Skip to content
KernelIndex
Search⌘K
Sign in

Megatron fused

NVIDIA Megatron-LM · python · BSD-3-Clause

Use it

Source available upstream · BSD-3-ClauseView source →

No install recipe recorded for this revision.

interfacepython
Compatibility
measured onNVIDIA H100
declared hardwaredeclared only
architectures—
dtypes

Benchmark evidence

36 measurements across 1 GPU, fastest first.

Operation / workload
Hardware
Latency
Rank
Observed
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]
NVIDIA H100
312.2µs
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]
NVIDIA H100
527.4µs
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]
NVIDIA H100
531.1µs
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]
NVIDIA H100
558.5µs
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]
NVIDIA H100
635.9µs
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]
NVIDIA H100
767.6µs
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]
NVIDIA H100
769.3µs
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]
NVIDIA H100
1.05ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]
NVIDIA H100
1.05ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]
NVIDIA H100
1.08ms
#3 of 3
2026-06-15
Run →
Show all 36 measurements ›
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]
NVIDIA H100
1.35ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]
NVIDIA H100
1.45ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]
NVIDIA H100
1.46ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]
NVIDIA H100
2.07ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]
NVIDIA H100
2.16ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]
NVIDIA H100
2.16ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]
NVIDIA H100
2.78ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]
NVIDIA H100
2.78ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]
NVIDIA H100
2.83ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]
NVIDIA H100
4.02ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]
NVIDIA H100
4.45ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]
NVIDIA H100
4.45ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]
NVIDIA H100
5.41ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]
NVIDIA H100
5.41ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]
NVIDIA H100
5.76ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]
NVIDIA H100
7.93ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]
NVIDIA H100
8.98ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]
NVIDIA H100
8.98ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]
NVIDIA H100
10.7ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]
NVIDIA H100
10.7ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]
NVIDIA H100
11.8ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]
NVIDIA H100
15.8ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]
NVIDIA H100
18.2ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]
NVIDIA H100
18.2ms
#2 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]
NVIDIA H100
21.2ms
#3 of 3
2026-06-15
Run →
Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]
NVIDIA H100
21.2ms
#3 of 3
2026-06-15
Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:d9055243d96f775b3343dd8a494ac60fd04f4df9fdd5506871ddbce7ad70a00c
license declaredBSD-3-Clause
license concludedBSD-3-Clause
imported2026-08-20

Best evidence level for this revision: reported

JSON