NVIDIA H100 80GB HBM3 · median of 10 runs after 3 warm-ups · measured 2026-09-28T18:38:52+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 PyTorch1.77× fasterlatency

• Linnet, CUDA graphs0.33 mstransformers, compiled0.59 ms

vs torch.onnx, on ONNX Runtime1.02× fasterlatency

• Linnet, ONNX Runtime f320.64 mstorch.onnx export0.65 ms

vs torch.onnx, behind Triton0.98× slowerlatency

• Triton, Linnet's ONNX3.18 msTriton, torch.onnx export3.10 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)1.21 msthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)0.59 ms2.0x the reference's speedDistance from reference: 0.0313

Linnet torch (generated source)1.04 ms0.57× the speed of transformers, compiledDistance from reference: 0

Linnet torch (torch.compile, inductor)0.63 ms0.94× the speed of transformers, compiledDistance from reference: 0.0313

Linnet torch (CUDA graphs)0.33 ms1.77× the speed of transformers, compiledDistance from reference: 0.0313

Linnet JAX (XLA, StableHLO)0.64 ms1.9x the reference's speedDistance 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)1.10 ms1.1x the reference's speedDistance 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 Runtime0.65 ms1.9x the reference's speedDistance from reference: 0.0894the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)0.64 ms1.02× the speed of torch.onnx exportDistance from reference: 0.0894f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)0.53 ms2.3x the reference's speedDistance from reference: 0.102f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)1.05 ms1.2x 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)0.27 ms4.6x the reference's speedDistance from reference: 0.0895f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.21 ms5.7x the reference's speedDistance from reference: 0.104f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.45 ms2.7x 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)3.10 ms0.4x the reference's speedDistance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)3.18 ms0.98× the speed of Triton, torch.onnx exportDistance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)2.76 ms0.4x the reference's speedDistance from reference: 0.0156over 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)18,541/sthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)23,490/s1.3x the reference's speedDistance from reference: 0.0313

Linnet torch (generated source)18,580/s0.79× the speed of transformers, compiledDistance from reference: 0

Linnet torch (torch.compile, inductor)24,100/s1.03× the speed of transformers, compiledDistance from reference: 0.0313

Linnet torch (CUDA graphs)25,051/s1.07× the speed of transformers, compiledDistance from reference: 0.0313

Linnet JAX (XLA, StableHLO)32,230/s1.7x the reference's speedDistance 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)31,053/s1.7x the reference's speedDistance 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 Runtime15,125/s0.8x the reference's speedDistance from reference: 0.0894the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)14,894/s0.98× the speed of torch.onnx exportDistance from reference: 0.0894f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)19,847/s1.1x the reference's speedDistance from reference: 0.102f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)11,198/s0.6x 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)36,114/s1.9x the reference's speedDistance from reference: 0.0895f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)59,978/s3.2x the reference's speedDistance from reference: 0.104f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)18,561/s1.0x 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)1,454/s0.08x the reference's speedDistance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)1,243/s0.85× the speed of Triton, torch.onnx exportDistance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)1,220/s0.07x the reference's speedDistance from reference: 0.0156over 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)3.95 sthe referenceDistance from reference: 0 (the reference)

transformers (torch.compile)3.80 s1.0x the reference's speedDistance from reference: 0.0313

Linnet torch (generated source)0.51 s7.45× the speed of transformers, compiledDistance from reference: 0

Linnet torch (torch.compile, inductor)0.63 s6.05× the speed of transformers, compiledDistance from reference: 0.0313

Linnet torch (CUDA graphs)0.62 s6.14× the speed of transformers, compiledDistance from reference: 0.0313

