NVIDIA H100 80GB HBM3 · median of 10 runs after 3 warm-ups · measured 2026-09-28T19:01:59+00:00

Linnet 0.1.0, PyTorch 2.14.0, JAX 0.11.2, transformers 5.17.0, diffusers 0.40.0, ONNX Runtime 1.30.0, Python 3.12.3, driver 580.126.09

Linnet against the stack it replaces

Each side in its fastest configuration, on the same GPU and checkpoint. A speed-up is how many times the other's speed; an export is measured against the original checkpoint in the same engine.

vs the reference, in PyTorch2.42× fasterlatency

• Linnet, CUDA graphs0.81 mstransformers, compiled1.97 ms

vs KerasHub, in JAX5.63× fasterlatency

• Linnet, XLA (generated JAX)1.00 msKerasHub5.61 ms

vs torch.onnx, on ONNX Runtime1.14× fasterlatency

• Linnet, ONNX Runtime f322.13 mstorch.onnx export2.43 ms

vs torch.onnx, behind Triton1.02× fasterlatency

• Triton, Linnet's ONNX3.81 msTriton, torch.onnx export3.87 ms

Best in the chartLinnetReference

Each row also carries its distance from reference: the largest absolute difference between its output and transformers (eager)'s on the same input. bf16 outputs differ by rounding (about 0.1 on logits near 16), so a small number is expected; it is there so that a fast wrong answer cannot look like a win.

Latency

ms, lower is better; the best bar is red.

transformers (eager)2.31 msthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)1.97 ms1.2x the reference's speedDistance from reference: 0.0234

Linnet torch (generated source)2.16 ms0.91× the speed of transformers, compiledDistance from reference: 0.0313

Linnet torch (torch.compile, inductor)1.68 ms1.17× the speed of transformers, compiledDistance from reference: 0.0195

Linnet torch (CUDA graphs)0.81 ms2.42× the speed of transformers, compiledDistance from reference: 0.0195

KerasHub (JAX)5.61 ms0.4x the reference's speedDistance from reference: 0.0625bf16, under jax.jit; ; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)1.37 ms4.09× the speed of KerasHubDistance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)1.00 ms5.63× the speed of KerasHubDistance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime2.43 ms1.0x the reference's speedDistance from reference: 0.0396the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)2.13 ms1.14× the speed of torch.onnx exportDistance from reference: 0.0397f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)1.91 ms1.2x the reference's speedDistance from reference: 0.041f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)1.76 ms1.3x the reference's speedDistance from reference: 0.0313bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)0.95 ms2.4x the reference's speedDistance from reference: 0.0403f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.61 ms3.8x the reference's speedDistance from reference: 0.041f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.64 ms3.6x the reference's speedDistance from reference: 0.0313bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)3.87 ms0.6x the reference's speedDistance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)3.81 ms1.02× the speed of Triton, torch.onnx exportDistance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)3.21 ms0.7x the reference's speedDistance from reference: 0.0313over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight

Throughput

/s, higher is better; the best bar is red.

transformers (eager)8,811/sthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)4,916/s0.6x the reference's speedDistance from reference: 0.0234

Linnet torch (generated source)8,215/s0.93× the speed of transformersDistance from reference: 0.0313

Linnet torch (torch.compile, inductor)9,646/s1.09× the speed of transformersDistance from reference: 0.0195

Linnet torch (CUDA graphs)10,007/s1.14× the speed of transformersDistance from reference: 0.0195

KerasHub (JAX)1,814/s0.2x the reference's speedDistance from reference: 0.0625bf16, under jax.jit; ; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)6,835/s3.77× the speed of KerasHubDistance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)7,063/s3.89× the speed of KerasHubDistance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime2,296/s0.3x the reference's speedDistance from reference: 0.0396the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)2,657/s1.16× the speed of torch.onnx exportDistance from reference: 0.0397f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)4,406/s0.5x the reference's speedDistance from reference: 0.041f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)4,090/s0.5x the reference's speedDistance from reference: 0.0313bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)4,340/s0.5x the reference's speedDistance from reference: 0.0403f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)10,975/s1.2x the reference's speedDistance from reference: 0.041f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)10,282/s1.2x the reference's speedDistance from reference: 0.0313bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)1,391/s0.2x the reference's speedDistance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)1,241/s0.89× the speed of Triton, torch.onnx exportDistance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)1,272/s0.1x the reference's speedDistance from reference: 0.0313over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight

Load time

s, lower is better; the best bar is red.

transformers (eager)9.36 sthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)3.83 s2.4x the reference's speedDistance from reference: 0.0234

Linnet torch (generated source)1.06 s3.61× the speed of transformers, compiledDistance from reference: 0.0313

Linnet torch (torch.compile, inductor)1.70 s2.26× the speed of transformers, compiledDistance from reference: 0.0195

Linnet torch (CUDA graphs)1.17 s3.27× the speed of transformers, compiledDistance from reference: 0.0195

