MORE FROM AMD GPUS. PROVEN ON REAL MODELS.
Paiton finds what slows your model down, replaces the generic runtime path,
and accelerates language, image and video workloads without retraining.
MEASURED ON REAL GENERATION AND INFERENCE PATHS.
The gap is rarely the model. It is usually the runtime.
Paiton tunes the exact path your workload uses, then proves the gain with benchmarks.
Latest benchmark: Radeon AI PRO R9700
Qwen3.8 delivers 400.7 aggregate output tok/s on one 32 GB R9700 with ROCm 10 and vLLM 0.29. The 65K-preset run covered 290 requests including warmups, with thinking and prefix caching off. This is a separate workload, not a matched speedup over the earlier release.
Benchmark + demoVideo diffusion: MI355X vs B200
Paiton-Diffusers removed runtime waste from the generation path and let the hardware show up.
Read Wan2.2 BenchmarkYou do not need a new model to improve unit economics. You need the runtime, kernels, and deployment path tuned around the workload you already run.
TRANSLATE THROUGHPUT INTO REUSABLE CAPACITY.
EDITABLE PLANNING SCENARIO
Faster inference is more than a benchmark number. It can reduce the GPU time required for the same workload,
or increase the output of hardware you already own.
Start with this illustrative scenario, then replace the assumptions with the uplift measured during your Paiton benchmark.
LANGUAGE, IMAGES, VIDEO WITH SOUND, AND CUSTOM STACKS.
Whether you run LLMs, image and video diffusion, MoE, or custom architectures,
Paiton starts where money is lost: latency, throughput, memory pressure, and cost per useful output.
More tokens, less waiting
Serve more work from one GPU. The latest Qwen3.8 run reached 400.7 aggregate output tok/s across eight concurrent requests on R9700. Throughput across users is distinct from the speed of a single response.
No retraining required
Keep the model and improve the path around it with AMD-aware operators and custom .so files.
Runtime work that pays back
ComfyUI workflows for FLUX.2 klein images and MiniMax H3 video with sound on R9700, Wan2.2 video on Instinct, and vLLM for language inference.
Language Models
Llama, Qwen, Deepseek, Gemma, Mistral, and CodeLlama with vLLM support, FP8 precision, and custom kernels when the gain is worth it.
Diffusion & Video
MiniMax H3 with native audio, Wan2.2, FLUX.2 klein, Stable Diffusion, SDXL and ControlNet, with workload-specific runtime and memory tuning.
Advanced Models
MoE, MLA, multimodal systems, proprietary architectures, expert kernels, novel attention, and custom inference stacks.
MEASURE. TUNE. PROVE.
Bring the workload, the target GPU, and the performance goal.
We turn that into a focused path from measurement to production-ready gain.
Run the real workload and identify where latency, memory pressure, or cost per useful output is leaking value.
Build the AMD-specific path with custom .so files, FP8 precision, kernel fusion, and the same model weights.
Deploy the optimized path on AMD Instinct or Radeon AI PRO hardware and compare the result against the baseline.
FP8 can reduce memory pressure while preserving accuracy. Tensor parallelism spreads large models like Llama-3.1-405B across multiple AMD GPUs. Custom kernels target the bottlenecks generic runtimes leave behind.
MI355X
288GB HBM3E
MI325X / MI300X / MI300A
256GB HBM3E, 192GB HBM3, 128GB HBM3 APU
MI250X / MI250 / MI210
Datacenter AMD GPU support
MI100
First compute-optimized generation
Radeon AI PRO R9700
Qualified Qwen3.8, Ornith 1.5, FLUX.2 klein and MiniMax H3 profiles on 32GB GDDR6
RX 7900 XTX / RX 7900 XT
Supported for selected inference workloads
RX 6800 XT / RX 6900 XT
Supported with workload-specific qualification
Paiton supports both AMD Instinct datacenter accelerators and Radeon GPUs. Every model, runtime, precision, and hardware combination is qualified against a defined deployment scope before delivery.
PRACTICAL LOCAL AI, SHARED WITH THE COMMUNITY.
Better inference should reach beyond our customer projects. Alongside our open-source vLLM plugin, our free RDNA community packages help developers run language models, generate images and create video with sound on their own hardware.
Explore the code, run the supported models locally, and share what you learn. It is our way of helping more people get more from the GPUs they already own.
Plan your on-prem AIvLLM integration, 65K throughput and 200K chat presets, ComfyUI workflows and model-specific setup guides.
Explore the packages on GitHubQwen3.8 includes structured tool calls and a public experimental chat profile with prefix caching and a 220K context option. Long-context chat uses one active request and has tight VRAM headroom; release presets default to caching off. Separate packages cover FLUX.2 klein images and MiniMax H3 video with sound. Check each guide for limits and licenses. The Paiton compiler remains private.
400.7 aggregate tok/s
New 65K-preset results, public 200K/220K chat instructions and experimental prefix caching. The article also preserves the original 57% gain against matched Radiance + DFlash2 in the separate 188-request, 8K-context comparison, plus the demo recording.
Benchmark + demo16.7% less waiting
A 15-second clip with native stereo audio, saved in 5m 33s on average versus 6m 39s for matched stock. Turbo8 on one 32 GB R9700, with warm, complete-request timings.
Watch the clips16.2% less generation time
1.054 seconds per image with a warm pipeline, 33.4% lower peak Torch allocation, and a free local ComfyUI workflow. Tested at 1024 × 1024, four steps and batch one.
Read Benchmark+27% output
Paiton with DFlash delivers 44.63 output tokens per second on R9700, 27% ahead of tuned stock vLLM in the tested single-user workload.
Read Benchmark+21% output
Higher interactive, coding, and long-context throughput from the same 32GB Radeon GPU.
Read Benchmark405B model
Faster startup and inference for a frontier-scale open model.
Read Case Study$/1M tokens
MI300X plus Paiton runtime on long-context MoE economics.
Read Benchmark
Share the model, target GPU, and production goal.
Paiton turns that into a benchmark-led optimization plan.
