NVIDIA H100 80GB HBM3 · median of 10 runs after 3 warm-ups · measured 2026-09-28T19:15:03+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 PyTorch1.29× fasterlatency

• Linnet, CUDA graphs1.31 mstransformers, compiled1.69 ms

vs torch.onnx, on ONNX Runtime1.01× fasterlatency

• Linnet, ONNX Runtime f326.29 mstorch.onnx export6.38 ms

vs torch.onnx, behind Triton1.17× fasterlatency

• Triton, Linnet's ONNX27.1 msTriton, torch.onnx export31.6 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)2.95 msthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)1.69 ms1.7x the reference's speedDistance from reference: 1.53

Linnet torch (generated source)2.13 ms0.79× the speed of transformers, compiledDistance from reference: 0.625

Linnet torch (torch.compile, inductor)1.51 ms1.12× the speed of transformers, compiledDistance from reference: 2.23

Linnet torch (CUDA graphs)1.31 ms1.29× the speed of transformers, compiledDistance from reference: 2.23

Linnet JAX (XLA, StableHLO)2.50 ms1.2x the reference's speedDistance from reference: 4.16bf16, 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)2.26 ms1.3x the reference's speedDistance from reference: 1.78bf16, 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 Runtime6.38 ms0.5x the reference's speedDistance from reference: 1.09the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)6.29 ms1.01× the speed of torch.onnx exportDistance from reference: 1.09f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)3.95 ms0.7x the reference's speedDistance from reference: 1.24f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)4.92 ms0.6x the reference's speedDistance from reference: 6.88bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)4.32 ms0.7x the reference's speedDistance from reference: 1.09f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)1.16 ms2.5x the reference's speedDistance from reference: 1.3f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)1.26 ms2.3x the reference's speedDistance from reference: 1.03bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)31.6 ms0.09x the reference's speedDistance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)27.1 ms1.17× the speed of Triton, torch.onnx exportDistance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)25.6 ms0.1x the reference's speedDistance from reference: 1.05over 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)1,170/sthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)1,333/s1.1x the reference's speedDistance from reference: 1.53

Linnet torch (generated source)1,064/s0.80× the speed of transformers, compiledDistance from reference: 0.625

Linnet torch (torch.compile, inductor)1,415/s1.06× the speed of transformers, compiledDistance from reference: 2.23

Linnet torch (CUDA graphs)1,405/s1.05× the speed of transformers, compiledDistance from reference: 2.23

Linnet JAX (XLA, StableHLO)603/s0.5x the reference's speedDistance from reference: 4.16bf16, 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)1,137/s1.0x the reference's speedDistance from reference: 1.78bf16, 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 Runtime253/s0.2x the reference's speedDistance from reference: 1.09the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)254/s1.01× the speed of torch.onnx exportDistance from reference: 1.09f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)452/s0.4x the reference's speedDistance from reference: 1.24f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)355/s0.3x the reference's speedDistance from reference: 6.88bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)300/s0.3x the reference's speedDistance from reference: 1.09f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)1,471/s1.3x the reference's speedDistance from reference: 1.3f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)1,412/s1.2x the reference's speedDistance from reference: 1.03bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)48.8/s0.04x the reference's speedDistance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)47.3/s0.97× the speed of Triton, torch.onnx exportDistance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)47.7/s0.04x the reference's speedDistance from reference: 1.05over 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)6.78 sthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)3.67 s1.8x the reference's speedDistance from reference: 1.53

Linnet torch (generated source)1.13 s3.25× the speed of transformers, compiledDistance from reference: 0.625

Linnet torch (torch.compile, inductor)1.29 s2.85× the speed of transformers, compiledDistance from reference: 2.23

Linnet torch (CUDA graphs)1.18 s3.11× the speed of transformers, compiledDistance from reference: 2.23

