Skip to content
KernelIndex
Search⌘K
Sign in

gpt-o3

Competition author · flashinfer-ai/flashinfer-trace · Apache-2.0
Records held
30as of 2026-10-02 00:50 UTC
Entries
41
Runs
1,033
Is this you? Claim this project ›

A claim grants attribution and metadata maintenance, never the right to edit evidence. Evidence claims are reviewed by a maintainer.

Current records

Operation · workload
History
Record
Implementation
Hardware
Since
RMSNorm h512bf16 · [512] · batch 7
6.19µs
1.7% faster · was 6.29 µs
gpt-o3 / triton19c647
B200
2025-10-16
RMSNorm h512bf16 · [512] · batch 1
6.21µs
4.6% faster · was 6.50 µs
gpt-o3 / triton19c647
B200
2025-10-16
RMSNorm h512bf16 · [512] · batch 539
6.19µs
2.5% faster · was 6.35 µs
gpt-o3 / triton19c647
B200
2025-10-16
RMSNorm h128bf16 · [128] · batch 136
6.17µs
0.9% faster · was 6.23 µs
gpt-o3 / cuda64f9f5
B200
2025-10-16
RMSNorm h128bf16 · [128] · batch 2528
8.01µs
0.8% faster · was 8.07 µs
gpt-o3 / triton35b90e
B200
2025-10-16
RMSNorm h128bf16 · [128] · batch 192
6.18µs
0.2% faster · was 6.19 µs
gpt-o3 / cuda64f9f5
B200
2025-10-16
Fused add RMSNorm h2048bf16 · [2048] · batch 6
6.21µs
8.2% faster · was 6.77 µs
gpt-o3 / triton2e18c2
B200
2025-10-16
Fused add RMSNorm h2048bf16 · [2048] · batch 1
6.25µs
56.4% faster · was 14.3 µs
gpt-o3 / triton2e18c2
B200
2025-10-16
Fused add RMSNorm h2048bf16 · [2048] · batch 79
6.19µs
5.2% faster · was 6.53 µs
gpt-o3 / triton2e18c2
B200
2025-10-16
Fused add RMSNorm h2048bf16 · [2048] · batch 64
6.20µs
8.3% faster · was 6.76 µs
gpt-o3 / triton2e18c2
B200
2025-10-16
Show all 30 records ›
GQA paged prefill causal h32 kv8 d128 ps1bf16 · [2, 32, 128]
31.5µs
gpt-o3 / cudad4241d
B200
2025-10-21
GQA paged decode h32 kv4 d128 ps1bf16 · [1, 32, 128] · num pages 10 · num kv indices 9
19.4µs
22.0% faster · was 24.9 µs
gpt-o3 / cuda5548d0
B200
2026-03-28
GQA paged decode h32 kv4 d128 ps1bf16 · [1, 32, 128] · num pages 17 · num kv indices 2
13.3µs
23.8% faster · was 17.5 µs
gpt-o3 / cuda5548d0
B200
2026-03-28
GEMM n5120 k2048fp16 · [12853, 2048]
194.3µs
39.0% faster · was 318.6 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [11006, 2048]
167.5µs
35.9% faster · was 261.4 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [8828, 2048]
130.7µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [128, 2048]
12.8µs
52.0% faster · was 26.7 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [12251, 2048]
180.0µs
37.6% faster · was 288.4 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [63, 2048]
12.3µs
53.5% faster · was 26.5 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [2, 2048]
12.1µs
54.6% faster · was 26.6 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [93, 2048]
12.7µs
52.4% faster · was 26.6 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [289, 2048]
16.2µs
39.3% faster · was 26.6 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [5, 2048]
12.2µs
81.2% faster · was 65.1 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [25, 2048]
12.7µs
52.1% faster · was 26.5 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [8, 2048]
12.1µs
81.2% faster · was 64.8 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [1, 2048]
12.3µs
39.8% faster · was 20.5 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [34, 2048]
12.5µs
38.8% faster · was 20.4 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [952, 2048]
22.2µs
42.9% faster · was 38.9 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [16, 2048]
12.3µs
39.5% faster · was 20.4 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16
GEMM n5120 k2048fp16 · [172, 2048]
14.6µs
29.3% faster · was 20.7 µs
gpt-o3 / cudaaf0f3d
B200
2025-10-16