KerasHub (JAX)5.27 s1.8x the reference's speedDistance from reference: 0.0625bf16, under jax.jit; ; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)0.56 s9.44× the speed of KerasHubDistance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)0.51 s10× the speed of KerasHubDistance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime17.2 s0.5x the reference's speedDistance from reference: 0.0396the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)0.30 s58× the speed of torch.onnx exportDistance from reference: 0.0397f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)0.28 s34x the reference's speedDistance from reference: 0.041f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)0.29 s32x the reference's speedDistance from reference: 0.0313bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)0.29 s32x the reference's speedDistance from reference: 0.0403f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.28 s33x the reference's speedDistance from reference: 0.041f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.29 s32x the reference's speedDistance from reference: 0.0313bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)10.0 s0.9x the reference's speedDistance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)16.1 s0.62× the speed of Triton, torch.onnx exportDistance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)6.02 s1.6x the reference's speedDistance from reference: 0.0313over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight

Peak GPU memory

What the driver reports the process holding at its peak, in GiB; lower is better; the best bar is red.

transformers (eager)1.04 GiB (1,060 MiB)Distance from reference: 0 (the reference)

transformers (torch.compile)1.14 GiB (1,166 MiB)Distance from reference: 0.0234

Linnet torch (generated source)1.02 GiB (1,042 MiB)Distance from reference: 0.0313

Linnet torch (torch.compile, inductor)1.04 GiB (1,062 MiB)Distance from reference: 0.0195

Linnet torch (CUDA graphs)1.16 GiB (1,190 MiB)Distance from reference: 0.0195

KerasHub (JAX)0.41 GiB (415 MiB)Distance from reference: 0.0625bf16, under jax.jit; ; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)0.63 GiB (640 MiB)Distance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)0.61 GiB (624 MiB)Distance from reference: 0.0313bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime2.34 GiB (2,392 MiB)Distance from reference: 0.0396the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)2.42 GiB (2,474 MiB)Distance from reference: 0.0397f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)1.65 GiB (1,692 MiB)Distance from reference: 0.041f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)1.67 GiB (1,708 MiB)Distance from reference: 0.0313bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)3.07 GiB (3,140 MiB)Distance from reference: 0.0403f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)2.87 GiB (2,940 MiB)Distance from reference: 0.041f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)2.87 GiB (2,940 MiB)Distance from reference: 0.0313bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)2.38 GiB (2,435 MiB)Distance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)2.37 GiB (2,429 MiB)Distance from reference: 0.0413over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)2.81 GiB (2,880 MiB)Distance from reference: 0.0313over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight

Every number

Method	Latency	Throughput	Load time	Peak GPU memory	Against the stack it replaces	Distance from reference	Notes
PyTorch
transformers	2.31 ms	8,811/s	9.36 s	1.04 GiB		0 (the reference)
transformers, compiled	1.97 ms	4,916/s	3.83 s	1.14 GiB		0.0234
Linnet, generated PyTorch	2.16 ms	8,215/s	1.06 s	1.02 GiB	0.91× transformers, compiled	0.0313
Linnet, inductor	1.68 ms	9,646/s	1.70 s	1.04 GiB	1.17× transformers, compiled	0.0195
Linnet, CUDA graphs	0.81 ms	10,007/s	1.17 s	1.16 GiB	2.42× transformers, compiled	0.0195
JAX
KerasHub	5.61 ms	1,814/s	5.27 s	0.41 GiB		0.0625	bf16, under jax.jit; ; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions
Linnet, XLA (StableHLO)	1.37 ms	6,835/s	0.56 s	0.63 GiB	4.09× KerasHub	0.0313	bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions
Linnet, XLA (generated JAX)	1.00 ms	7,063/s	0.51 s	0.61 GiB	5.63× KerasHub	0.0313	bf16, like the torch rows; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions
ONNX Runtime and TensorRT
torch.onnx export	2.43 ms	2,296/s	17.2 s	2.34 GiB		0.0396	the reference model's own ONNX export, f32 like Linnet's;
Linnet, ONNX Runtime f32	2.13 ms	2,657/s	0.30 s	2.42 GiB	1.14× torch.onnx export	0.0397	f32; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, ONNX Runtime f16	1.91 ms	4,406/s	0.28 s	1.65 GiB		0.041	f16; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, ONNX Runtime bf16	1.76 ms	4,090/s	0.29 s	1.67 GiB		0.0313	bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, TensorRT f32	0.95 ms	4,340/s	0.29 s	3.07 GiB		0.0403	f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Linnet, TensorRT f16	0.61 ms	10,975/s	0.28 s	2.87 GiB		0.041	f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Linnet, TensorRT bf16	0.64 ms	10,282/s	0.29 s	2.87 GiB		0.0313	bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Triton Inference Server
Triton, torch.onnx export	3.87 ms	1,391/s	10.0 s	2.38 GiB		0.0413	over HTTP; f32; throughput with 4 requests of batch 32 in flight
Triton, Linnet's ONNX	3.81 ms	1,241/s	16.1 s	2.37 GiB	1.02× Triton, torch.onnx export	0.0413	over HTTP; f32; throughput with 4 requests of batch 32 in flight
Triton, Linnet Python backend	3.21 ms	1,272/s	6.02 s	2.81 GiB		0.0313	over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight