ESA Benchmarks — ESA Documentation

CURRENT SNAPSHOT

Benchmarks tied to the implementation that produced them.

These are selected internal/research ESA measurements already surfaced on the MLBricks homepage. They are not universal performance guarantees.

THROUGHPUT29.57Mtok/s · ESA training · 64K context

MEMORY1,325.3 MBESA reported peak · 64K context

PPL9.08best ESA perplexity · 16K context

DECODE6,580.16tok/s · ESA average over 512, 1K, 4K, and 8K

AT A GLANCE

ESA vs FA4 throughput + quality

Context	ESA throughput	FA4 throughput	Speedup	ESA PPL	FA4 PPL
1K	16.22M	4.38M	3.70×	9.50	12.09
16K	28.62M	7.06M	4.05×	9.08	11.88
64K	29.57M	2.67M	11.06×	9.18	11.95

ESA vs FA4 peak memory

Context	ESA peak	FA4 peak	Memory reduction
1K	182.4 MB	978.1 MB	81.4% lower
16K	454.5 MB	12,213.4 MB	96.3% lower
64K	1,325.3 MB	48,166.6 MB	97.2% lower

A focused comparison of ESA and FA4 using the measured rows from the current H100 benchmark report. Internal Thunder C16 Compile Reduce Overhead results are presented here as ESA, while FA4 uses the native FA4 baseline.

Method mapping:ESA= Thunder C16 Compile Reduce OH (compass 16);FA4= FA4 Native. Training results use NVIDIA H100, batch 64, FP16, and three-seed means. Throughput is million tokens per second; reported peak memory is MB.

CONTEXT 1,024

ESA vs FA4

NVIDIA H100 · training · batch 64 · FP16 · three-seed mean · parameters 555,520 · tokens trained 13.11M.

Method / runtime	Validation loss	PPL	Training throughput	Reported peak memory
ESA · C16 Compile Reduce OH	2.2502	9.50	16.22M	182.4 MB
FA4 · Native	2.4912	12.09	4.38M	978.1 MB

1K takeaway:ESA reaches 3.70× the FA4 training throughput and reports 81.4% lower peak memory, with PPL 9.50 versus 12.09 for FA4.

CONTEXT 16,384

ESA vs FA4

NVIDIA H100 · training · batch 64 · FP16 · three-seed mean · parameters 2,521,600 · tokens trained 209.72M.

Method / runtime	Validation loss	PPL	Training throughput	Reported peak memory
ESA · C16 Compile Reduce OH	2.2062	9.08	28.62M	454.5 MB
FA4 · Native	2.4747	11.88	7.06M	12,213.4 MB

16K takeaway:ESA reaches 4.05× the FA4 training throughput and reports 96.3% lower peak memory, with PPL 9.08 versus 11.88 for FA4.

CONTEXT 65,536

ESA vs FA4

NVIDIA H100 · training · batch 64 · FP16 · three-seed mean · parameters 8,813,056 · tokens trained 838.86M.

Method / runtime	Validation loss	PPL	Training throughput	Reported peak memory
ESA · C16 Compile Reduce OH	2.2166	9.18	29.57M	1,325.3 MB
FA4 · Native	2.4807	11.95	2.67M	48,166.6 MB

64K takeaway:ESA reaches 11.06× the FA4 training throughput and reports 97.2% lower peak memory, with PPL 9.18 versus 11.95 for FA4.

DECODER BENCHMARK

Decode throughput and recurrent-state memory

Prompt tokens	ESA Reduce OH tok/s	Attention KV tok/s	ESA speedup	ESA state	Attention KV memory
512	6,658.09	1,103.19	6.04×	0.000488 MB	0.750000 MB
1,024	6,574.91	1,118.56	5.88×	0.000488 MB	1.250000 MB
4,096	6,518.19	1,109.38	5.88×	0.000488 MB	4.250000 MB
8,192	6,569.44	1,062.37	6.18×	0.000488 MB	8.250000 MB

Decode takeaway:Across the four displayed prompt lengths (512, 1K, 4K, and 8K), ESA Compile Reduce Overhead averages6,580.16 tok/sversus1,098.38 tok/sfor native Attention KV — a5.99×ratio of average throughput. ESA recurrent state remains0.000488 MBacross the displayed prompt lengths.

INTERPRETATION

These values describe the displayed H100 training configurations only. Performance can change with model size, dataset, context, batch, precision, backend, compile mode, software version, and hardware. “Reported peak memory” should not be interpreted as total GPU memory unless the measurement scope is defined by the benchmark procedure.

← Examples ESA product →