Linnet JAX (XLA, StableHLO)0.46 s15x the reference's speedDistance from reference: 4.16bf16, 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)0.54 s12x the reference's speedDistance from reference: 1.78bf16, 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 Runtime17.3 s0.4x the reference's speedDistance from reference: 1.09the 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: 1.09f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)0.31 s22x the reference's speedDistance from reference: 1.24f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)0.29 s23x the reference's speedDistance from reference: 6.88bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)0.28 s24x the reference's speedDistance from reference: 1.09f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.34 s20x the reference's speedDistance from reference: 1.3f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.28 s24x the reference's speedDistance from reference: 1.03bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)12.0 s0.6x the reference's speedDistance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)15.0 s0.80× the speed of Triton, torch.onnx exportDistance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)6.02 s1.1x the reference's speedDistance from reference: 1.05over 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)2.11 GiB (2,164 MiB)Distance from reference: 0 (the reference)

transformers (torch.compile)1.83 GiB (1,870 MiB)Distance from reference: 1.53

Linnet torch (generated source)1.83 GiB (1,870 MiB)Distance from reference: 0.625

Linnet torch (torch.compile, inductor)1.62 GiB (1,662 MiB)Distance from reference: 2.23

Linnet torch (CUDA graphs)1.77 GiB (1,816 MiB)Distance from reference: 2.23

Linnet JAX (XLA, StableHLO)3.34 GiB (3,419 MiB)Distance from reference: 4.16bf16, 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)1.26 GiB (1,288 MiB)Distance from reference: 1.78bf16, 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 Runtime19.9 GiB (20,406 MiB)Distance from reference: 1.09the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)15.9 GiB (16,330 MiB)Distance from reference: 1.09f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)8.42 GiB (8,620 MiB)Distance from reference: 1.24f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)8.47 GiB (8,672 MiB)Distance from reference: 6.88bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)9.50 GiB (9,724 MiB)Distance from reference: 1.09f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)9.15 GiB (9,372 MiB)Distance from reference: 1.3f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)9.15 GiB (9,372 MiB)Distance from reference: 1.03bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Triton Inference Server (ONNX Runtime backend, torch.onnx export)19.9 GiB (20,351 MiB)Distance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)19.9 GiB (20,355 MiB)Distance from reference: 1.03over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)3.86 GiB (3,948 MiB)Distance from reference: 1.05over 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	2.95 ms	1,170/s	6.78 s	2.11 GiB		0 (the reference)
transformers, compiled	1.69 ms	1,333/s	3.67 s	1.83 GiB		1.53
Linnet, generated PyTorch	2.13 ms	1,064/s	1.13 s	1.83 GiB	0.79× transformers, compiled	0.625
Linnet, inductor	1.51 ms	1,415/s	1.29 s	1.62 GiB	1.12× transformers, compiled	2.23
Linnet, CUDA graphs	1.31 ms	1,405/s	1.18 s	1.77 GiB	1.29× transformers, compiled	2.23
JAX
Linnet, XLA (StableHLO)	2.50 ms	603/s	0.46 s	3.34 GiB		4.16	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)	2.26 ms	1,137/s	0.54 s	1.26 GiB		1.78	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	6.38 ms	253/s	17.3 s	19.9 GiB		1.09	the reference model's own ONNX export, f32 like Linnet's;
Linnet, ONNX Runtime f32	6.29 ms	254/s	0.30 s	15.9 GiB	1.01× torch.onnx export	1.09	f32; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, ONNX Runtime f16	3.95 ms	452/s	0.31 s	8.42 GiB		1.24	f16; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, ONNX Runtime bf16	4.92 ms	355/s	0.29 s	8.47 GiB		6.88	bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, TensorRT f32	4.32 ms	300/s	0.28 s	9.50 GiB		1.09	f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Linnet, TensorRT f16	1.16 ms	1,471/s	0.34 s	9.15 GiB		1.3	f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Linnet, TensorRT bf16	1.26 ms	1,412/s	0.28 s	9.15 GiB		1.03	bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Triton Inference Server
Triton, torch.onnx export	31.6 ms	48.8/s	12.0 s	19.9 GiB		1.03	over HTTP; f32; throughput with 4 requests of batch 32 in flight
Triton, Linnet's ONNX	27.1 ms	47.3/s	15.0 s	19.9 GiB	1.17× Triton, torch.onnx export	1.03	over HTTP; f32; throughput with 4 requests of batch 32 in flight
Triton, Linnet Python backend	25.6 ms	47.7/s	6.02 s	3.86 GiB		1.05	over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight