Kwker benchmark results: text embeddings (KwkEncoder vs OpenVINO and torch.compile)
====================================================================================

Published on https://kwker.io/benchmarks/#enc. Raw numbers as recorded.

Machine     Intel Xeon with AMX (bf16 / int8), Emerald Rapids class cloud VM, 4 cores used by every column
Date        October 2026
OpenVINO    2026.4, bf16 inference precision hint (int8: NNCF post-training quantization, transformer preset,
            8 calibration batches of the same shape)
PyTorch     torch.compile(backend="kwker") in its "medium" precision mode (bf16 matrix products, float32 elsewhere)
Kwker       KwkEncoder (kwker.encode): bf16 operands on AMX tiles, float32 accumulation, LayerNorm / softmax / GELU in
            float32; padded batches skip the padding (unpad=True, the default)
Protocol    interleaved rounds of every column, best round per column, milliseconds per batch. Batched rows are
            sentences of L/2..L tokens padded to L. Output error against float32 eager PyTorch, relative to the largest
            output: OpenVINO bf16 5e-3, KwkEncoder 5e-3 (MiniLM: 1.7e-3 both).

Random weights, bf16
                        OpenVINO bf16   torch.compile(backend="kwker") bf16   KwkEncoder bf16   vs OpenVINO bf16
BERT-base 1 x 128       17.6 ms         18.6 ms                               10.7 ms           1.64x
BERT-base 8 x 128       114 ms          102 ms                                65.8 ms           1.74x
BERT-base 1 x 512       93.1 ms         85.2 ms                               44.4 ms           2.10x
MiniLM-L6 1 x 128       3.52 ms         4.74 ms                               1.88 ms           1.88x
MiniLM-L6 32 x 64       33.3 ms         37.7 ms                               26.8 ms           1.24x

Without skipping the padding (unpad=False), same run: BERT-base 8 x 128 79.7 ms (61.1 ms with), MiniLM 32 x 64 33.4 ms
(23.4 ms with).

Real checkpoints (bert-base-uncased, all-MiniLM-L6-v2), bf16 and int8; error in parentheses
                        OpenVINO bf16       OpenVINO int8       KwkEncoder bf16     KwkEncoder int8     int8 vs OpenVINO int8
BERT-base 1 x 128       18.3 ms (1.1e-2)    12.1 ms (7.9e-2)    10.9 ms (8.7e-3)    9.25 ms (1.4e-1)    1.31x
BERT-base 8 x 128       96.3 ms             72.7 ms (1.3e-1)    57.6 ms             42.2 ms (2.7e-1)    1.72x
MiniLM-L6 1 x 128       3.73 ms             3.20 ms (5.0e-2)    2.03 ms             1.49 ms (3.0e-2)    2.15x
MiniLM-L6 32 x 64       28.0 ms             30.9 ms (1.1e-1)    20.7 ms             15.9 ms (5.3e-2)    1.94x

KwkEncoder bf16 alone is faster than OpenVINO int8 in every row. Uncalibrated int8 has a larger error than OpenVINO's
on BERT-base (outlier channels in its LayerNorm and GELU outputs); check accuracy on your own data before using int8.

This VM's AMX throughput dips for a few milliseconds at random times (another tenant on the core): single timings vary,
best-of rounds are stable.