951,"image":1952,"originalUrl":1953,"categories":1954},"Paiton FP8 Beats NVIDIA’s H200 on AMD’s MI300X","Compare Paiton on AMD MI300X with NVIDIA H200 for Llama 3.1 70B FP8, including throughput, first-token delay and latency across batch sizes.","2025-06-08T19:12:40","paiton-fp8-beats-nvidias-h200-on-amds-mi300x","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fblognewfp8.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-fp8-beats-nvidias-h200-on-amds-mi300x\u002F",[1526,1527,1308,1528,1863,1955,1814,1737,1698,1699,1700,1943,1956,1957],"Cold Start Optimization","Model Serving","vLLM Optimization",{"path":1959,"title":1960,"description":1961,"date":1962,"slug":1963,"image":1964,"originalUrl":1965,"categories":1966},"\u002Fblog\u002Fmi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm","MI300X vs H200 vs RX 7900 XTX vs Tenstorrent n300s with vLLM","Compare MI300X, H200, RX 7900 XTX and Tenstorrent n300s on Llama 3 8B with vLLM, including throughput, modeled token costs and hardware limits.","2025-05-09T14:03:58","mi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fcomparisontenstor.webp","https:\u002F\u002Feliovp.com\u002Fmi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm\u002F",[1526,1527,1308,1809,1528,1530,1531,1533,1967,1968],"RX7900XTX","tenstorrent",{"path":1970,"title":1971,"description":1972,"date":1973,"slug":1974,"image":1975,"originalUrl":1976,"categories":1977},"\u002Fblog\u002Fclusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights","ClusterP&L: Financial Modeling for GPU Clusters","Explore how ClusterP&L models GPU cluster costs, profitability and investment scenarios, with ROI metrics, risk simulations and exportable reports.","2025-05-03T10:52:22","clusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fcomparisonscenarios.webp","https:\u002F\u002Feliovp.com\u002Fclusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights\u002F",[1526,1527,1603,1809,509,499,1978,1533,1979],"MI325x","pnl calculator",{"path":1981,"title":1982,"description":1983,"date":1984,"slug":1985,"image":1986,"originalUrl":1987,"categories":1988},"\u002Fblog\u002Fcranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200","AMD MI300X vs. NVIDIA H200: Qwen3-32B with Paiton","Compare Qwen3-32B benchmarks on Paiton-optimized AMD MI300X and NVIDIA H200, covering throughput, latency and hardware costs.","2025-05-02T21:10:30","cranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002F3ac59a73-2466-4422-b7e5-ef2e4a8ca58e.webp","https:\u002F\u002Feliovp.com\u002Fcranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200\u002F",[1526,1527,1308,1529,1530,499,1989,1533,1308,1534],"MI300",{"path":1991,"title":1992,"description":1993,"date":1994,"slug":1995,"image":1996,"originalUrl":1997,"categories":1998},"\u002Fblog\u002Fpower-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw","Modular Data Centers for NVIDIA NVL: 1 to 2 MW","Explore modular data center designs for NVIDIA NVL systems, covering power capacity, liquid cooling, redundancy and deployment planning.","2025-05-02T14:09:59","power-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Feliovp_critical-1mw-pod_rev-2_transparent.webp","https:\u002F\u002Feliovp.com\u002Fpower-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw\u002F",[1526,1603,1999,1604,1848,1720,1608,1849,2000,2001,1852,2002],"1-2MW Data Center","NVIDIA Blackwell","NVIDIA NVL","Precision Cooling",{"path":2004,"title":2005,"description":2006,"date":2007,"slug":2008,"image":2009,"originalUrl":2010,"categories":2011},"\u002Fblog\u002Fexamining-ai-agents-in-the-medical-field-ai-that-speaks-dicom","Examining AI agents in the medical field: AI that speaks DICOM","Explore a local AI agent demo for DICOM workflows, from patient and study retrieval to a comparison of vision models using anonymized medical images.","2025-04-11T14:45:48","examining-ai-agents-in-the-medical-field-ai-that-speaks-dicom","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fhealthcareblog-1.webp","https:\u002F\u002Feliovp.com\u002Fexamining-ai-agents-in-the-medical-field-ai-that-speaks-dicom\u002F",[1526,1527,1809,1528,1529,1530,2012],"Healthcare",{"path":2014,"title":2015,"description":2016,"date":2017,"slug":2018,"image":2019,"originalUrl":2020,"categories":2021},"\u002Fblog\u002Feliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs","U.S. Tariffs and AI Supply Chain Resilience: April 2025","Read ElioVP's April 2025 perspective on U.S. tariffs and supply chain resilience for AI servers, HPC systems and modular data centers.","2025-04-04T10:01:27","eliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ftariffsshipping.webp","https:\u002F\u002Feliovp.com\u002Feliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs\u002F",[1526,1769,1529,1530,2022,2023,2024,2025],"import","Taiwan","Tariffs","Trump",{"path":2027,"title":2028,"description":2029,"date":2030,"slug":2031,"image":2032,"originalUrl":2033,"categories":2034},"\u002Fblog\u002Fwhy-ai-agents-are-the-future","Why