Sign InOpen Brain
AI EngineerVideoSource Linked

Can LLMs Write Fast Multi-GPU Kernels? — Simran Arora, Together AI

A multi-GPU kernel benchmark shows coding agents can improve with tools and iteration, but still struggle with communication scheduling, partitioning, and transfer choices.

AI Engineer · Aug 27, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Parallel Kernel Bench asks models to rewrite distributed PyTorch references as performant multi-GPU CUDA kernels across **87 problems**. A Gemini 3 Pro agent with a local shell improved from **24 to 35 solved tasks**.

Practical Implication

Treat tool access and iterative compilation as useful but insufficient. Only **26 tasks exceeded the reference speed**, and models still struggled with collective ordering, data partitioning, scheduling, and choosing transfer mechanisms.

Agent-Ready Context
Parallel Kernel Bench asks models to rewrite distributed PyTorch references as performant multi-GPU CUDA kernels across **87 problems**. A Gemini 3 Pro agent with a local shell improved from **24 to 35 solved tasks**.

Treat tool access and iterative compilation as useful but insufficient. Only **26 tasks exceeded the reference speed**, and models still struggled with collective ordering, data partitioning, scheduling, and choosing transfer mechanisms.

Performance plateaued as more time was added, even when the relevant kernel principles were supplied in context. The results cover this benchmark and harness, so they do not establish performance for every model, topology, or production workload.
Connected Context · Feed7 Judgment

Parallel Kernel Bench isolates a demanding boundary for coding agents: shell access and iterative compilation raise solved tasks, but producing kernels that outperform references remains substantially harder, and extra time or supplied principles does not prevent a plateau. This reinforces correctness-plus-performance validation and narrows conclusions to one benchmark, agent, harness, and distributed-kernel workload.

Can Large Language Models Recover Semantic Optimization Opportunities That Compilers Miss?Both show that generating an optimization is not enough: artifacts require independent correctness and performance checks, and successful optimization on controlled tasks does not imply dependable production improvement.State of Data — Sean Cai, Independent / State of DataThe benchmark’s gains are explicitly tied to a local-shell iterative harness, reinforcing the candidate’s warning that observed agent performance can shift with scaffolding and should be tested across harnesses.Benchmarking Coding Agents on New vs Legacy Codebases — Denys Linkov, WisedocsBoth distinguish plausible or completed output from accepted success: Parallel Kernel Bench checks distributed correctness and speed, while the candidate calls for end-to-end acceptance criteria in repository work.When Will The Benchmaxxing Plague End? — Nick Heiner, Surge AIThe benchmark reports separate solved-task and reference-speed outcomes and limits its scope, aligning with the candidate’s demand that verifier conditions and task quality be visible before treating a score as general capability.
Context Map
benchmarkcoding#agent-evals#benchmark-integrity#coding-agents
Uncertainty
Performance plateaued as more time was added, even when the relevant kernel principles were supplied in context. The results cover this benchmark and harness, so they do not establish performance for every model, topology, or production workload.