[{"data":1,"prerenderedAt":1201},["ShallowReactive",2],{"latest-news-3":3},[4,746,1024],{"id":5,"title":6,"body":7,"categories":728,"date":735,"description":736,"extension":737,"image":31,"meta":738,"navigation":739,"originalUrl":740,"path":741,"seo":742,"slug":743,"stem":744,"__hash__":745},"blog\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt.md","AI Data Center Power Requirements: The GPU\u002FMW Illusion",{"type":8,"value":9,"toc":712},"minimark",[10,18,25,32,35,38,41,44,47,50,53,56,61,64,71,74,77,80,169,172,175,181,187,190,193,196,199,230,233,236,239,245,251,254,265,268,271,274,277,280,283,286,289,292,298,304,307,310,342,345,348,351,354,357,363,366,369,372,375,378,416,419,422,425,428,431,434,437,440,443,448,454,457,460,492,495,500,503,506,511,514,517,520,526,529,552,555,561,567,570,573,576,579,582,585,588,591,594,597,603,606,609,612,615,618,621,624,627,630,633,639,646,649,652,655,661,664,667,670,676,679,682,688,691,694,697,703,706,709],[11,12,13,17],"p",{},[14,15,16],"strong",{},"Excerpt:"," Two data-center proposals can quote the same megawatt capacity while describing fundamentally different amounts of usable AI compute. The difference is often hidden in power boundaries, PUE assumptions and excluded infrastructure.",[19,20,22],"h2",{"id":21},"the-gpu-per-megawatt-illusion-why-more-gpus-on-paper-can-mean-less-compute-in-reality",[14,23,24],{},"The GPU-per-Megawatt Illusion: Why More GPUs on Paper Can Mean Less Compute in Reality",[11,26,27],{},[28,29],"img",{"alt":30,"src":31},"AI data center proposals compared by deployable compute rather than theoretical GPU count","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fgpu-per-megawatt-illusion.webp",[11,33,34],{},"Two AI data-center proposals are placed side by side.",[11,36,37],{},"Both claim the same number of megawatts. One lists more GPUs, offers a lower cost per accelerator and appears to deliver superior density.",[11,39,40],{},"Then the assumptions are opened.",[11,42,43],{},"One proposal uses average accelerator consumption. The other evaluates the complete rack and the operating conditions the facility must actually support.",[11,45,46],{},"One applies an attractive PUE figure. The other separates total facility power from usable ICT capacity.",[11,48,49],{},"One reserves power and space for networking, storage, management, cooling support, maintenance and expansion. The other mainly counts GPUs.",[11,51,52],{},"The higher number may not represent better engineering. It may simply represent a narrower calculation boundary.",[11,54,55],{},"This is the GPU-per-megawatt illusion: the belief that the largest theoretical accelerator count inside a nominal power envelope must be the best AI data-center design.",[11,57,58],{},[14,59,60],{},"It is not.",[11,62,63],{},"The relevant metric is deployable compute per megawatt: the infrastructure the complete facility can power, cool, connect, feed with data, maintain and operate reliably at the same time.",[65,66,68],"h3",{"id":67},"how-more-gpus-appear-on-paper",[14,69,70],{},"How More GPUs Appear on Paper",[11,72,73],{},"“GPUs per megawatt” is commercially attractive because it is easy to compare. The problem is that neither side of the ratio has a universal definition.",[11,75,76],{},"A megawatt may mean utility input, total facility power, data-hall power or usable ICT power at the rack. A GPU count may mean accelerator modules, complete servers, installed racks or fully operational cluster capacity.",[11,78,79],{},"That creates several ways for a proposal to look denser without necessarily delivering more usable compute.",[81,82,83,99],"table",{},[84,85,86],"thead",{},[87,88,89,93,96],"tr",{},[90,91,92],"th",{},"Category",[90,94,95],{},"Headline calculation",[90,97,98],{},"Complete engineering model",[100,101,102,114,125,136,147,158],"tbody",{},[87,103,104,108,111],{},[105,106,107],"td",{},"Power boundary",[105,109,110],{},"Treats most facility power as available to compute",[105,112,113],{},"Separates utility, facility and usable ICT capacity",[87,115,116,119,122],{},[105,117,118],{},"GPU power basis",[105,120,121],{},"Uses average consumption or isolated GPU TDP",[105,123,124],{},"Uses complete server or rack input and a defined workload envelope",[87,126,127,130,133],{},[105,128,129],{},"PUE",[105,131,132],{},"Uses a target, annual average or best-case value",[105,134,135],{},"Uses a stated design point and checks other operating conditions",[87,137,138,141,144],{},[105,139,140],{},"Supporting ICT",[105,142,143],{},"Minimizes or excludes network, storage and management",[105,145,146],{},"Includes the systems required to operate the cluster",[87,148,149,152,155],{},[105,150,151],{},"Operating state",[105,153,154],{},"Assumes every component is available",[105,156,157],{},"Shows capacity during maintenance and the stated failure condition",[87,159,160,163,166],{},[105,161,162],{},"Reserve",[105,164,165],{},"Allocates every available kilowatt",[105,167,168],{},"Preserves justified operational and expansion headroom",[11,170,171],{},"The