Benchmarks
Round-ups of published results for LLM inference, image generation and other AI workloads. Each chart uses one source, names it, and says how the figures were measured.
How are these chosen?
Each round-up uses one named source per chart and one workload with the same settings, so figures from different test setups are never mixed. We rewrite the results in our own words, say whether the source tested a reference card or a partner model, and show the read date in the table below. Read how we research for the full method.
AI workstations and PCs2 articles
Which sources do the benchmark charts use?
Every chart names its source. This table lists each source used on the site, what its figures measure and when we last read it.
| Source | What it measures | Last read | Link |
|---|---|---|---|
| AMD Radeon RX 9070 XT product page | Peak throughput: Peak theoretical | See the chart notes | amd.com (opens in a new tab) |
| arXiv 2601.09527 (Knoop and Holtmann, Private LLM Inference on Consumer Blackwell GPUs) | Aggregate serving throughput: Qwen3-8B NVFP4 | 17 November 2026 | arxiv.org (opens in a new tab) |
| Back2Gaming | token_generation_decode, images_per_minute: Qwen3.8-27B (UD-IQ4_XS) chat decode, 512-token answer; Qwen3.5-9B (Q4_K_M) chat decode, 512-token answer; Stable Diffusion XL 1.0 base, ComfyUI, 1024x1024, 30 steps, Euler, batch 1 | See the chart notes | back2gaming.com (opens in a new tab) |
| Blender Open Data | Blender Open Data median score: Blender Benchmark (Cycles) median per device | 17 November 2026 | opendata.blender.org (opens in a new tab) |
| Chimolog (chimolog.co) | Generation time: Wan2.2 I2V A14B image-to-video 81 frames | See the chart notes | chimolog.co (opens in a new tab) |
| ComputingForGeeks "Best GPU for ComfyUI & Stable Diffusion by VRAM (2026)" | time per image: ComfyUI SDXL 1024x1024, 30 steps; ComfyUI FLUX.1-dev fp8 1024x1024, 20 steps | See the chart notes | computingforgeeks.com (opens in a new tab) |
| ComputingForGeeks "RTX 5060 Ti vs 5070 Ti: Local AI Benchmarks and Which to Buy" | token generation: Ollama token generation, Llama 3.1 8B Q4_K_M; Ollama token generation, Qwen2.5 14B Q4_K_M | See the chart notes | computingforgeeks.com (opens in a new tab) |
| ComputingForGeeks, RTX Pro 4000 vs RTX Pro 5000 Blackwell local AI benchmarks | text_generation_single_user: Qwen2.5 32B Q4_K_M via Ollama | See the chart notes | computingforgeeks.com (opens in a new tab) |
| Geekbench Browser | GPU compute score: Geekbench 7 OpenCL | See the chart notes | browser.geekbench.com (opens in a new tab) |
| Hardware Corner | Text generation, Prompt processing: Qwen3.5-27B distilled 6-bit, 4K context | See the chart notes | hardware-corner.net (opens in a new tab) |
| Hardware Corner (GPU ranking for LLMs) | Text generation, token_generation, prompt_processing, token generation speed, prompt processing speed, token generation, prompt processing: Qwen3 14B Q4_K_XL; Qwen3 8B Q4_K_XL; Qwen3 8B 4-bit, token generation at 16K context and more | 17 November 2026 | hardware-corner.net (opens in a new tab) |
| Hardware Corner, First DGX Spark LLM benchmarks, 15 Oct 2025 (numbers from llama.cpp maintainer) | tg32 decode, pp2048 prefill: gpt-oss 120B MXFP4 text generation; gpt-oss 120B MXFP4 prompt processing | See the chart notes | hardware-corner.net (opens in a new tab) |
