Skip to content
KernelIndex
Search⌘K
Sign in

deepseek-v3

best known per operation on the selected GPU
21 operations · 8 families · 1,285 eligible runs
GPUNVIDIA B200969AMD Instinct MI300X316

0 of 21 operations have a usable best known on AMD Instinct MI300X · 21 without one

Best known on AMD Instinct MI300X

Operation · workload
Best known
Record
Trust
Observed
moe1 operation
AMD DeepSeek-style MoE layerfp16 · [1, 2048, 7168]
amd-mixture-of-experts · submission 30702not usable
4.11ms
Reported · license unknown · source
2025-06-02stale

No entry in this cohort passes the usability policy (no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 117 more entries in this cohort.

source mirroredlicense unknownno install recipeCohort on the operation page →View source →Vendor the source →Run detail →
mla-attention1 operation
AMD MLA decodebf16 · [128, 1, 7168]
submission 31179not usable
1.78ms
Reported · license unknown · source
2025-06-03stale

No entry in this cohort passes the usability policy (no install recipe, license unknown); the fastest known is shown. Measured by the source's own harness. Ranked by its metric under ranking-v1. 78 more entries in this cohort.

source mirroredlicense unknownno install recipeCohort on the operation page →View source →Vendor the source →Run detail →

Coverage gaps on AMD Instinct MI300X

Operation
Family
Status
GEMM n256 k7168
gemm
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
FP8 MLA attention output projection
mla-paged-attention
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
MLA fused QKV RoPE split
mla-paged-attention
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
MLA paged decode h16 ckv512 kpe64 ps1
mla-paged-attention
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
MLA paged decode h16 ckv512 kpe64 ps64
mla-paged-attention
0 eligible runs · gap
MLA paged prefill causal h16 ckv512 kpe64 ps1
mla-paged-attention
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
MLA paged prefill causal h16 ckv512 kpe64 ps64
mla-paged-attention
0 eligible runs · gap
MLA ragged prefill causal h16 qk192 vo128
mla-ragged
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
FP8 shared expert MLP
moe
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
MoE expert computation
moe
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
MoE FP8 block scale ds routing topk8 ng8 kg4 e32 h7168 i2048
moe
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
FP8 multi latent attention QKV projection
quantization
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
Fused add RMSNorm h7168
rmsnorm
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
RMSNorm h1536
rmsnorm
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
RMSNorm h512
rmsnorm
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
RMSNorm h7168
rmsnorm
0 eligible runs on AMD Instinct MI300X · measured on NVIDIA B200
Top k sampling from probs v129280
sampling
0 eligible runs · gap
Top k top p sampling from probs v129280
sampling
0 eligible runs · gap
Top p sampling from probs v129280
sampling
0 eligible runs · gap
AMD DeepSeek-style MoE layer
moe
measured · no usable entry (no install recipe, license unknown)
AMD MLA decode
mla-attention
measured · no usable entry (no install recipe, license unknown)

No eligible evidence yet for these operations. Challenges →

Related tags: deepseek-v3-1-terminus · deepseek-v3-2

Sources: FlashInfer-Bench (2026-04-06) · Apache-2.0 · NVIDIA SOL-ExecBench (2026-08-24) · GPU MODE KernelBot (2025-08-09) · June 9 Researcher Reciprocity License v1.0How comparison worksJSON