NVIDIA H100 80GB HBM3 · median of 10 runs after 3 warm-ups · measured 2026-09-28T18:30:34+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 PyTorch2.28× fasterlatency

• Linnet, CUDA graphs1.01 mstransformers, compiled2.31 ms

vs torch.onnx, on ONNX Runtime0.99× slowerlatency

• Linnet, ONNX Runtime f323.63 mstorch.onnx export3.59 ms

vs torch.onnx, behind Triton1.03× fasterlatency

• Triton, Linnet's ONNX6.60 msTriton, torch.onnx export6.79 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)9.19 msthe referenceDistance from reference: 0 (the reference)unpadded batches of exactly 256 tokens (this card takes no padding mask)

transformers (torch.compile)2.31 ms4.0x the reference's speedDistance from reference: 2.13unpadded batches of exactly 256 tokens (this card takes no padding mask)

Linnet torch (generated source)5.78 ms0.40× the speed of transformers, compiledDistance from reference: 1.63unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)2.15 ms1.07× the speed of transformers, compiledDistance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (CUDA graphs)1.01 ms2.28× the speed of transformers, compiledDistance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet JAX (XLA, StableHLO)1.99 ms4.6x the reference's speedDistance from reference: 1.47unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)1.81 ms5.1x the reference's speedDistance from reference: 2.21unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime3.59 ms2.6x the reference's speedDistance from reference: 2.34the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 256 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)3.63 ms0.99× the speed of torch.onnx exportDistance from reference: 2.39f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)2.79 ms3.3x the reference's speedDistance from reference: 2.19f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)3.24 ms2.8x the reference's speedDistance from reference: 2.25bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)1.71 ms5.4x the reference's speedDistance from reference: 2.35f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)1.25 ms7.3x the reference's speedDistance from reference: 2.16f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)1.36 ms6.8x the reference's speedDistance from reference: 1.94bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Triton Inference Server (ONNX Runtime backend, torch.onnx export)6.79 ms1.4x the reference's speedDistance from reference: 2.34over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)6.60 ms1.03× the speed of Triton, torch.onnx exportDistance from reference: 2.39over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)3.97 ms2.3x the reference's speedDistance from reference: 1.25over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 64 in flight

Throughput

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

transformers (eager)1,358/sthe referenceDistance from reference: 0 (the reference)unpadded batches of exactly 256 tokens (this card takes no padding mask)

transformers (torch.compile)4,488/s3.3x the reference's speedDistance from reference: 2.13unpadded batches of exactly 256 tokens (this card takes no padding mask)

Linnet torch (generated source)2,780/s0.62× the speed of transformers, compiledDistance from reference: 1.63unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)6,042/s1.35× the speed of transformers, compiledDistance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (CUDA graphs)6,153/s1.37× the speed of transformers, compiledDistance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet JAX (XLA, StableHLO)3,252/s2.4x the reference's speedDistance from reference: 1.47unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)3,506/s2.6x the reference's speedDistance from reference: 2.21unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime1,302/s1.0x the reference's speedDistance from reference: 2.34the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 256 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)1,347/s1.03× the speed of torch.onnx exportDistance from reference: 2.39f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)2,096/s1.5x the reference's speedDistance from reference: 2.19f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)2,040/s1.5x the reference's speedDistance from reference: 2.25bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)2,035/s1.5x the reference's speedDistance from reference: 2.35f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)5,759/s4.2x the reference's speedDistance from reference: 2.16f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)5,456/s4.0x the reference's speedDistance from reference: 1.94bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Triton Inference Server (ONNX Runtime backend, torch.onnx export)267/s0.2x the reference's speedDistance from reference: 2.34over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)263/s0.99× the speed of Triton, torch.onnx exportDistance from reference: 2.39over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)270/s0.2x the reference's speedDistance from reference: 1.25over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 64 in flight

Load time

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

transformers (eager)3.18 sthe referenceDistance from reference: 0 (the reference)unpadded batches of exactly 256 tokens (this card takes no padding mask)

transformers (torch.compile)1.39 s2.3x the reference's speedDistance from reference: 2.13unpadded batches of exactly 256 tokens (this card takes no padding mask)

Linnet torch (generated source)1.39 s1.00× the speed of transformers, compiledDistance from reference: 1.63unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)1.03 s1.36× the speed of transformers, compiledDistance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (CUDA graphs)1.36 s1.03× the speed of transformers, compiledDistance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet JAX (XLA, StableHLO)0.58 s5.5x the reference's speedDistance from reference: 1.47unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)0.63 s5.1x the reference's speedDistance from reference: 2.21unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime31.8 s0.10x the reference's speedDistance from reference: 2.34the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 256 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)0.29 s110× the speed of torch.onnx exportDistance from reference: 2.39f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)0.30 s11x the reference's speedDistance from reference: 2.19f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)0.28 s11x the reference's speedDistance from reference: 2.25bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)0.30 s11x the reference's speedDistance from reference: 2.35f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.28 s11x the reference's speedDistance from reference: 2.16f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.28 s11x the reference's speedDistance from reference: 1.94bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Triton Inference Server (ONNX Runtime backend, torch.onnx export)12.0 s0.3x the reference's speedDistance from reference: 2.34over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)18.1 s0.66× the speed of Triton, torch.onnx exportDistance from reference: 2.39over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)6.02 s0.5x the reference's speedDistance from reference: 1.25over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 64 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)1.88 GiB (1,922 MiB)Distance from reference: 0 (the reference)unpadded batches of exactly 256 tokens (this card takes no padding mask)

transformers (torch.compile)1.34 GiB (1,374 MiB)Distance from reference: 2.13unpadded batches of exactly 256 tokens (this card takes no padding mask)