AI Agents Are the Future","Explore AI agents for ERP, CRM, finance and customer support, with practical use cases and a path from workflow assessment to pilot and deployment.","2025-03-23T22:06:59","why-ai-agents-are-the-future","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ferp2.jpeg","https:\u002F\u002Feliovp.com\u002Fwhy-ai-agents-are-the-future\u002F",[1526,1527,1809,1529,2035,2036],"AI Agents","ERP",{"path":2038,"title":2039,"description":2040,"date":2041,"slug":2042,"image":2043,"originalUrl":2044,"categories":2045},"\u002Fblog\u002Fthe-rise-of-open-source-ai-model-optimization","The Rise of Open-Source AI Model Optimization","Explore open-source AI optimization trends, from quantization and mixture-of-experts models to hardware-aware tuning, RAG and edge deployment.","2025-03-22T20:59:23","the-rise-of-open-source-ai-model-optimization","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Friseofopensource.jpeg","https:\u002F\u002Feliovp.com\u002Fthe-rise-of-open-source-ai-model-optimization\u002F",[1526,1527,1769,2046,1530,1738,1533],"AI news",{"path":2048,"title":2049,"description":2050,"date":2051,"slug":2052,"image":2053,"originalUrl":2054,"categories":2055},"\u002Fblog\u002Fintroducing-our-benchmarking-tool-powered-by-dstack","Introducing Our Benchmarking Tool: Powered by dstack","Explore our dstack-powered tool for reproducible vLLM benchmarks, automated parameter sweeps and performance reports across local and cloud GPUs.","2025-03-20T14:21:59","introducing-our-benchmarking-tool-powered-by-dstack","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fbenchmarktool.jpeg","https:\u002F\u002Feliovp.com\u002Fintroducing-our-benchmarking-tool-powered-by-dstack\u002F",[1526,1527,1308,1529,1530,2056,2057,1531,1308],"benchmark","LLM",{"path":2059,"title":2060,"description":2061,"date":2062,"slug":2063,"image":2064,"originalUrl":2065,"categories":2066},"\u002Fblog\u002Foptimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200","Optimizing QwQ-32B (by Qwen): AMD MI300X vs. NVIDIA H200","Compare QwQ-32B throughput and latency on AMD MI300X with Paiton and NVIDIA H200, from small batches to higher concurrency.","2025-03-19T21:41:44","optimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaiton4.jpeg","https:\u002F\u002Feliovp.com\u002Foptimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200\u002F",[1526,1527,1308],{"path":2068,"title":2069,"description":2070,"date":2071,"slug":2072,"image":2073,"originalUrl":2074,"categories":2075},"\u002Fblog\u002Feliovp-featured-on-amd-tech-talk-podcast","Eliovp Featured on AMD “Tech Talk” Podcast","Listen to Elio Van Puyvelde and Jim Greene on AMD's Tech Talk podcast, discussing ElioVP's origins and its AI hardware and software services.","2025-03-19T07:53:39","eliovp-featured-on-amd-tech-talk-podcast","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ftechtalkjimgreene.jpeg","https:\u002F\u002Feliovp.com\u002Feliovp-featured-on-amd-tech-talk-podcast\u002F",[1526,1530,2076,2077,2078],"Jim Greene","Podcast","Tech Talk",{"path":2080,"title":2081,"description":2082,"date":2083,"slug":2084,"image":2085,"originalUrl":2086,"categories":2087},"\u002Fblog\u002Ffurther-optimizing-amd-powered-inference-with-paiton","Further Optimizing AMD-Powered Inference with Paiton","Explore Paiton's DeepSeek R1 Distill Llama 8B benchmarks on AMD MI300X, focusing on throughput and first-token latency at smaller batch sizes.","2025-03-13T06:18:30","further-optimizing-amd-powered-inference-with-paiton","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost3.webp","https:\u002F\u002Feliovp.com\u002Ffurther-optimizing-amd-powered-inference-with-paiton\u002F",[1526,1527,1308,1530,2088,2089,499,1531,1978,1308,1534],"Deepseek","H100",{"path":2091,"title":2092,"description":2093,"date":2094,"slug":2095,"image":2096,"originalUrl":2097,"categories":2098},"\u002Fblog\u002Fa-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b","Paiton Benchmarks: DeepSeek R1 Distill Llama 3.1 8B","Compare stock and Paiton-optimized DeepSeek R1 Distill Llama 3.1 8B on AMD MI300X, with throughput and latency benchmarks across batch sizes.","2025-01-31T09:11:02","a-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost2.webp","https:\u002F\u002Feliovp.com\u002Fa-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b\u002F",[1526,1527,1308,1530,2088,2089,499,1531,1978,1308,1534],{"path":2100,"title":2101,"description":2102,"date":2103,"slug":2104,"image":2105,"originalUrl":2106,"categories":2107},"\u002Fblog\u002Fai-model-optimization-with-paiton","AI Model Optimization with Paiton","Learn how Paiton uses model compilation, custom kernels and kernel fusion to optimize AI inference on AMD GPUs.","2025-01-30T19:53:25","ai-model-optimization-with-paiton","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost1.webp","https:\u002F\u002Feliovp.com\u002Fai-model-optimization-with-paiton\u002F",[1526,1527,1308,1530,2089,499,1531,1978,1308,1534],1793962342818]