Skip to main content

Evidence

Evidence explorer

Historical source evidence (legacy, read-only): what each test measures, which GPUs it covers and how fresh it is. Upgrade decisions use UOS workload indices only; these records are not decision inputs. Results from different sources are never averaged together.

38 tests in Local AI inference. What the evidence labels mean

Evidence tests
TestDomain · workloadMetricGPUsFreshnessEvidenceLegacy admission
LocalScore · repo bd6ffe9 · pp16+tg1536LocalScoreLocal AI inferencepp16+tg1536prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp16+tg1536LocalScoreLocal AI inferencepp16+tg1536generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp16+tg1536LocalScoreLocal AI inferencepp16+tg1536ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp64+tg1024LocalScoreLocal AI inferencepp64+tg1024prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp64+tg1024LocalScoreLocal AI inferencepp64+tg1024generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp64+tg1024LocalScoreLocal AI inferencepp64+tg1024ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp384+tg1152LocalScoreLocal AI inferencepp384+tg1152prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp384+tg1152LocalScoreLocal AI inferencepp384+tg1152generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp384+tg1152LocalScoreLocal AI inferencepp384+tg1152ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1024+tg16LocalScoreLocal AI inferencepp1024+tg16prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1024+tg16LocalScoreLocal AI inferencepp1024+tg16generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1024+tg16LocalScoreLocal AI inferencepp1024+tg16ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1024+tg1024LocalScoreLocal AI inferencepp1024+tg1024prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1024+tg1024LocalScoreLocal AI inferencepp1024+tg1024generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1024+tg1024LocalScoreLocal AI inferencepp1024+tg1024ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1280+tg3072LocalScoreLocal AI inferencepp1280+tg3072prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1280+tg3072LocalScoreLocal AI inferencepp1280+tg3072generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp1280+tg3072LocalScoreLocal AI inferencepp1280+tg3072ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp2048+tg256LocalScoreLocal AI inferencepp2048+tg256prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp2048+tg256LocalScoreLocal AI inferencepp2048+tg256generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp2048+tg256LocalScoreLocal AI inferencepp2048+tg256ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp2048+tg768LocalScoreLocal AI inferencepp2048+tg768prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp2048+tg768LocalScoreLocal AI inferencepp2048+tg768generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp2048+tg768LocalScoreLocal AI inferencepp2048+tg768ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp4096+tg256LocalScoreLocal AI inferencepp4096+tg256ttft ms (ms)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp4096+tg256LocalScoreLocal AI inferencepp4096+tg256prompt tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
LocalScore · repo bd6ffe9 · pp4096+tg256LocalScoreLocal AI inferencepp4096+tg256generation tokens per second (tok/s)2Latest measurement Mar 19, 2025Standardized test — The same benchmark suite and version run on each GPU.Supporting only
Q6_K_S · 8K · ROCm · llama-benchllama-benchLocal AI inferenceQ6_K_S · ROCm · 8K contextgeneration tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Met legacy decision rules
Q6_K_S · 8K · ROCm · llama-benchllama-benchLocal AI inferenceQ6_K_S · ROCm · 8K contextprompt tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Met legacy decision rules
Q6_K_S · 32K · ROCm · llama-benchllama-benchLocal AI inferenceQ6_K_S · ROCm · 32K contextgeneration tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Met legacy decision rules
Q6_K_S · 32K · ROCm · llama-benchllama-benchLocal AI inferenceQ6_K_S · ROCm · 32K contextprompt tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Met legacy decision rules
Q6_K_S · 64K · ROCm · llama-benchllama-benchLocal AI inferenceQ6_K_S · ROCm · 64K contextgeneration tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Met legacy decision rules
Q6_K_S · 64K · ROCm · llama-benchllama-benchLocal AI inferenceQ6_K_S · ROCm · 64K contextprompt tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Met legacy decision rules
Qwen3.8-27B-Q6_K_S.gguf · Q6_K_S · 8K · ROCm · llama-benchllama-benchLocal AI inferenceQwen3.8-27B-Q6_K_S.gguf · Q6_K_S · ROCm · 8K contextgeneration tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Supporting only
Qwen3.8-27B-Q6_K_S.gguf · Q6_K_S · 8K · ROCm · llama-benchllama-benchLocal AI inferenceQwen3.8-27B-Q6_K_S.gguf · Q6_K_S · ROCm · 8K contextprompt tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Supporting only
Qwen3.8-27B-Q6_K_S.gguf · Q6_K_S · 8K · ROCm · llama-serverllama-serverLocal AI inferenceQwen3.8-27B-Q6_K_S.gguf · Q6_K_S · ROCm · 8K contextttft ms (ms)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Supporting only
Qwen3.8-27B-Q6_K_S.gguf · Q6_K_S · 8K · ROCm · llama-serverllama-serverLocal AI inferenceQwen3.8-27B-Q6_K_S.gguf · Q6_K_S · ROCm · 8K contextgeneration tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Supporting only
Qwen3.8-27B-Q6_K_S.gguf · Q6_K_S · 8K · ROCm · llama-serverllama-serverLocal AI inferenceQwen3.8-27B-Q6_K_S.gguf · Q6_K_S · ROCm · 8K contextprompt tokens per second (tok/s)1Latest measurement Sep 25, 2026Measured — Measured on the exact workload (model, settings, and runtime) being compared.Supporting only

Local AI and GPU compute records measured by the UpgradeOrSave Bench Agent are listed per test. For run-level detail, open the benchmark console.