Linnet torch (generated source)1.27 GiB (1,304 MiB)Distance from reference: 1.63unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)1.28 GiB (1,306 MiB)Distance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet torch (CUDA graphs)1.41 GiB (1,442 MiB)Distance from reference: 1.25unpadded batches of exactly 256 tokens; bf16 on cuda

Linnet JAX (XLA, StableHLO)0.91 GiB (927 MiB)Distance from reference: 1.47unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, generated source)1.21 GiB (1,234 MiB)Distance from reference: 2.21unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

transformers -> torch.onnx -> ONNX Runtime6.80 GiB (6,960 MiB)Distance from reference: 2.34the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 256 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)4.71 GiB (4,822 MiB)Distance from reference: 2.39f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)2.75 GiB (2,812 MiB)Distance from reference: 2.19f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)2.75 GiB (2,812 MiB)Distance from reference: 2.25bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)4.02 GiB (4,116 MiB)Distance from reference: 2.35f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)3.68 GiB (3,768 MiB)Distance from reference: 2.16f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)3.68 GiB (3,770 MiB)Distance from reference: 1.94bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Triton Inference Server (ONNX Runtime backend, torch.onnx export)6.94 GiB (7,105 MiB)Distance from reference: 2.34over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)6.90 GiB (7,065 MiB)Distance from reference: 2.39over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)3.13 GiB (3,202 MiB)Distance from reference: 1.25over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 64 in flight

Every number

Method	Latency	Throughput	Load time	Peak GPU memory	Against the stack it replaces	Distance from reference	Notes
PyTorch
transformers	9.19 ms	1,358/s	3.18 s	1.88 GiB		0 (the reference)	unpadded batches of exactly 256 tokens (this card takes no padding mask)
transformers, compiled	2.31 ms	4,488/s	1.39 s	1.34 GiB		2.13	unpadded batches of exactly 256 tokens (this card takes no padding mask)
Linnet, generated PyTorch	5.78 ms	2,780/s	1.39 s	1.27 GiB	0.40× transformers, compiled	1.63	unpadded batches of exactly 256 tokens; bf16 on cuda
Linnet, inductor	2.15 ms	6,042/s	1.03 s	1.28 GiB	1.07× transformers, compiled	1.25	unpadded batches of exactly 256 tokens; bf16 on cuda
Linnet, CUDA graphs	1.01 ms	6,153/s	1.36 s	1.41 GiB	2.28× transformers, compiled	1.25	unpadded batches of exactly 256 tokens; bf16 on cuda
JAX
Linnet, XLA (StableHLO)	1.99 ms	3,252/s	0.58 s	0.91 GiB		1.47	unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions
Linnet, XLA (generated JAX)	1.81 ms	3,506/s	0.63 s	1.21 GiB		2.21	unpadded batches of exactly 256 tokens; bf16 on cuda; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions
ONNX Runtime and TensorRT
torch.onnx export	3.59 ms	1,302/s	31.8 s	6.80 GiB		2.34	the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 256 tokens
Linnet, ONNX Runtime f32	3.63 ms	1,347/s	0.29 s	4.71 GiB	0.99× torch.onnx export	2.39	f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, ONNX Runtime f16	2.79 ms	2,096/s	0.30 s	2.75 GiB		2.19	f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, ONNX Runtime bf16	3.24 ms	2,040/s	0.28 s	2.75 GiB		2.25	bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, TensorRT f32	1.71 ms	2,035/s	0.30 s	4.02 GiB		2.35	f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, TensorRT f16	1.25 ms	5,759/s	0.28 s	3.68 GiB		2.16	f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, TensorRT bf16	1.36 ms	5,456/s	0.28 s	3.68 GiB		1.94	bf16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Triton Inference Server
Triton, torch.onnx export	6.79 ms	267/s	12.0 s	6.94 GiB		2.34	over HTTP; f32; throughput with 4 requests of batch 64 in flight
Triton, Linnet's ONNX	6.60 ms	263/s	18.1 s	6.90 GiB	1.03× Triton, torch.onnx export	2.39	over HTTP; f32; throughput with 4 requests of batch 64 in flight
Triton, Linnet Python backend	3.97 ms	270/s	6.02 s	3.13 GiB		1.25	over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 64 in flight