NVIDIA H100 80GB HBM3 · median of 10 runs after 3 warm-ups · measured 2026-09-28T18:07:07+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 PyTorch2.27× fasterlatency

• Linnet, CUDA graphs0.74 mstransformers, compiled1.67 ms

vs KerasHub, in JAX5.05× fasterlatency

• Linnet, XLA (generated JAX)0.93 msKerasHub4.67 ms

vs torch.onnx, on ONNX Runtime1.36× fasterlatency

• Linnet, ONNX Runtime f321.52 mstorch.onnx export2.06 ms

vs torch.onnx, behind Triton1.13× fasterlatency

• Triton, Linnet's ONNX3.24 msTriton, torch.onnx export3.66 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)3.60 msthe referenceDistance from reference: 0 (the reference)unpadded batches of exactly 128 tokens (this card takes no padding mask)

transformers (torch.compile)1.67 ms2.2x the reference's speedDistance from reference: 0.109unpadded batches of exactly 128 tokens (this card takes no padding mask)

Linnet torch (generated source)2.79 ms0.60× the speed of transformers, compiledDistance from reference: 0.117unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)1.75 ms0.95× the speed of transformers, compiledDistance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (CUDA graphs)0.74 ms2.27× the speed of transformers, compiledDistance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

KerasHub (JAX)4.67 ms0.8x the reference's speedDistance from reference: 0.227bf16, under jax.jit; unpadded batches of exactly 128 tokens; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)1.56 ms3.01× the speed of KerasHubDistance from reference: 0.109unpadded batches of exactly 128 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)0.93 ms5.05× the speed of KerasHubDistance from reference: 0.25unpadded batches of exactly 128 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 Runtime2.06 ms1.8x the reference's speedDistance from reference: 0.114the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 128 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)1.52 ms1.36× the speed of torch.onnx exportDistance from reference: 0.115f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)1.25 ms2.9x the reference's speedDistance from reference: 0.113f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)1.25 ms2.9x the reference's speedDistance from reference: 0.125bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)0.74 ms4.9x the reference's speedDistance from reference: 0.112f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)0.53 ms6.8x the reference's speedDistance from reference: 0.113f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.55 ms6.6x the reference's speedDistance from reference: 0.109bf16; 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)3.66 ms1.0x the reference's speedDistance from reference: 0.114over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)3.24 ms1.13× the speed of Triton, torch.onnx exportDistance from reference: 0.115over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)2.71 ms1.3x the reference's speedDistance from reference: 0.141over 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)11,043/sthe referenceDistance from reference: 0 (the reference)unpadded batches of exactly 128 tokens (this card takes no padding mask)

transformers (torch.compile)14,663/s1.3x the reference's speedDistance from reference: 0.109unpadded batches of exactly 128 tokens (this card takes no padding mask)

Linnet torch (generated source)13,867/s0.95× the speed of transformers, compiledDistance from reference: 0.117unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)17,002/s1.16× the speed of transformers, compiledDistance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (CUDA graphs)17,770/s1.21× the speed of transformers, compiledDistance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

KerasHub (JAX)5,159/s0.5x the reference's speedDistance from reference: 0.227bf16, under jax.jit; unpadded batches of exactly 128 tokens; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)13,223/s2.56× the speed of KerasHubDistance from reference: 0.109unpadded batches of exactly 128 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)11,749/s2.28× the speed of KerasHubDistance from reference: 0.25unpadded batches of exactly 128 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 Runtime4,943/s0.4x the reference's speedDistance from reference: 0.114the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 128 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)5,274/s1.07× the speed of torch.onnx exportDistance from reference: 0.115f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)9,557/s0.9x the reference's speedDistance from reference: 0.113f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)9,728/s0.9x the reference's speedDistance from reference: 0.125bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)9,036/s0.8x the reference's speedDistance from reference: 0.112f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)19,916/s1.8x the reference's speedDistance from reference: 0.113f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)18,310/s1.7x the reference's speedDistance from reference: 0.109bf16; 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)579/s0.05x the reference's speedDistance from reference: 0.114over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)594/s1.03× the speed of Triton, torch.onnx exportDistance from reference: 0.115over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)558/s0.05x the reference's speedDistance from reference: 0.141over 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)1.44 sthe referenceDistance from reference: 0 (the reference)unpadded batches of exactly 128 tokens (this card takes no padding mask)

transformers (torch.compile)1.63 s0.9x the reference's speedDistance from reference: 0.109unpadded batches of exactly 128 tokens (this card takes no padding mask)

Linnet torch (generated source)1.09 s1.32× the speed of transformersDistance from reference: 0.117unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)1.11 s1.29× the speed of transformersDistance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (CUDA graphs)1.12 s1.29× the speed of transformersDistance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

KerasHub (JAX)6.31 s0.2x the reference's speedDistance from reference: 0.227bf16, under jax.jit; unpadded batches of exactly 128 tokens; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)0.52 s12× the speed of KerasHubDistance from reference: 0.109unpadded batches of exactly 128 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)0.58 s11× the speed of KerasHubDistance from reference: 0.25unpadded batches of exactly 128 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 Runtime17.5 s0.08x the reference's speedDistance from reference: 0.114the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 128 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)0.33 s53× the speed of torch.onnx exportDistance from reference: 0.115f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)0.56 s2.6x the reference's speedDistance from reference: 0.113f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)0.29 s5.0x the reference's speedDistance from reference: 0.125bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)0.28 s5.1x the reference's speedDistance from reference: 0.112f32; 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 s5.1x the reference's speedDistance from reference: 0.113f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)0.31 s4.7x the reference's speedDistance from reference: 0.109bf16; 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)9.10 s0.2x the reference's speedDistance from reference: 0.114over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)16.1 s0.57× the speed of Triton, torch.onnx exportDistance from reference: 0.115over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)6.02 s0.2x the reference's speedDistance from reference: 0.141over 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)1.13 GiB (1,162 MiB)Distance from reference: 0 (the reference)unpadded batches of exactly 128 tokens (this card takes no padding mask)

transformers (torch.compile)1.12 GiB (1,142 MiB)Distance from reference: 0.109unpadded batches of exactly 128 tokens (this card takes no padding mask)

Linnet torch (generated source)1.08 GiB (1,108 MiB)Distance from reference: 0.117unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (torch.compile, inductor)1.04 GiB (1,060 MiB)Distance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

Linnet torch (CUDA graphs)1.12 GiB (1,148 MiB)Distance from reference: 0.141unpadded batches of exactly 128 tokens; bf16 on cuda

KerasHub (JAX)0.51 GiB (525 MiB)Distance from reference: 0.227bf16, under jax.jit; unpadded batches of exactly 128 tokens; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions

Linnet JAX (XLA, StableHLO)0.72 GiB (734 MiB)Distance from reference: 0.109unpadded batches of exactly 128 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)0.71 GiB (727 MiB)Distance from reference: 0.25unpadded batches of exactly 128 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 Runtime3.30 GiB (3,382 MiB)Distance from reference: 0.114the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 128 tokens

Linnet ONNX f32 -> ONNX Runtime (CUDA)2.53 GiB (2,586 MiB)Distance from reference: 0.115f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (CUDA)1.71 GiB (1,748 MiB)Distance from reference: 0.113f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (CUDA)1.71 GiB (1,748 MiB)Distance from reference: 0.125bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f32 -> ONNX Runtime (TensorRT)3.21 GiB (3,282 MiB)Distance from reference: 0.112f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX f16 -> ONNX Runtime (TensorRT)2.97 GiB (3,042 MiB)Distance from reference: 0.113f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length

Linnet ONNX bf16 -> ONNX Runtime (TensorRT)2.97 GiB (3,038 MiB)Distance from reference: 0.109bf16; 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)2.37 GiB (2,429 MiB)Distance from reference: 0.114over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (ONNX Runtime backend, Linnet ONNX)2.37 GiB (2,425 MiB)Distance from reference: 0.115over HTTP; f32; throughput with 4 requests of batch 64 in flight

Triton Inference Server (Python backend, Linnet torch)2.81 GiB (2,880 MiB)Distance from reference: 0.141over 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	3.60 ms	11,043/s	1.44 s	1.13 GiB		0 (the reference)	unpadded batches of exactly 128 tokens (this card takes no padding mask)
transformers, compiled	1.67 ms	14,663/s	1.63 s	1.12 GiB		0.109	unpadded batches of exactly 128 tokens (this card takes no padding mask)
Linnet, generated PyTorch	2.79 ms	13,867/s	1.09 s	1.08 GiB	0.60× transformers, compiled	0.117	unpadded batches of exactly 128 tokens; bf16 on cuda
Linnet, inductor	1.75 ms	17,002/s	1.11 s	1.04 GiB	0.95× transformers, compiled	0.141	unpadded batches of exactly 128 tokens; bf16 on cuda
Linnet, CUDA graphs	0.74 ms	17,770/s	1.12 s	1.12 GiB	2.27× transformers, compiled	0.141	unpadded batches of exactly 128 tokens; bf16 on cuda
JAX
KerasHub	4.67 ms	5,159/s	6.31 s	0.51 GiB		0.227	bf16, under jax.jit; unpadded batches of exactly 128 tokens; peak memory is JAX's allocator peak; the driver shows its pool, which grows in whole regions
Linnet, XLA (StableHLO)	1.56 ms	13,223/s	0.52 s	0.72 GiB	3.01× KerasHub	0.109	unpadded batches of exactly 128 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)	0.93 ms	11,749/s	0.58 s	0.71 GiB	5.05× KerasHub	0.25	unpadded batches of exactly 128 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	2.06 ms	4,943/s	17.5 s	3.30 GiB		0.114	the reference model's own ONNX export, f32 like Linnet's; unpadded batches of exactly 128 tokens
Linnet, ONNX Runtime f32	1.52 ms	5,274/s	0.33 s	2.53 GiB	1.36× torch.onnx export	0.115	f32; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, ONNX Runtime f16	1.25 ms	9,557/s	0.56 s	1.71 GiB		0.113	f16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, ONNX Runtime bf16	1.25 ms	9,728/s	0.29 s	1.71 GiB		0.125	bf16; ONNX Runtime's CUDA execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, TensorRT f32	0.74 ms	9,036/s	0.28 s	3.21 GiB		0.112	f32; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, TensorRT f16	0.53 ms	19,916/s	0.28 s	2.97 GiB		0.113	f16; ONNX Runtime's TensorRT execution provider; first calls build the sessions; unpadded batches of exactly the sequence length
Linnet, TensorRT bf16	0.55 ms	18,310/s	0.31 s	2.97 GiB		0.109	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	3.66 ms	579/s	9.10 s	2.37 GiB		0.114	over HTTP; f32; throughput with 4 requests of batch 64 in flight
Triton, Linnet's ONNX	3.24 ms	594/s	16.1 s	2.37 GiB	1.13× Triton, torch.onnx export	0.115	over HTTP; f32; throughput with 4 requests of batch 64 in flight
Triton, Linnet Python backend	2.71 ms	558/s	6.02 s	2.81 GiB		0.141	over HTTP; bf16, CUDA graphs; throughput with 4 requests of batch 64 in flight