deepseek-v3
0 of 21 operations have a usable best known on NVIDIA B200 · 21 without one
Best known on NVIDIA B200
FP8 MLA attention output projectionsuite of 16 casesRIAC as wellnot usable387.8µsReported · license unknown · no source2026-08-21
No entry in this cohort passes the usability policy (no public source, no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 37 more entries in this cohort.
MLA fused QKV RoPE splitsuite of 16 casesRIAC as wellnot usable153.6µsReported · license unknown · no source2026-08-19
No entry in this cohort passes the usability policy (no public source, no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 23 more entries in this cohort.
MLA paged decode h16 ckv512 kpe64 ps1bf16 · [1, 16, 64] · num kv indices 8flashinfer / wrapper03f7b0not usable15.1µsReported · Apache-2.0 · source2025-10-21stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 3 more entries in this cohort.
MLA paged prefill causal h16 ckv512 kpe64 ps1bf16 · [3, 16, 64]flashinfer / wrapperea3787not usable16.3µsReported · Apache-2.0 · source2025-10-21stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 4 more entries in this cohort.
Fused add RMSNorm h7168bf16 · [7168] · batch 18gemini-2.5-pro / triton05pwmxnot usable8.19µsReported · Apache-2.0 · source2025-10-16stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 6 more entries in this cohort.
RMSNorm h1536bf16 · [1536] · batch 32gemini-2.5-pro / cuda208a66not usable6.26µsReproduction-ready · Apache-2.0 · source2025-10-16stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 7 more entries in this cohort.
RMSNorm h512bf16 · [512] · batch 32gemini-2.5-pro / tritonnicff0not usable6.18µsReported · Apache-2.0 · source2025-10-16stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 6 more entries in this cohort.
RMSNorm h7168bf16 · [7168] · batch 32gpt-5 / cuda8fff8anot usable8.05µsReproduction-ready · Apache-2.0 · source2025-10-16stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 6 more entries in this cohort.
GEMM n256 k7168fp16 · [16, 7168]torch / matmul67278enot usable15.7µsReported · Apache-2.0 · source2025-10-20stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 6 more entries in this cohort.
FP8 shared expert MLPsuite of 16 casesAmir M. Mir | SF Tensornot usable164.6µsReported · license unknown · no source2026-08-01
No entry in this cohort passes the usability policy (no public source, no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 22 more entries in this cohort.
MoE expert computationsuite of 16 casesDatabricksnot usable3.86msReported · license unknown · no source2026-07-10
No entry in this cohort passes the usability policy (no public source, no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 13 more entries in this cohort.
MoE FP8 block scale ds routing topk8 ng8 kg4 e32 h7168 i2048bf16/fp32/fp8_e4m3 · [256] · seq len 1gpt-5-2025-08-07 / cudaa2d8canot usable1.51msReproduction-ready · Apache-2.0 · source2025-10-16stale
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 2 more entries in this cohort.
FP8 multi latent attention QKV projectionsuite of 16 casesLasernot usable351.4µsReported · license unknown · no source2026-08-23
No entry in this cohort passes the usability policy (no public source, no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 23 more entries in this cohort.
MLA ragged prefill causal h16 qk192 vo128bf16 · [86, 16, 192]flashinfer / wrapperd90b98not usable10.2µsReported · Apache-2.0 · source2026-04-06
No entry in this cohort passes the usability policy (no install recipe); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 1 more entry in this cohort.
Coverage gaps on NVIDIA B200
No eligible evidence yet for these operations. Challenges →
Related tags: deepseek-v3-1-terminus · deepseek-v3-2