issue is not necessarily false arithmetic. It is inconsistent scope.",[11,173,174],{},"A proposal does not become more efficient because network switches, storage arrays, control nodes, cooling-support loads or maintenance capacity have been moved outside the calculation.",[65,176,178],{"id":177},"average-consumption-is-not-design-capacity",[14,179,180],{},"Average Consumption Is Not Design Capacity",[11,182,183],{},[28,184],{"alt":185,"src":186},"Average GPU consumption compared with design capacity for AI data center planning","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Faverage-consumption-design-capacity.webp",[11,188,189],{},"Average GPU consumption is useful for forecasting energy purchases, operating expenditure and expected utilization.",[11,191,192],{},"It is not automatically the correct basis for sizing every electrical and thermal component.",[11,194,195],{},"Large transformer-training clusters can move between workload states in near unison. Uptime Institute reports that this can create frequent step changes in power demand, sometimes every second or two. The size of those changes depends on the hardware, cluster configuration, workload and power-management settings. This does not mean that every GPU runs continuously at one theoretical peak. Nor does it mean that every facility should be oversized using one arbitrary multiplier.",[11,197,198],{},"A credible design evaluates:",[200,201,202,206,209,212,215,218,221,224,227],"ul",{},[203,204,205],"li",{},"Maximum sustained workload",[203,207,208],{},"Repeated power excursions",[203,210,211],{},"Synchronization and workload diversity",[203,213,214],{},"Manufacturer limits",[203,216,217],{},"UPS, breaker and busway overload characteristics",[203,219,220],{},"Cooling-system response",[203,222,223],{},"Power-capping or smoothing policies",[203,225,226],{},"Maintenance and redundancy states",[203,228,229],{},"The required reliability level",[11,231,232],{},"Average power answers an energy question.",[11,234,235],{},"Design capacity answers an operating question.",[11,237,238],{},"The facility must support the agreed workload without relying on unplanned throttling, repeated overload or the later removal of racks.",[65,240,242],{"id":241},"pue-is-not-electrical-headroom",[14,243,244],{},"PUE Is Not Electrical Headroom",[11,246,247],{},[28,248],{"alt":249,"src":250},"PUE and ICT capacity relationship for AI data center power planning","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fpue-electrical-headroom.webp",[11,252,253],{},"The Green Grid defines Power Usage Effectiveness as:",[255,256,262],"pre",{"className":257,"code":259,"language":260,"meta":261},[258],"language-text","PUE = total data-center energy ÷ ICT-equipment energy\n","text","",[263,264,259],"code",{"__ignoreMap":261},[11,266,267],{},"PUE describes facility overhead relative to the IT load. It is not a generic safety factor, and it does not replace electrical coordination, transient analysis, redundancy planning or a complete ICT load schedule.",[11,269,270],{},"Formally, PUE is an energy metric. Applying a design-point PUE to power can be useful during concept-stage capacity planning, but it remains an approximation. Detailed design still requires component-level electrical and mechanical load schedules.",[11,272,273],{},"A measured annual PUE is also not necessarily the same as a target PUE, a design-point PUE or performance during the most demanding ambient and load condition.",[11,275,276],{},"Consider a deliberately simplified example.",[11,278,279],{},"A 10 MW facility allocation converted using a PUE of 1.10 appears to provide 9.09 MW for ICT equipment.",[11,281,282],{},"At a design-point PUE of 1.20, the same allocation provides 8.33 MW.",[11,284,285],{},"That is roughly 760 kW of difference before any allowance is made for networking, storage, management systems or operational reserve.",[11,287,288],{},"The example does not suggest that either PUE is correct for a particular project. It demonstrates why the basis must be disclosed.",[11,290,291],{},"ElioVP’s approach is not to claim that one “worst-case PUE” solves every engineering question. It is to use conservative and transparent design assumptions, then separately validate power distribution, cooling, workload behaviour, maintenance conditions and operational headroom.",[65,293,295],{"id":294},"gpus-are-not-the-complete-ai-platform",[14,296,297],{},"GPUs Are Not the Complete AI Platform",[11,299,300],{},[28,301],{"alt":302,"src":303},"Complete AI platform infrastructure beyond GPU accelerator count","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fcomplete-ai-platform.webp",[11,305,306],{},"A GPU does not train or serve a model by itself.",[11,308,309],{},"Useful AI infrastructure also requires:",[200,311,312,315,318,321,324,327,330,333,336,339],{},[203,313,314],{},"Host CPUs and system memory",[203,316,317],{},"Scale-up interconnects",[203,319,320],{},"Scale-out Ethernet or InfiniBand",[203,322,323],{},"Fabric switches and optical