Challenges
what the index has no good answer for yet
187 open challenges
Priority gaps
high-demand families with zero runs on a priority GPUOperation / family
GPU
What is missing
Since / asks
Model gaps
measured on one priority GPU but not the otherOperation / family
GPU
What is missing
Since / asks
Attention softmax dropout value matmul backward
H100
seed-seed-oss-36b-instruct · measured on B200 only
—
Fused vision cu seqlens attention with 2d RoPE backward
H100
qwen3-vl-4b-instruct · measured on B200 only
—
Unbeaten baselines
the source's own reference; nobody has entered against itOperation / family
GPU
What is missing
Since / asks
Sparse GEMM NVFP4 n2048 k2048
B200
int8 · [2048, 1024] · m=128 · cuda_sparse_nvfp4_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Sparse GEMM NVFP4 n2048 k2048
B200
int8 · [2048, 1024] · m=64 · cuda_sparse_nvfp4_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM NVFP4 m contiguous g4 n2048 k2048
B200
fp32 · [512, 1024] · cuda_nvfp4_grouped_naive_g4_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM MXFP4 out NVFP4 n2048 k2048
B200
bf16 · [256, 2048] · mm_fp4_out_nvfp4_flashinfer_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM MXFP4 out NVFP4 n2048 k2048
B200
bf16 · [128, 2048] · mm_fp4_out_nvfp4_flashinfer_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 fp6 n2048 k2048
B200
int8 · [128, 1024] · cuda_fp4_fp6_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 fp6 n2048 k2048
B200
int8 · [64, 1024] · cuda_fp4_fp6_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Sparse GEMM fp4 FP8 n2048 k2048
B200
fp8_e4m3 · [2048, 2048] · m=128 · cuda_sparse_fp4_fp8_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Show all 60 challenges ›Showing all 60 challenges ⌄
Sparse GEMM fp4 FP8 n2048 k2048
B200
fp8_e4m3 · [2048, 2048] · m=64 · cuda_sparse_fp4_fp8_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 FP8 n2048 k2048
B200
fp8_e4m3 · [128, 1024] · cuda_fp4_fp8_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 FP8 n2048 k2048
B200
fp8_e4m3 · [64, 1024] · cuda_fp4_fp8_naive_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 e2m1 n4096 k2048
B200
int8 · [1024, 1024] · cublaslt_fp4_e2m1_scaled_mm_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 e2m1 n4096 k2048
B200
int8 · [256, 1024] · cublaslt_fp4_e2m1_scaled_mm_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 e2m1 n4096 k2048
B200
int8 · [128, 1024] · cublaslt_fp4_e2m1_scaled_mm_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 e2m1 n2048 k2048
B200
int8 · [1024, 1024] · cublaslt_fp4_e2m1_scaled_mm_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 e2m1 n2048 k2048
B200
int8 · [256, 1024] · cublaslt_fp4_e2m1_scaled_mm_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM fp4 e2m1 n2048 k2048
B200
int8 · [128, 1024] · cublaslt_fp4_e2m1_scaled_mm_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM FP8 fp4 m masked g4 n4096 k2048
B200
fp32/fp8_e4m3 · [4, 512, 2048] · deepgemm_m_grouped_fp8_fp4_masked_g4_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM FP8 fp4 m masked g4 n2048 k2048
B200
fp32/fp8_e4m3 · [4, 512, 2048] · deepgemm_m_grouped_fp8_fp4_masked_g4_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM FP8 fp4 m contiguous g4 n4096 k2048
B200
fp32/fp8_e4m3 · [1920, 2048] · deepgemm_m_grouped_fp8_fp4_contiguous_g4_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM FP8 fp4 m contiguous g4 n4096 k2048
B200
fp32/fp8_e4m3 · [960, 2048] · deepgemm_m_grouped_fp8_fp4_contiguous_g4_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM FP8 fp4 m contiguous g4 n2048 k2048
B200
fp32/fp8_e4m3 · [1920, 2048] · deepgemm_m_grouped_fp8_fp4_contiguous_g4_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM FP8 fp4 m contiguous g4 n2048 k2048
B200
fp32/fp8_e4m3 · [960, 2048] · deepgemm_m_grouped_fp8_fp4_contiguous_g4_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM mxfp8 MXFP4 g4 n4096 k2048
B200
fp8_e4m3 · [1024, 2048] · group_gemm_mxfp4_flashinfer_g4_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM mxfp8 MXFP4 g4 n4096 k2048
B200
fp8_e4m3 · [512, 2048] · group_gemm_mxfp4_flashinfer_g4_n4096_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM mxfp8 MXFP4 g4 n2048 k2048
B200
fp8_e4m3 · [1024, 2048] · group_gemm_mxfp4_flashinfer_g4_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Grouped GEMM mxfp8 MXFP4 g4 n2048 k2048
B200
fp8_e4m3 · [512, 2048] · group_gemm_mxfp4_flashinfer_g4_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM FP8 fp4 n4096 k4096
B200
fp32/fp8_e4m3 · [1024, 4096] · deepgemm_fp8_fp4_gemm_nt_n4096_k4096 is the source baseline and nobody has beaten it
2026-06-06
GEMM FP8 fp4 n4096 k4096
B200
fp32/fp8_e4m3 · [256, 4096] · deepgemm_fp8_fp4_gemm_nt_n4096_k4096 is the source baseline and nobody has beaten it
2026-06-06
GEMM FP8 fp4 n4096 k4096
B200
fp32/fp8_e4m3 · [128, 4096] · deepgemm_fp8_fp4_gemm_nt_n4096_k4096 is the source baseline and nobody has beaten it
2026-06-06
GEMM FP8 fp4 n2048 k2048
B200
fp32/fp8_e4m3 · [1024, 2048] · deepgemm_fp8_fp4_gemm_nt_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM FP8 fp4 n2048 k2048
