AMD Radeon R9700
A consumer- and workstation-grade discrete graphics card from AMD that supports the ROCm computing ecosystem and can be used for local large model inference.
Core Facts
Timeline (last 90 days)
The quantized Qwen3 MoE model can run on an R9700 with 32GB VRAM via the Straighta deployment solution
When running the Qwen3 27B NVFP4 quantized model on a single AMD Radeon R9700, performance doubled across all metrics after optimization.
The R9700 achieved a median decoding speed of 153.1 tok/s on JSON tasks.
The R9700 prefill throughput peaks at 3619 tok/s at a depth of 16000 tokens, drops back to 3192 tok/s at a depth of 64000 tokens, a decay of approximately 12%.
In the R9700 concurrency test, a single request achieves 120 tok/s, 2 concurrent requests 215, 4 concurrent 322, and 8 concurrent 471 tok/s, showing non-linear scaling.
The R9700's decoding speed varied significantly by task: math 140.0, summarization 141.7, file_edit 138.0, chat 67.1, prose 69.2, code 120.5, and reasoning 123.9 tok/s.
The core of the build is four AMD Radeon Pro V620 graphics cards, totaling approximately $1,400, providing 128GB of combined VRAM.
All Facts (7)
The core of the build is four AMD Radeon Pro V620 graphics cards, totaling approximately $1,400, providing 128GB of combined VRAM.
65%UnverifiedThe quantized Qwen3 MoE model can run on an R9700 with 32GB VRAM via the Straighta deployment solution
50%UnverifiedThe R9700 achieved a median decoding speed of 153.1 tok/s on JSON tasks.
50%UnverifiedThe R9700 prefill throughput peaks at 3619 tok/s at a depth of 16000 tokens, drops back to 3192 tok/s at a depth of 64000 tokens, a decay of approximately 12%.
50%UnverifiedIn the R9700 concurrency test, a single request achieves 120 tok/s, 2 concurrent requests 215, 4 concurrent 322, and 8 concurrent 471 tok/s, showing non-linear scaling.
50%UnverifiedThe R9700's decoding speed varied significantly by task: math 140.0, summarization 141.7, file_edit 138.0, chat 67.1, prose 69.2, code 120.5, and reasoning 123.9 tok/s.
50%UnverifiedWhen running the Qwen3 27B NVFP4 quantized model on a single AMD Radeon R9700, performance doubled across all metrics after optimization.
50%