DeepSeek_Simulator H800: DeepSeek-V3 decode 2585 tok/s/device (d=48 b_mla=256, single-batch overlap)
2585.0 tok/s per device (simulated decode aggregate, EP degree d=48, per-rank batch 256)
Fields
id
benchmarks/dsim-h800-deepseekv3-d48-b256-single-tpsMetric
tps_aggregate
Value
2,585
Unit
tok/s per device (simulated decode aggregate, EP degree d=48, per-rank batch 256)
Model
deepseek-ai/DeepSeek-V3
Methodology
MEASURED-KERNEL SIMULATION (shenh10/DeepSeek_Simulator, commit 7c691017). DeepSeek V3/R1 decode phase: real DeepGEMM + FlashMLA kernels timed on H800-80G silicon, then recombined into throughput for expert-parallel configs that do not fit one node. Table values are tok/s PER DEVICE aggregate decode (attention-DP, MoE-EP; 'd' = data/expert-parallel degree, 'b_mla' = per-rank batch). Two-microbatch overlapping = DeepSeek official overlap; single-batch = compute-comm overlap. NOT an end-to-end serving number - a model-driven ceiling, filed as draft. Row: d=48, tp=1, b_mla=256; single-batch overlap peak (b_mla=256 knee).
Measured by
third_party
Reproducible
yes
notes
SIMULATED ceiling - real measured kernels (DeepGEMM/FlashMLA on H800) but a recombination model for the cluster-level number, not a serving trace. Do NOT quote alongside end-to-end sglang/vLLM numbers; compare within the simulator's own rows.
Notes
SIMULATED ceiling - real measured kernels (DeepGEMM/FlashMLA on H800) but a recombination model for the cluster-level number, not a serving trace. Do NOT quote alongside end-to-end sglang/vLLM numbers; compare within the simulator's own rows.