transceivers",[203,325,326],{},"High-performance storage",[203,328,329],{},"Dataset-ingestion and checkpointing capacity",[203,331,332],{},"Management and control-plane nodes",[203,334,335],{},"DPUs and security services",[203,337,338],{},"Monitoring and orchestration",[203,340,341],{},"Customer and out-of-band networks",[11,343,344],{},"Liquid-cooled systems also require a complete thermal chain: rack interfaces, CDUs, pumps, facility-water systems, controls and external heat rejection.",[11,346,347],{},"Every one of those systems consumes power, occupies space, generates heat and adds cost.",[11,349,350],{},"A proposal that assigns almost the complete ICT envelope to accelerator power has not made the supporting platform disappear. It has left it outside the headline number.",[11,352,353],{},"A powered accelerator can still be commercially unproductive if the fabric is congested, storage cannot feed the workload, cooling limitations force power caps or the cluster cannot remain available during maintenance.",[11,355,356],{},"A powered GPU is not necessarily a productive GPU.",[65,358,360],{"id":359},"amd-helios-makes-gpu-only-calculations-obsolete",[14,361,362],{},"AMD Helios Makes GPU-Only Calculations Obsolete",[11,364,365],{},"AMD Helios is the clearest current example of why accelerator-only capacity planning no longer works.",[11,367,368],{},"The Helios rack-scale design integrates 72 AMD Instinct MI455X GPUs with AMD EPYC “Venice” CPUs, Pensando networking, UALink scale-up connectivity, rack-level power, liquid cooling and ROCm software. It uses the double-wide OCP Open Rack Wide format.",[11,370,371],{},"This is not a conventional rack containing 72 independent GPUs.",[11,373,374],{},"It is a coordinated rack-scale computer.",[11,376,377],{},"A credible Helios deployment must include:",[200,379,380,383,386,389,392,395,398,401,404,407,410,413],{},[203,381,382],{},"Complete rack input",[203,384,385],{},"Host CPUs and memory",[203,387,388],{},"Scale-up and scale-out fabrics",[203,390,391],{},"External network-switch capacity",[203,393,394],{},"DPUs and management infrastructure",[203,396,397],{},"Storage and checkpointing",[203,399,400],{},"Liquid-cooling distribution",[203,402,403],{},"Residual room cooling",[203,405,406],{},"Facility heat rejection",[203,408,409],{},"Physical service clearances",[203,411,412],{},"Maintenance isolation",[203,414,415],{},"Software and power-management policies",[11,417,418],{},"In July 2026, AMD and Schneider Electric released a jointly developed Helios infrastructure reference design supporting rack densities up to 246 kW and modular clusters up to 10.4 MW of IT load.",[11,420,421],{},"The 246 kW value is a supported density in that reference design. It should not be interpreted as a statement that every Helios rack will continuously consume exactly 246 kW.",[11,423,424],{},"Its importance is architectural.",[11,426,427],{},"At that density, the electrical system, liquid-cooling plant, building geometry, controls and operational model must be engineered together.",[11,429,430],{},"A bidder cannot calculate credible Helios capacity by multiplying 72 accelerators by an individual GPU power figure. That would omit the host systems, fabric, networking, cooling and supporting infrastructure that turn those accelerators into a working AI platform.",[11,432,433],{},"AMD has announced that Helios shipments to customers, including Microsoft, will begin during the second half of 2026. The infrastructure requirement is therefore immediate, not theoretical.",[11,435,436],{},"NVIDIA Vera Rubin confirms the same direction.",[11,438,439],{},"NVIDIA describes Vera Rubin as a five-rack platform operating as one AI supercomputer. The platform combines NVL72 compute with dedicated CPU, storage, networking and operational infrastructure.",[11,441,442],{},"Different ecosystems are reaching the same conclusion:",[11,444,445],{},[14,446,447],{},"The data center is becoming part of the computer.",[65,449,451],{"id":450},"why-the-realistic-proposal-may-look-more-expensive",[14,452,453],{},"Why the Realistic Proposal May Look More Expensive",[11,455,456],{},"A complete proposal may show fewer GPUs and a higher initial cost because it includes more of the real project.",[11,458,459],{},"It may include:",[200,461,462,465,468,471,474,477,480,483,486,489],{},[203,463,464],{},"Network fabrics",[203,466,467],{},"Storage systems",[203,469,470],{},"Management infrastructure",[203,472,473],{},"Complete liquid cooling",[203,475,476],{},"Heat-rejection equipment",[203,478,479],{},"Maintainable power paths",[203,481,482],{},"Monitoring and controls",[203,484,485],{},"Commissioning",[203,487,488],{},"Service space",[203,490,491],{},"Expansion capacity",[11,493,494],{},"A narrower proposal can appear cheaper because those requirements have been excluded, deferred, assigned to the customer or left for detailed design.",[11,496,497],{},[14,498,499],{},"The costs do not disappear.",[11,501,502],{},"They return later as additional switchgear, larger