Open in the records ledger →

Measured entries

Implementation
Operation
Runtime
Best latency
Evidence
Availability
gpt-o3 / cuda64f9f5gpt-o3
RMSNorm h128
cuda
6.17µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / triton19c647gpt-o3
RMSNorm h512
triton
6.19µs
Reported
Apache-2.0 · source
gpt-o3 / triton2e18c2gpt-o3
Fused add RMSNorm h2048
triton
6.19µs
Reported
Apache-2.0 · source
gpt-o3 / triton35b90egpt-o3
RMSNorm h128
triton
6.30µs
Reported
Apache-2.0 · source
gpt-o3 / cuda3def09gpt-o3
Fused add RMSNorm h2048
cuda
7.28µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cudaa1fa19gpt-o3
RMSNorm h512
cuda
9.32µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / triton793f87gpt-o3
RMSNorm h4096
triton
10.1µs
Reported
Apache-2.0 · source
gpt-o3 / tritonc1e819gpt-o3
Fused add RMSNorm h4096
triton
10.2µs
Reported
Apache-2.0 · source
gpt-o3 / tritond1dccegpt-o3
RMSNorm h1536
triton
10.3µs
Reported
Apache-2.0 · source
gpt-o3 / tritonc3c0ccgpt-o3
GQA paged decode h32 kv8 d128 ps1
triton
10.3µs
Reported
Apache-2.0 · source
Show all 40 entries ›
gpt-o3 / cudaec73eagpt-o3
GQA paged decode h32 kv8 d128 ps1
cuda
10.4µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cudaaf0f3dgpt-o3
GEMM n5120 k2048
cuda
12.1µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / triton38f281gpt-o3
RMSNorm h2048
triton
12.3µs
Reported
Apache-2.0 · source
gpt-o3 / cuda5548d0gpt-o3
GQA paged decode h32 kv4 d128 ps1
cuda
13.3µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cuda620cb5gpt-o3
Fused add RMSNorm h7168
cuda
16.0µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cudaa7bbcfgpt-o3
Fused add RMSNorm h4096
cuda
16.4µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cuda025ec1gpt-o3
GQA ragged prefill causal h32 kv8 d128
cuda
18.6µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cudad4241dgpt-o3
GQA paged prefill causal h32 kv8 d128 ps1
cuda
31.5µs
Reported
Apache-2.0 · source
gpt-o3 / tritone5fd98gpt-o3
Fused add RMSNorm h7168
triton
32.7µs
Reported
Apache-2.0 · source
gpt-o3 / cuda9abd34gpt-o3
GQA ragged prefill causal h32 kv4 d128
cuda
47.7µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / tritona796abgpt-o3
GEMM n128 k2048
triton
51.4µs
Reported
Apache-2.0 · source
gpt-o3 / triton951f7egpt-o3
RMSNorm h7168
triton
55.3µs
Reported
Apache-2.0 · source
gpt-o3 / triton4c17a1gpt-o3
MLA paged decode h16 ckv512 kpe64 ps1
triton
59.1µs
Reported
Apache-2.0 · source
gpt-o3_cuda_696722gpt-o3
GEMM n28672 k4096
cuda
59.6µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / triton63bfe6gpt-o3
GEMM n5120 k2048
triton
61.8µs
Reported
Apache-2.0 · source
gpt-o3 / tritonfc01fagpt-o3
GEMM n6144 k4096
triton
69.9µs
Reported
Apache-2.0 · source
gpt-o3 / triton4c9c32gpt-o3
GEMM n28672 k4096
triton
85.3µs
Reported
Apache-2.0 · source
gpt-o3 / tritonb12b97gpt-o3
GQA paged decode h32 kv4 d128 ps1
triton
86.6µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / triton25db20gpt-o3
GQA ragged prefill causal h32 kv8 d128
triton
93.6µs
Reported
Apache-2.0 · source
gpt-o3 / tritonef0d11gpt-o3
GEMM n4096 k4096
triton
103.6µs
Reported
Apache-2.0 · source
gpt-o3 / triton6fd1efgpt-o3
GQA paged prefill causal h32 kv4 d128 ps1
triton
143.7µs
Reported
Apache-2.0 · source
gpt-o3 / tritonad56c1gpt-o3
MLA paged prefill causal h16 ckv512 kpe64 ps1
triton
144.3µs
Reported
Apache-2.0 · source
gpt-o3 / triton93df2bgpt-o3
GEMM n2048 k4096
triton
144.6µs
Reported
Apache-2.0 · source
gpt-o3 / cuda1dc3e4gpt-o3
MLA paged prefill causal h16 ckv512 kpe64 ps1
cuda
189.1µs
Reproduction-ready
Apache-2.0 · source
gpt-o3_triton_189201gpt-o3
GEMM n256 k7168
triton
290.3µs
Reported
Apache-2.0 · source
gpt-o3_triton_a47205gpt-o3
GEMM n4096 k14336
triton
304.7µs
Reported
Apache-2.0 · source
gpt-o3_cuda_270394gpt-o3
GEMM n128 k2048
cuda
347.0µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cuda5a050dgpt-o3
GEMM n2048 k4096
cuda
404.8µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / cuda7a2145gpt-o3
GEMM n256 k7168
cuda
523.3µs
Reproduction-ready
Apache-2.0 · source
gpt-o3 / triton2b4be8gpt-o3
GQA paged prefill causal h32 kv8 d128 ps1
triton
4.71ms
Reported
Apache-2.0 · source

Showing 40 of 41; the full list is in the API dossier below.

Record activity

Nov
Dec
Jan
Feb
4Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Sources: FlashInfer-Bench (2026-03-28) · Apache-2.0last observed 2026-03-28JSON