B200
fp32/fp8_e4m3 · [256, 2048] · deepgemm_fp8_fp4_gemm_nt_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
GEMM FP8 fp4 n2048 k2048
B200
fp32/fp8_e4m3 · [128, 2048] · deepgemm_fp8_fp4_gemm_nt_n2048_k2048 is the source baseline and nobody has beaten it
2026-06-06
Jensen-Shannon divergence loss
B200
fp32 · [8192, 128256] · PyTorch is the source baseline and nobody has beaten it
2026-05-27
Jensen-Shannon divergence loss
B200
fp32 · [4096, 128256] · PyTorch is the source baseline and nobody has beaten it
2026-05-27
Jensen-Shannon divergence loss
B200
fp32 · [2048, 128256] · PyTorch is the source baseline and nobody has beaten it
2026-05-27
Jensen-Shannon divergence loss
B200
fp32 · [1024, 128256] · PyTorch is the source baseline and nobody has beaten it
2026-05-27
Jensen-Shannon divergence loss
B200
fp32 · [1024, 128256] · PyTorch is the source baseline and nobody has beaten it
2026-05-27
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [4, 32, 128] · num_pages=389 · num_kv_indices=232 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [4, 32, 128] · num_pages=349 · num_kv_indices=212 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [4, 32, 128] · num_pages=398 · num_kv_indices=268 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [128, 32, 128] · num_pages=52033 · num_kv_indices=47438 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [128, 32, 128] · num_pages=51792 · num_kv_indices=46414 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [16, 32, 128] · num_pages=1735 · num_kv_indices=1145 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [16, 32, 128] · num_pages=1686 · num_kv_indices=1033 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [16, 32, 128] · num_pages=1794 · num_kv_indices=1257 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [127, 32, 128] · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [64, 32, 128] · num_pages=29441 · num_kv_indices=27222 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [64, 32, 128] · num_pages=29189 · num_kv_indices=26582 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [63, 32, 128] · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [32, 32, 128] · num_pages=15178 · num_kv_indices=14103 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [32, 32, 128] · num_pages=15114 · num_kv_indices=13783 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [31, 32, 128] · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [8, 32, 128] · num_pages=902 · num_kv_indices=576 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [8, 32, 128] · num_pages=839 · num_kv_indices=528 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA paged decode h32 kv8 d128 ps64
B200
bf16 · [8, 32, 128] · num_pages=910 · num_kv_indices=640 · flashinfer_wrapper_ad4135 is the source baseline and nobody has beaten it
2026-04-30
GQA ragged prefill causal h32 kv8 d64
B200
bf16 · [8819, 8, 64] · flashinfer_wrapper_44b034 is the source baseline and nobody has beaten it
2026-04-20
GQA ragged prefill causal h32 kv8 d64
B200
bf16 · [10930, 8, 64] · flashinfer_wrapper_44b034 is the source baseline and nobody has beaten it
2026-04-20
Unchallenged records
single-entry cohortsStale records
not re-observed in 180 daysOperation / family
GPU
What is missing
Since / asks
AMD MXFP4 GEMM
AMD Instinct MI355X
suite of 6 cases · submission 728835 last observed 2026-04-04
2026-04-04
Fused MoE expert MLP
H100
bf16 · [8192, 2048] · experts=128 · Liger last observed 2026-04-02
2026-04-02
Show all 60 challenges ›Showing all 60 challenges ⌄
Fused MoE expert MLP
H100
bf16 · [8192, 2048] · experts=128 · Liger last observed 2026-04-02
2026-04-02
Fused MoE expert MLP
H100
bf16 · [8192, 2048] · experts=128 · Liger last observed 2026-04-02
2026-04-02
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [413, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [79, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [1834, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [1825, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [4743, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [1307, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [2406, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [170, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [559, 24, 128] · num_pages=5878 · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [332, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [987, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [4812, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [1728, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [82, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [247, 24, 128] · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
GQA paged prefill causal h24 kv4 d128 ps1
B200
bf16 · [559, 24, 128] · num_pages=718 · flashinfer_wrapper_0b1670 last observed 2026-03-31
2026-03-31
Priority coverage
FamilyH100B200All GPUs
Closing a challenge means publishing evidence: a benchmark run with its workload, protocol, environment, and source.Contribute evidence →JSON