busways, extra network rows, storage upgrades, cooling-plant expansion, permanent power caps, rack depopulation or delayed commissioning.",[11,504,505],{},"The cheapest proposal is sometimes simply the proposal with the most costs still hidden.",[11,507,508],{},[14,509,510],{},"Headroom is not automatically waste.",[11,512,513],{},"It may be required for workload changes, maintenance, demanding environmental conditions, partial cooling availability, additional networking and storage, commissioning uncertainty or future rack generations.",[11,515,516],{},"Excessive reserve can also strand capital. The answer is not blind oversizing.",[11,518,519],{},"It is evidence-based engineering with clearly disclosed assumptions.",[65,521,523],{"id":522},"what-buyers-should-ask-before-comparing-proposals",[14,524,525],{},"What Buyers Should Ask Before Comparing Proposals",[11,527,528],{},"Before accepting a GPU-per-megawatt figure, buyers should ask:",[200,530,531,534,537,540,543,546,549],{},[203,532,533],{},"What exactly does the quoted megawatt represent?",[203,535,536],{},"Is GPU power based on component TDP, average consumption, complete server input or complete rack input?",[203,538,539],{},"Is the PUE annual, measured, targeted, seasonal or a defined design point?",[203,541,542],{},"Are networking, storage, management and cooling-support systems included in power, space and cost?",[203,544,545],{},"What capacity remains during maintenance or the stated failure condition?",[203,547,548],{},"How are sustained workloads, recurring power changes and power-management policies evaluated?",[203,550,551],{},"Can the quoted configuration actually be commissioned and operated without unplanned derating?",[11,553,554],{},"If bidders cannot answer those questions using equivalent boundaries, their GPU-density figures should not be compared.",[65,556,558],{"id":557},"eliovp-designs-for-deployable-compute",[14,559,560],{},"ElioVP Designs for Deployable Compute",[11,562,563],{},[28,564],{"alt":565,"src":566},"ElioVP deployable compute capacity model for AI infrastructure planning","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fdeployable-compute-model.webp",[11,568,569],{},"At ElioVP, capacity planning begins with the complete AI platform rather than an isolated accelerator specification.",[11,571,572],{},"That means connecting facility power, liquid cooling, rack architecture, networking, storage, management infrastructure and workload behaviour in one capacity model.",[11,574,575],{},"It also means making assumptions and exclusions visible before contract signature, not after commissioning.",[11,577,578],{},"ElioVP’s capabilities span ModFlex modular data centers, high-density hardware, storage, advanced networking and AMD workload optimization through Paiton.",[11,580,581],{},"That combination is particularly relevant to Helios.",[11,583,584],{},"Being ready for AMD Helios is not only about providing enough electrical power or installing liquid-cooling pipes. It requires an understanding of the relationship between MI455X compute, EPYC host systems, UALink, Pensando networking, storage, ROCm, workload behaviour and the physical facility.",[11,586,587],{},"ElioVP’s engineering team also includes an Uptime Institute Accredited Tier Specialist, bringing formal Tier Standards knowledge into decisions concerning resilience, maintainability and operational requirements.",[11,589,590],{},"Being ready for AMD Helios and NVIDIA Vera Rubin does not mean claiming that every final platform can be installed in any building without project-specific validation.",[11,592,593],{},"It means being prepared to engineer the complete rack- and pod-scale system around the real site, workload, cooling architecture, power topology, regulatory environment and expansion plan.",[11,595,596],{},"At ElioVP, we would rather explain why a realistic number is lower today than explain why an unrealistic number cannot be delivered tomorrow.",[65,598,600],{"id":599},"productive-compute-is-the-metric-that-matters",[14,601,602],{},"Productive Compute Is the Metric That Matters",[11,604,605],{},"AI data center power requirements cannot be reduced to the largest GPU number that fits into a spreadsheet.",[11,607,608],{},"Average power is not design capacity.",[11,610,611],{},"PUE is not a safety margin.",[11,613,614],{},"Accelerator TDP is not complete rack input.",[11,616,617],{},"Installed GPUs are not necessarily deployable GPUs.",[11,619,620],{},"Do not ask only how many GPUs fit inside a megawatt.",[11,622,623],{},"Ask how many GPUs the complete facility can power, cool, connect, feed with data, maintain and operate reliably, at the same time.",[11,625,626],{},"That number may be lower than the most aggressive headline.",[11,628,629],{},"It is also far more likely to survive detailed engineering, commissioning and real workloads.",[11,631,632],{},"ElioVP can independently review an AI data-center capacity model, normalize its assumptions and identify which requirements have, or have not, been