Google’s JAXBench benchmark lifts TPU kernel optimization with 50 real workloads
omarsar0 · x · 2026-07-27
Google’s JAXBench gives TPUs a real benchmark for autonomous kernel optimization
Google, Harvard, and UC Berkeley introduce JAXBench, a TPU-native benchmark suite for autonomous kernel optimization. The suite contains 50 real JAX workloads drawn from production-style architectures such as Llama-3.1, DeepSeek-V3, Mixtral, Mamba-2, and AlphaFold2.
Highlights:
- 17 operators are translated from KernelBench and validated for correctness.
- 8 more ship with hand-tuned Pallas kernels from Tokamax as expert baselines.
- With Gemini 3 Flash, curated TPU documentation improves per-sample correctness from 5.8% to 37.3%.
- The system solves 48 of 50 benchmarks at a 1.28x geomean speedup.
- Beam search further raises performance to 1.36x.
The paper’s main takeaway is that correctness is heavily documentation-dependent, while speed is largely a search problem. The benchmark, harness, and baseline results are released for open-source contribution.
More from Infra
- fmgo: call Apple's on-device Foundation Models from Go with no CGO and no Swift — Super_Run_8466 · 2026-09-23
- Huawei unveils Peerium architecture: nested BSP unifies million processors into one computer — Dr_Singularity · 2026-09-23
- Grok explains why DeepSeek picked DualPipe + ZeRO-1 over ZeRO-3 on 2048 H800s — TheZachMueller · 2026-09-23
- AI costs fall 47% per quarter, 4x faster than DNA sequencing: Epoch AI — daveholtz · 2026-09-23
- M5 Ultra LLM test: 4x faster prompt processing, but double the power draw — DigitalguyCH · 2026-09-23
- $500 of Dell OptiPlexes become a diskless netboot lab where AI agents can't brick the hardware — colinmcnamara · 2026-09-23