W7800 LLM benchmarks
259 results · 579 measurements · 9 sources
Device specifications
| Variant | Memory | Memory type | Bus | Bandwidth | FP16 matrix | INT8 matrix |
|---|---|---|---|---|---|---|
| W7800 48GB ↗ | 48 GB | GDDR6 | 384-bit | 864 GB/s | 90.4 TFLOPS | 90.4 TOPS |
| W7800 ↗ | 32 GB | GDDR6 | 256-bit | 576 GB/s | 90.4 TFLOPS | 90.4 TOPS |
Manufacturer specifications. Shared system RAM is not dedicated VRAM. Unreported values remain —; measured LLM results retain their reported configuration.
Selected reported tests
24 of 259 resultsSource-reported results, selected across reports by documentation quality. Match configurations before comparing values.
| View ReportGeneric LLM | Qwen3.8-27BConflicting evidence | HumanEval100 problems | 98 %Reported pass@1 · n = 100 | — | 42.1 tok/sGeneration · median · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
|---|---|---|---|---|---|---|
| View ReportCustom evaluation | Qwen3.8-Flash-Next-GSQ-RCO | GSM8K200 questions; seed42 | 95 %Accuracy · 190 / 200 · Reported correctness; exact extraction/scorer not published 50.68 tok/sThroughput · aggregate | — | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | Muse-Glimmer-30BConflicting evidence | HumanEval100 problems | 97 %Reported pass@1 · n = 100 | — | 27.51 tok/sGeneration · mean · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | Devstral-2-123B-Instruct-2512Conflicting evidence | HumanEval100 problems | 88 %Reported pass@1 · n = 100 | — | 7.05 tok/sGeneration · mean · per request | 2 × W7800Layer split · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | WikiText128K configured context | 9.7297 perplexityWord perplexity · Perplexity · Lower is better | — | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | MLPerf Client 1.0 LLM suite (selected models unreported) | MLPerf Client1.0 | — | — | 87.5 tok/sGeneration · unknown | 1 × W7800Single GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27BConflicting evidence | Token speed | — | — | ~24 tok/sGeneration · range lower · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View Reportllama-bench | Qwen3.8-27B | Token speed | — | — | 22.99 tok/sGeneration · per request | 1 × W7800Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | Token speedABC synthetic wall-rate | 31 tok/sThroughput · per request | — | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27BConflicting evidence | HumanEval100 problems | 99 %Reported pass@1 · n = 100 | — | 25.5 tok/sGeneration · median · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-Flash-Next-GSQ-RCO | Token speed | — | 1,043.8 tok/sPrefill · per request | 57.9 tok/sGeneration · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | Qwen3-Coder-30B-A3B-InstructConflicting evidence | HumanEval100 problems | 99 %Reported pass@1 · n = 100 | — | 96.97 tok/sGeneration · mean · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | Devstral-Small-2-24B-Instruct-2512Conflicting evidence | HumanEval100 problems | 88 %Reported pass@1 · n = 100 | — | 31.57 tok/sGeneration · mean · per request | 2 × W7800Layer split · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | GSM8KSize not reported128K configured context | 83.7 %Accuracy · Flexible extraction · Mean | — | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | MLPerf Client 1.0 LLM suite (selected models unreported) | MLPerf Client1.0 | 631 msTime to first token · Unknown | — | — | 1 × W7800Single GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27BConflicting evidence | Token speed | — | — | ~25 tok/sGeneration · range upper · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | Token speed | — | — | 22.4 tok/sGeneration · per request | 1 × W7800Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | Token speedABC synthetic wall-rate | 34.4 tok/sThroughput · per request | — | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-Flash-Next-GSQ-RCO | Token speed | — | 1,085.2 tok/sPrefill · per request | 55.2 tok/sGeneration · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | Muse-Glimmer-30BConflicting evidence | Token speedsynth-bulk-dispatcher-impl | — | — | 26.88 tok/sGeneration · mean · per request | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | gemma-4-26B-A4B-itConflicting evidence | HumanEval100 problems | 96 %Reported pass@1 · n = 100 | — | 71.18 tok/sGeneration · mean · per request | 2 × W7800Layer split · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | GSM8KSize not reported128K configured context | 84.12 %Accuracy · Strict extraction · Mean | — | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View Reportllama-bench | Qwen3.8-27BConflicting evidence | Token speed | — | 504 tok/sPrefill · per request | — | 1 × W7800Single GPU · 48 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | Qwen3.8-27B | Token speed | — | — | ~2.8 tok/sGeneration · range lower · unknown ~2.9 tok/sGeneration · range upper · unknown | 1 × W7800Single GPU · 32 GB VRAM/GPUConfiguration
|
Models and software in this collection
Model families: Devstral 2, Devstral Small 2, GLM 4.5, GLM 4.7, GPT-OSS, Gemma 4, Mistral Small 4, Muse Glimmer, Qwen3, Qwen3.8
Engines: MLPerf Client, Ollama, SGLang, SGLang gfx1100-support fork, Strata, llama-bench, llama.cpp, llama.cpp / llama-server, vLLM
Methodology and evidence standards · Download the research dataset