included before a theoretical GPU count becomes a costly physical constraint.",[19,634,636],{"id":635},"frequently-asked-questions",[14,637,638],{},"Frequently Asked Questions",[640,641,643],"h4",{"id":642},"how-many-gpus-can-an-ai-data-center-support-per-megawatt",[14,644,645],{},"How Many GPUs Can an AI Data Center Support per Megawatt?",[11,647,648],{},"There is no universal number.",[11,650,651],{},"The result depends on the accelerator architecture, complete server or rack configuration, facility-power boundary, PUE, workload, networking, storage, cooling, maintenance conditions, redundancy and justified reserve.",[11,653,654],{},"For rack-scale platforms such as AMD Helios, capacity should be calculated using the complete system and its supporting infrastructure, not by dividing one megawatt by the power rating of an individual GPU.",[640,656,658],{"id":657},"should-an-ai-data-center-be-sized-using-average-or-peak-gpu-power",[14,659,660],{},"Should an AI Data Center Be Sized Using Average or Peak GPU Power?",[11,662,663],{},"Neither value is sufficient by itself.",[11,665,666],{},"Average power is useful for estimating energy use and operating cost. Engineering design must also evaluate maximum sustained demand, recurring excursions, workload synchronization, equipment limits, power-management policies and the required reliability objective.",[11,668,669],{},"The correct basis is a validated operating envelope rather than one average or one theoretical maximum.",[640,671,673],{"id":672},"does-pue-include-networking-and-storage",[14,674,675],{},"Does PUE Include Networking and Storage?",[11,677,678],{},"Networking and storage belong inside the ICT load. They are not optional facility overhead that can be ignored when calculating available GPU capacity.",[11,680,681],{},"PUE describes the relationship between total facility energy and ICT-equipment energy. It does not determine how the ICT envelope should be divided between GPUs, CPUs, networking, storage and management infrastructure.",[640,683,685],{"id":684},"why-do-gpu-capacity-estimates-differ-between-proposals",[14,686,687],{},"Why Do GPU-Capacity Estimates Differ Between Proposals?",[11,689,690],{},"They often use different definitions.",[11,692,693],{},"One bidder may start with utility power while another begins with usable ICT capacity. They may also use different PUE assumptions, GPU power bases, workload profiles, support-system allowances, maintenance states and expansion reserves.",[11,695,696],{},"The proposal with the higher GPU count may be more efficient, or it may simply have counted less of the complete platform.",[640,698,700],{"id":699},"what-does-helios-ready-mean-for-a-data-center",[14,701,702],{},"What Does “Helios-Ready” Mean for a Data Center?",[11,704,705],{},"A Helios-ready design must consider the complete rack-scale architecture: GPUs, CPUs, UALink connectivity, Pensando networking, rack power, liquid cooling, external fabric, storage, management and ROCm operation.",[11,707,708],{},"It must also accommodate the double-wide Open Rack Wide format, service access, facility-water interfaces, heat rejection and the required maintenance conditions.",[11,710,711],{},"It should not mean that a generic rack position has simply been labelled “AI-ready.”",{"title":261,"searchDepth":713,"depth":713,"links":714},2,[715,727],{"id":21,"depth":713,"text":24,"children":716},[717,719,720,721,722,723,724,725,726],{"id":67,"depth":718,"text":70},3,{"id":177,"depth":718,"text":180},{"id":241,"depth":718,"text":244},{"id":294,"depth":718,"text":297},{"id":359,"depth":718,"text":362},{"id":450,"depth":718,"text":453},{"id":522,"depth":718,"text":525},{"id":557,"depth":718,"text":560},{"id":599,"depth":718,"text":602},{"id":635,"depth":713,"text":638},[729,730,731,732,733,734],"All","AI Infrastructure","Data Centers","ModFlex","HPC","AMD Helios","2026-07-27T23:52:00","Why do AI data-center proposals quote different GPU capacities? See how PUE, peak loads, storage, networking and cooling determine deployable compute.","md",{},true,null,"\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt",{"title":6,"description":736},"ai-data-center-power-requirements-gpu-per-megawatt","blog\u002Fai-data-center-power-requirements-gpu-per-megawatt","HB9AuItl8hAHri3naGXPH9VwbXO_7rXv27WS5KdVSqk",{"id":747,"title":748,"body":749,"categories":989,"date":1014,"description":1015,"extension":737,"image":1016,"meta":1017,"navigation":739,"originalUrl":1018,"path":1019,"seo":1020,"slug":1021,"stem":1022,"__hash__":1023},"blog\u002Fblog\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x.md","Paiton Returns to Its Diffusion Roots: Optimizing Wan2.2-T2V-A14B on AMD MI355X",{"type":8,"value":750,"toc":981},[751,754,761,765,772,784,787,791,798,841,846,851,854,860,863,883,886,889,895,898,901,904,907,910,913,930,933,939,942,945,948,953,956,959,962,966,969,972,975,978],[11,752,753],{},"When we first started building Paiton, one of our earliest focus areas was optimizing diffusion models. Stable Diffusion XL was one of the first large models where we showed that fused operators, efficient execution, and hardware-aware kernels could make a real difference.",[11,755,756,757,760],{},"Now we are returning to those origins.With the growing interest in text-to-video generation, we have added support for ",[14,758,759],{},"Wan-AI\u002FWan2.2-T2V-A14B",", a large text-to-video diffusion model. This is an important step for Paiton because it shows that our compiler and runtime approach is not limited to LLM inference. Paiton is built to optimize real AI workloads across model families, including diffusion, video generation, and multimodal systems.",[65,762,764],{"id":763},"benchmark-setup","Benchmark setup",[11,766,767,768,771],{},"For this comparison, we used the ",[14,769,770],{},"Hugging Face Diffusers"," library to run the Wan2.2-T2V-A14B model on two high-end accelerators:",[200,773,774,779],{},[203,775,776],{},[14,777,778],{},"AMD MI355X",[203,780,781],{},[14,782,783],{},"NVIDIA B200",[11,785,786],{},"The goal was simple: compare generation time for the same model using the same Diffusers-based workflow. Using our Paiton-Diffusers plugin on the AMD MI355X.",[65,788,790],{"id":789},"results","Results",[11,792,793,794,797],{},"Based on the average generation times, the AMD MI355X completed generation in ",[14,795,796],{},"17.6% less time"," than the NVIDIA B200.",[81,799,800,819],{},[84,801,802],{},[87,803,804,809,814],{},[90,805,806],{},[14,807,808],{},"GPU",[90,810,811],{},[14,812,813],{},"Average generation time",[90,815,816],{},[14,817,818],{},"Result",[100,820,821,831],{},[87,822,823,825,828],{},[105,824,783],{},[105,826,827],{},"6.672s",[105,829,830],{},"Baseline",[87,832,833,835,838],{},[105,834,778],{},[105,836,837],{},"5.501s",[105,839,840],{},"17.6% faster generation time",[11,842,843],{},[28,844],{"alt":261,"src":845},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fwan22_benhmark_evidence.jpg",[11,847,848],{},[28,849],{"alt":261,"src":850},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fwan22_barchart.jpg",[11,852,853],{},"This is exactly the type of workload Paiton was built for: large models, heavy GPU execution, and expensive inference paths where every optimization matters.",[65,855,857],{"id":856},"why-diffusion-models-still-matter",[14,858,859],{},"Why diffusion models still matter",[11,861,862],{},"Large language models have received most of the attention in recent years, but diffusion models remain one of the most important classes of AI workloads. Image generation, video generation, 3D generation, and multimodal pipelines all depend on many of the same performance-critical patterns:",[200,864,865,868,871,874,877,880],{},[203,866,867],{},"large matrix operations",[203,869,870],{},"normalization layers",[203,872,873],{},"attention blocks",[203,875,876],{},"memory-heavy tensor transformations",[203,878,879],{},"repeated denoising steps",[203,881,882],{},"operator scheduling overhead",[11,884,885],{},"These are exactly the areas where Paiton can provide value.",[11,887,888],{},"Our work on Wan2.2-T2V-A14B is a continuation of what we started with SDXL: taking complex diffusion pipelines and making them run faster through compilation, kernel optimization, and hardware-aware execution.",[65,890,892],{"id":891},"we-do-not-aim-for-day-zero-support",[14,893,894],{},"We do not aim for day-zero support",[11,896,897],{},"We want to be clear about our philosophy.",[11,899,900],{},"Paiton does not try to provide day-zero support for every new model the moment it appears online. That is not how we work.",[11,902,903],{},"We are European. We like to take our time.",[11,905,906],{},"Our goal is not to be first with a fragile implementation. Our goal is to publish support when the model is stable, the runtime is tested, and the performance work has been done properly. We would rather spend more time understanding the model architecture, profiling the actual bottlenecks, and optimizing the execution path than rush out a superficial integration.",[11,908,909],{},"For us, support does not simply mean “it runs.”",[11,911,912],{},"Support means:",[200,914,915,918,921,924,927],{},[203,916,917],{},"the model runs correctly",[203,919,920],{},"the model is profiled properly",[203,922,923],{},"the expensive operations are understood",[203,925,926],{},"the implementation is optimized",[203,928,929],{},"the result is useful in production",[11,931,932],{},"That is the standard we want for Paiton.",[65,934,936],{"id":935},"what-this-means-for-paiton",[14,937,938],{},"What this means for Paiton",[11,940,941],{},"Supporting Wan2.2-T2V-A14B is another step toward making Paiton a broader AI inference optimization platform.",[11,943,944],{},"We started with diffusion models. We expanded into LLMs. Now we are bringing that experience back to video generation and modern diffusion workloads.",[11,946,947],{},"The result is a simple but important message:",[11,949,950],{},[14,951,952],{},"Paiton can help AMD GPUs compete at the highest level on real AI workloads.",[11,954,955],{},"In