Linnet JAX (XLA, StableHLO)0.37 s11x the reference's speedDistance 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)0.43 s9.1x the reference's speedDistance 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 Runtime11.9 s0.3x the reference's speedDistance from reference: 0.0894the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)0.30 s40× the speed of torch.onnx exportDistance from reference: 0.0894f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)0.29 s14x the reference's speedDistance from reference: 0.102f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)0.31 s13x 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 s14x the reference's speedDistance from reference: 0.0895f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.29 s14x the reference's speedDistance from reference: 0.104f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.33 s12x 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)4.06 s1.0x the reference's speedDistance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)9.08 s0.45× the speed of Triton, torch.onnx exportDistance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)6.02 s0.7x the reference's speedDistance from reference: 0.0156over 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)0.91 GiB (928 MiB)Distance from reference: 0 (the reference)

transformers (torch.compile)0.92 GiB (946 MiB)Distance from reference: 0.0313

Linnet torch (generated source)0.91 GiB (928 MiB)Distance from reference: 0

Linnet torch (torch.compile, inductor)0.86 GiB (878 MiB)Distance from reference: 0.0313

Linnet torch (CUDA graphs)0.95 GiB (968 MiB)Distance from reference: 0.0313

Linnet JAX (XLA, StableHLO)0.43 GiB (444 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)0.50 GiB (513 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 Runtime1.28 GiB (1,312 MiB)Distance from reference: 0.0894the reference model's own ONNX export, f32 like Linnet's;

Linnet ONNX f32 -> ONNX Runtime (CUDA)1.28 GiB (1,308 MiB)Distance from reference: 0.0894f32; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (CUDA)1.05 GiB (1,080 MiB)Distance from reference: 0.102f16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (CUDA)1.52 GiB (1,552 MiB)Distance from reference: 0.0313bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions

Linnet ONNX f32 -> ONNX Runtime (TensorRT)2.52 GiB (2,580 MiB)Distance from reference: 0.0895f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX f16 -> ONNX Runtime (TensorRT)2.40 GiB (2,454 MiB)Distance from reference: 0.104f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)2.58 GiB (2,638 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)1.50 GiB (1,541 MiB)Distance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)1.50 GiB (1,539 MiB)Distance from reference: 0.0894over HTTP; f32; throughput with 4 requests of batch 32 in flight

Triton Inference Server (Python backend, Linnet torch)2.38 GiB (2,436 MiB)Distance from reference: 0.0156over 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	1.21 ms	18,541/s	3.95 s	0.91 GiB		0 (the reference)
transformers, compiled	0.59 ms	23,490/s	3.80 s	0.92 GiB		0.0313
Linnet, generated PyTorch	1.04 ms	18,580/s	0.51 s	0.91 GiB	0.57× transformers, compiled	0
Linnet, inductor	0.63 ms	24,100/s	0.63 s	0.86 GiB	0.94× transformers, compiled	0.0313
Linnet, CUDA graphs	0.33 ms	25,051/s	0.62 s	0.95 GiB	1.77× transformers, compiled	0.0313
JAX
Linnet, XLA (StableHLO)	0.64 ms	32,230/s	0.37 s	0.43 GiB		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)	1.10 ms	31,053/s	0.43 s	0.50 GiB		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	0.65 ms	15,125/s	11.9 s	1.28 GiB		0.0894	the reference model's own ONNX export, f32 like Linnet's;
Linnet, ONNX Runtime f32	0.64 ms	14,894/s	0.30 s	1.28 GiB	1.02× torch.onnx export	0.0894	f32; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, ONNX Runtime f16	0.53 ms	19,847/s	0.29 s	1.05 GiB		0.102	f16; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, ONNX Runtime bf16	1.05 ms	11,198/s	0.31 s	1.52 GiB		0.0313	bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions
Linnet, TensorRT f32	0.27 ms	36,114/s	0.29 s	2.52 GiB		0.0895	f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Linnet, TensorRT f16	0.21 ms	59,978/s	0.29 s	2.40 GiB		0.104	f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Linnet, TensorRT bf16	0.45 ms	18,561/s	0.33 s	2.58 GiB		0.0313	bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions
Triton Inference Server
Triton, torch.onnx export	3.10 ms	1,454/s	4.06 s	1.50 GiB		0.0894	over HTTP; f32; throughput with 4 requests of batch 32 in flight
Triton, Linnet's ONNX	3.18 ms	1,243/s	9.08 s	1.50 GiB	0.98× Triton, torch.onnx export	0.0894	over HTTP; f32; throughput with 4 requests of batch 32 in flight
Triton, Linnet Python backend	2.76 ms	1,220/s	6.02 s	2.38 GiB		0.0156	over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 32 in flight