| Hardware Corner, GPT-OSS 120B: Offloading MoE Layers to CPU Boosts RTX 3090 and 5090 Performance (Allan Witt, updated 10 Nov 2025) | token generation, ~86 token prompt: gpt-oss-120b (F16 GGUF, MXFP4 weights) on RTX 5090 32GB; model does not fit in VRAM | See the chart notes | hardware-corner.net (opens in a new tab) |
| Hardware Corner: Dual RTX 5060 Ti 16GB vs RTX 3090 for local LLMs | Token generation: Qwen3 32B (dense); Qwen3 30B-A3B (MoE) | See the chart notes | hardware-corner.net (opens in a new tab) |
| HotHardware | benchmark_score: Procyon AI Text Generation, Mistral 7B score; Procyon AI Image Generation, SDXL FP16, overall score, ONNX | See the chart notes | hothardware.com (opens in a new tab) |
| InsiderLLM | Token generation speed: Llama 3.2 3B Q4_K_M, llama.cpp tg128; Gemma 3 4B Q4_K_M, llama.cpp tg128 | See the chart notes | insiderllm.com (opens in a new tab) |
| llama.cpp GitHub discussion (community-reported) | prompt_processing_pp512, token_generation_tg128, prompt processing speed, token generation speed, tg128: llama-bench, Llama 2 7B Q4_0, CUDA backend, pp512, flash attention on; llama-bench, Llama 2 7B Q4_0, CUDA backend, tg128, flash attention on; Llama 2 7B Q4_0 pp512 and more | 17 November 2026 | github.com (opens in a new tab) |
| llama.cpp GitHub discussion (community-reported) | token_generation_tg128: llama-bench, Llama 2 7B Q4_0, ROCm/HIP backend, tg128, flash attention on | 17 November 2026 | github.com (opens in a new tab) |
| llama.cpp GitHub discussion #10879 (community-reported) | Text generation, prompt_processing_pp512, token_generation_tg128: Llama 2 7B Q4_0 (3.56 GiB); llama-bench, Llama 2 7B Q4_0, Vulkan backend, pp512, flash attention on; llama-bench, Llama 2 7B Q4_0, Vulkan backend, tg128, flash attention on | 17 November 2026 | github.com (opens in a new tab) |
| llama.cpp GitHub discussion 16578 (community-reported, tester eugr, Oct-Nov 2025) | pp2048 prefill at depth 0, tg32 decode at depth 0, pp2048 prefill at depth 32768, tg32 decode at depth 32768: gpt-oss 120B MXFP4 prompt processing; gpt-oss 120B MXFP4 text generation | See the chart notes | github.com (opens in a new tab) |
| LMSYS Org, NVIDIA DGX Spark In-Depth Review (13 Oct 2025) | decode: GPT-OSS 20B MXFP4 (fits in every card tested) | See the chart notes | lmsys.org (opens in a new tab) |
| LMSYS Org, NVIDIA DGX Spark In-Depth Review (13 Oct 2025) | decode: Llama 3.1 8B FP8 text generation (total throughput) | See the chart notes | lmsys.org (opens in a new tab) |
| LocalScore | Text generation, Prompt processing: Llama 3.1 8B Instruct Q4_K_M | See the chart notes | localscore.ai (opens in a new tab) |
| LocalScore | Text generation, Prompt processing: Qwen2.5 14B Instruct Q4_K_M | See the chart notes | localscore.ai (opens in a new tab) |
| Maxim Saplin (DEV Community) | Token generation speed: Mistral 7B Q6_K (5.94 GB) CPU-only generation, prompt: Tell me about Mars; Llama 3.1 8B f16 (16.07 GB) CPU-only generation; Mistral 7B Q6_K CPU-only generation and more | See the chart notes | dev.to (opens in a new tab) |
| NanoReview | GB6 ML half precision score: Geekbench 6 ML, half precision | See the chart notes | nanoreview.net (opens in a new tab) |
| Notebookcheck | Time Spy graphics score, Blender render time: 3DMark Time Spy (graphics); Blender 3.3 Classroom CUDA | See the chart notes | notebookcheck.net (opens in a new tab) |
| Notebookcheck | Time Spy graphics score: 3DMark Time Spy (graphics) | See the chart notes | notebookcheck.net (opens in a new tab) |
| Notebookcheck | Time Spy graphics score: 3DMark Time Spy (graphics) | See the chart notes | notebookcheck.net (opens in a new tab) |
| Notebookcheck | Time Spy graphics score, Geekbench 6.7.1 GPU OpenCL score, Render time: 3DMark Time Spy (graphics); Geekbench 6.7.1 GPU OpenCL; Blender 3.3 Classroom CUDA | See the chart notes | notebookcheck.net (opens in a new tab) |
| Notebookcheck | Geekbench 6.7 multi-core, 3DMark Steel Nomad score: Geekbench 6.7 CPU multi-core; 3DMark Steel Nomad (GPU) | See the chart notes | notebookcheck.net (opens in a new tab) |
| NVIDIA Technical Blog (Accelerating AI Training with NVIDIA TF32 Tensor Cores) | Relative math throughput (FP32 = 1x): Peak math, data-centre GPU | See the chart notes | developer.nvidia.com (opens in a new tab) |
| NVIDIA Technical Blog (MLPerf Inference v5.0 entries) | Throughput: Llama 2 70B; Mixtral 8x7B | See the chart notes | developer.nvidia.com (opens in a new tab) |
| PassMark | GPU Compute average: PassMark GPU Compute (PerformanceTest) | See the chart notes | videocardbenchmark.net (opens in a new tab) |
| Phoronix | Prompt processing speed, Throughput, Transcription time, Text generation speed: llama.cpp CPU LLM inference; ONNX Runtime CPU inference; Whisper.cpp speech-to-text | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix | Throughput, Generation speed: OpenVINO CPU inference; OpenVINO GenAI CPU LLM inference; TensorFlow CPU inference | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix | Build time: General compute: LLVM compile; General compute: Linux kernel compile | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix | Throughput, Prompt processing, Text generation, Transcription time: PyTorch ResNet-50 inference; llama.cpp Llama-3.1-Tulu-3-8B Q8_0; OpenVINO Person Detection FP16 and more | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix | Throughput: OpenVINO CPU inference; TensorFlow CPU inference | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix | Throughput, Prompt processing speed: ONNX Runtime CPU inference; llama.cpp CPU LLM inference | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix | Build time: General compute: Linux kernel compile; General compute: LLVM compile | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix (Michael Larabel) | Text generation: Qwen3 8B Q8_0; gpt-oss-20b Q8_0 | 17 November 2026 | phoronix.com (opens in a new tab) |
| Phoronix (Michael Larabel), NVIDIA RTX PRO Blackwell Linux review | text_generation_tg128, prompt_processing_pp512: Llama-3.1-Tulu-3-8B Q8_0 on llama.cpp (CUDA for NVIDIA, ROCm HIP for AMD); Mistral-7B-Instruct-v0.3 Q8_0 on llama.cpp (CUDA for NVIDIA, ROCm HIP for AMD); gpt-oss-20b Q8_0 (MoE) on llama.cpp (CUDA for NVIDIA, ROCm HIP for AMD) | See the chart notes | phoronix.com (opens in a new tab) |
| Phoronix (Michael Larabel), NVIDIA RTX PRO Blackwell Linux review | text_generation_tg128: GLM-4.7-Flash IQ4_XS on llama.cpp (CUDA for NVIDIA, ROCm HIP for AMD) | See the chart notes | phoronix.com (opens in a new tab) |
| Puget Systems | token_generation_2nd_plus, time_to_first_token_geomean, token generation speed (2nd+ tokens), geomean time to first token: MLPerf Client v1.0 LLM tests, 2nd+ token generation rate (model not named on chart); MLPerf Client v1.0 LLM tests, geomean time to first token; MLPerf Client 1.0 LLM | See the chart notes | pugetsystems.com (opens in a new tab) |
| Puget Systems | token_generation, image_generation_iterations_per_second: llama.cpp b4493, Phi-3-mini-4k-instruct Q4_K_M, token generation (default 128 tokens); SD.Next (2025-01-29), SDXL Base 1.0, batch size 1 | See the chart notes | pugetsystems.com (opens in a new tab) |
| Puget Systems | Processing speed: Topaz Video 1.6.1 internal benchmark, Artemis 1X model, 1920x1080; Topaz Video 1.6.1 internal benchmark, Iris 2X model, 1920x1080; Topaz Video 1.6.1 internal benchmark, Proteus 2X model, 1920x1080 and more | See the chart notes | pugetsystems.com (opens in a new tab) |
| Puget Systems, Puget Mobile 17 vs M3 Max MacBook Pro 16 for AI workflows (28 May 2024) | token generation (tg128): Meta-Llama-3-8B-Instruct Q4_K_M (fits in 16GB VRAM); Meta-Llama-3-70B-Instruct Q4_K_M (too big for 16GB VRAM) | See the chart notes | pugetsystems.com (opens in a new tab) |
| PurePC | Image generation score, Average time per image: Procyon AI Image Generation, Stable Diffusion XL (FP16), TensorRT | 17 November 2026 | purepc.pl (opens in a new tab) |
| StorageReview | image generation time: UL Procyon AI Image Generation, Stable Diffusion XL (FP16) | See the chart notes | storagereview.com (opens in a new tab) |
| StorageReview | image generation time: UL Procyon AI Image Generation, Stable Diffusion XL (FP16); UL Procyon AI Image Generation, Stable Diffusion 1.5 (FP16) | See the chart notes | storagereview.com (opens in a new tab) |
| StorageReview | image generation time: UL Procyon AI Image Generation, Stable Diffusion XL (FP16) | See the chart notes | storagereview.com (opens in a new tab) |
| StorageReview | Text generation speed, Time per image, SPECworkstation AI and Machine Learning score: Procyon AI Text Generation: Mistral; Procyon AI Image Generation: Stable Diffusion 1.5; SPECworkstation 4.0.0 AI and Machine Learning vertical | See the chart notes | storagereview.com (opens in a new tab) |
| StorageReview | Text generation: Llama 3; Llama 2; Phi and more | See the chart notes | storagereview.com (opens in a new tab) |
| StorageReview | Overall score: Procyon AI Image Generation, Stable Diffusion XL (FP16) | See the chart notes | storagereview.com (opens in a new tab) |
| Tom's Hardware | Procyon benchmark score, tokens per second, benchmark_score, token_generation, token_generation_after_first_token_geomean, Procyon AI Vision score: UL Procyon Stable Diffusion XL; MLPerf Client 0.5, Llama 2 7B, token generation (LLM inference, DirectML); Procyon AI Image Generation, Stable Diffusion XL, benchmark score and more | See the chart notes | tomshardware.com (opens in a new tab) |
| Tom's Hardware | Procyon benchmark score, Procyon AI Vision score, Procyon score: UL Procyon Stable Diffusion 1.5; UL Procyon Stable Diffusion XL; UL Procyon AI Vision (inference) and more | See the chart notes | tomshardware.com (opens in a new tab) |
| Tom's Hardware (RTX 5060 Ti 16GB review, page 8) | token generation speed (MLPerf Client 0.5), Procyon Stable Diffusion 1.5 score, Procyon Stable Diffusion XL score: MLPerf Client 0.5 LLM text generation; UL Procyon AI Image Generation SD1.5; UL Procyon AI Image Generation SDXL | See the chart notes | tomshardware.com (opens in a new tab) |
| Ultralytics Docs | Inference time: YOLO26n COCO128 imgsz 640 | 17 November 2026 | docs.ultralytics.com (opens in a new tab) |