this benchmark, using the Diffusers library, the AMD MI355X achieved better generation time than the NVIDIA B200 on Wan2.2-T2V-A14B.",[11,957,958],{},"For customers building inference services around video generation, image generation, or large multimodal models, this matters. Faster generation means lower latency, better utilization, and lower cost per output.",[11,960,961],{},"Paiton is built for exactly that.",[65,963,965],{"id":964},"thank-you-to-amd-supermicro","Thank you to AMD & Supermicro",[11,967,968],{},"We also want to thank AMD and SuperMicro for their continued support.",[11,970,971],{},"Building an IR graph compiler and runtime system like Paiton requires close access to modern hardware, a strong software ecosystem, and technical support from people who understand the platform deeply. AMD’s support has helped us test, profile, and optimize Paiton on the latest AMD Instinct GPUs, including the MI355X.",[11,973,974],{},"For a European company building high-performance AI infrastructure, this support matters. It allows us to move faster, validate our work on real hardware, and show that AMD GPUs can compete strongly on demanding AI workloads such as large-scale video generation.",[11,976,977],{},"Paiton is independent technology, but having access to AMD’s hardware and ecosystem helps us push our optimization work further.",[11,979,980],{},"We appreciate the collaboration and look forward to continuing to build high-performance AI inference solutions on AMD GPUs.",{"title":261,"searchDepth":713,"depth":713,"links":982},[983,984,985,986,987,988],{"id":763,"depth":718,"text":764},{"id":789,"depth":718,"text":790},{"id":856,"depth":718,"text":859},{"id":891,"depth":718,"text":894},{"id":935,"depth":718,"text":938},{"id":964,"depth":718,"text":965},[729,990,991,992,993,994,995,996,997,998,999,1000,808,1001,1002,1003,1004,1005,1006,1007,991,1008,1009,1010,1011,1012,1013],"Artificial Intelligence","Paiton","14B","AMD","B200","Benchmarks","Blackwell","Compute","Diffusion","Eliovp","Generative AI","Hardware","Inference","Instinct","MI355x","NVidia","On-Premise","Optimization","Sovereign AI","T2V","Text-to-Video","Tuning","Video-Generation","Wan2.2","2026-06-10T14:04:04","When we first started building Paiton, one of our earliest focus areas was optimizing diffusion models. Stable Diffusion XL was one of the first large models where we showed that fused operators, efficient execution, and hardware-aware kernels could make a real difference. Now we are returning to those origins.With the growing interest in text-to-video generation, ...","\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","\u002Fblog\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x",{"title":748,"description":1015},"paiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","blog\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","gKa2ypyZGm5Bl81PxcNxXfaJdYNDf412P547Kv9m5XQ",{"id":1025,"title":1026,"body":1027,"categories":1189,"date":1192,"description":1193,"extension":737,"image":1065,"meta":1194,"navigation":739,"originalUrl":1195,"path":1196,"seo":1197,"slug":1198,"stem":1199,"__hash__":1200},"blog\u002Fblog\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure.md","From the Attic to the Front Page: ElioVP Recognized as a Pioneer in Chip Optimization & Data Center Infrastructure",{"type":8,"value":1028,"toc":1183},[1029,1052,1058,1061,1066,1072,1079,1085,1090,1104,1110,1116,1122,1129,1135,1141,1144,1153,1156,1162,1168,1171],[11,1030,1031,1032,1034,1035,1038,1039,1051],{},"It has been some incredible weeks for the team here at ",[14,1033,999],{},". We are extremely proud to share that our company was recently featured on the ",[14,1036,1037],{},"front page of"," ",[1040,1041,1045],"a",{"href":1042,"rel":1043},"https:\u002F\u002Fwww.tijd.be\u002F",[1044],"nofollow",[1046,1047,1048],"em",{},[14,1049,1050],{},"De Tijd",", Belgium’s leading business newspaper.",[11,1053,1054],{},[28,1055],{"alt":1056,"src":1057},"ElioVP founder holding De Tijd newspaper front page feature inside a modular data center server aisle","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fvettefotoschuin.jpg",[11,1059,1060],{},"Seeing our story, from our founder’s early days tinkering with wires in an attic to generating €215 million in revenue, printed in bold on the front page of the physical newspaper was a milestone moment. It is a testament to the hard work of our team and our relentless drive to squeeze every ounce of performance out of today’s hardware.",[11,1062,1063],{},[28,1064],{"alt":261,"src":1065},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fphysicalnewspaper.webp",[65,1067,1069],{"id":1068},"the-race-for-efficiency",[14,1070,1071],{},"The Race for Efficiency",[11,1073,1074,1075,1078],{},"The main feature, titled ",[1046,1076,1077],{},"“Vlaams bedrijf optimaliseert chips van AMD en Nvidia en draait daarmee 215 miljoen euro omzet,”"," extensively researching our origins.",[11,1080,1081,1082],{},"As the article highlights, we started by unlocking the