GPT-OSS 120B

Benchmark

The following data comes from ~/lzc-aipod-benchmark/gpt-oss-120b-moe/cumulative-report.md, corresponding to model gpt-oss-120b-mxfp4, runtime llama.cpp, and weight format GGUF.

Device	Runtime	Weight	Single Stream	Peak Aggregate	Power	GPU Temp
Thor	llama.cpp	GGUF mxfp4	34.615 tok/s (g512 / c1)	178.536 tok/s (g64 / c16)	31.793 W	67 C
Thor T5000	llama.cpp	GGUF mxfp4	35.088 tok/s (g1024 / c1)	330.602 tok/s (g64 / c48)	31.261 W for latest c1; 54.419 W at c48	56 C for latest c1; 57.593 C at c48
Jetson AGX Orin	-	-	no data	no data	-	-
Thor T4000	-	-	no data	no data	-	-

Notes

• THOR The current best single-stream result comes from the generation length matrix of g128 / g256 / g512 / g1024, where g512 / c1 is 34.615 tok/s.
• T5000 The current default single-path approach is fixed to g1024 / c1 / server_parallel=1 / 300s, with the latest retesting result being 35.088 tok/s.
• T5000's high-concurrency peak comes from c40 / c48 extended scanning, where g64 / c48 reaches 330.602 tok/s.
• In the current benchmark repository, Orin and T4000 lack the official gpt-oss-120b results, so only no data is retained.
• Data source: /home/catdog/lzc-aipod-benchmark/gpt-oss-120b-moe/cumulative-report.md