hidden potential of hardware from giants like AMD and Nvidia. By optimizing software to bypass standard factory limitations, we provided the speed and efficiency that the crypto and AI markets were starving for. As our founder Elio Van Puyvelde told the newspaper: ",[1046,1083,1084],{},"“In the race for efficiency, my solution is worth gold.”",[11,1086,1087],{},[28,1088],{"alt":261,"src":1089},"\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fderace.jpg",[11,1091,1092,1038,1095,1038,1101],{},[1046,1093,1094],{},"(You can read the full online article here:",[1040,1096,1099],{"href":1097,"rel":1098},"https:\u002F\u002Fwww.tijd.be\u002Fondernemen\u002Ftechnologie\u002Fvlaams-bedrijf-optimaliseert-chips-van-amd-en-nvidia-en-draait-daarmee-215-miljoen-euro-omzet\u002F10635877.html",[1044],[1046,1100,1050],{},[1046,1102,1103],{},")",[65,1105,1107],{"id":1106},"building-the-infrastructure-of-the-future",[14,1108,1109],{},"Building the Infrastructure of the Future",[11,1111,1112,1113,1115],{},"While we are famous for software optimization, the coverage didn’t stop there. We are particularly proud of the follow-up analysis ",[1046,1114,1050],{}," published just a week later regarding the booming data center industry.",[11,1117,1118,1119],{},"As AI models grow larger, traditional data centers are struggling to cope with the massive heat and power density required by modern GPUs. We realized early on that to truly support the next generation of computing, we couldn’t just fix the chips; ",[14,1120,1121],{},"we had to build the environment they live in.",[11,1123,1124,1125,1128],{},"We have pivoted significantly toward designing and building ",[14,1126,1127],{},"modular data centers",". These aren’t just standard server rooms; they are high-performance, prefabricated modules designed to handle the extreme energy densities of the AI revolution.",[11,1130,1131],{},[28,1132],{"alt":1133,"src":1134},"De Tijd article featuring ElioVP as a standard-setter in modular data center construction","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fenglishversionprofitereboomdatacenter.jpg",[65,1136,1138],{"id":1137},"setting-the-standard-for-the-industry",[14,1139,1140],{},"Setting the Standard for the Industry",[11,1142,1143],{},"Perhaps the proudest moment for us was seeing ElioVP recognized not just as a participant in this market, but as a pioneer.",[11,1145,1146,1147,1038,1150,1152],{},"In the follow-up article ",[1046,1148,1149],{},"“Belgische bedrijven profiteren van boom datacenters,”",[1046,1151,1050],{}," highlighted how the Flemish data center ecosystem is rapidly expanding. We are honored that the newspaper points to ElioVP as a trailblazer in this space, noting that our early innovations helped pave the way for other local players now entering the market.",[11,1154,1155],{},"It is a privilege to see that our specialized approach to modular construction and high-density cooling has become a blueprint for the industry. We are proud to see a strong tech ecosystem growing right here in Belgium, with ElioVP at the forefront.",[65,1157,1159],{"id":1158},"looking-ahead",[14,1160,1161],{},"Looking Ahead",[11,1163,1164,1165,1167],{},"We want to extend a huge thank you to the editorial team at ",[1046,1166,1050],{}," for telling our story, more specifically Emma Verplancke, and to photographer Jonas Lampens for the fantastic shots.",[11,1169,1170],{},"Most importantly, we want to thank our partners and our team. Whether we are optimizing a single GPU or deploying a modular data center to train the next Large Language Model, ElioVP is committed to staying ahead of the curve.",[11,1172,1173,1038,1176],{},[1046,1174,1175],{},"Read the follow-up article on data centers here:",[1040,1177,1180],{"href":1178,"rel":1179},"https:\u002F\u002Fwww.tijd.be\u002Fondernemen\u002Fict\u002Fbelgische-bedrijven-profiteren-van-boom-datacenters-vroeger-maakten-we-grondstof-voor-zeep-nu-koelen-we-servers\u002F10646132.html",[1044],[1046,1181,1182],{},"De Tijd – Data Center Boom",{"title":261,"searchDepth":713,"depth":713,"links":1184},[1185,1186,1187,1188],{"id":1068,"depth":718,"text":1071},{"id":1106,"depth":718,"text":1109},{"id":1137,"depth":718,"text":1140},{"id":1158,"depth":718,"text":1161},[729,990,1190,1191,993,1050,1190,1005],"Modular DC","Uncategorized","2026-02-10T20:48:12","It has been some incredible weeks for the team here at Eliovp. We are extremely proud to share that our company was recently featured on the front page of De Tijd, Belgium’s leading business newspaper. Seeing our story, from our founder’s early days tinkering with wires in an attic to generating €215 million in revenue, ...",{},"https:\u002F\u002Feliovp.com\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure\u002F","\u002Fblog\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure",{"title":1026,"description":1193},"from-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","blog\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","NG2Quz47clvVRf-TKggcyNYWxBK7poZzTQKGB0kT8GA",1785189546860]