Releases·Tool Release
Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph
In a YouTube talk, Morph founder Tejas Bhakta (formerly Tesla inference optimization) describes how his team combined agent-written custom GPU kernels with bare-metal hardware tweaks to run models three times faster on cheaper GPUs. He argues GPU kernels suit autoresearch because correctness and speed are easy to verify, with humans supplying the ideas and agents doing the tuning.
35 engagement1 sourceSat, Sep 26, 2026, 03:00 PM
AI Overview
- Morph founder Tejas Bhakta says agent-written kernels plus bare-metal tweaks made models 3x faster on cheaper GPUs
- Bhakta's rule: humans supply the big ideas, like a DeepSeek attention step loading excess context; agents tune block sizes and parameters
- GPU kernels suit autoresearch because correctness and speed are easy to verify
- Agents need hardware and model context; reward hacking risks include disabling CUDA graphs to speed one kernel while slowing the whole model
- Bare-metal tweaks add about 25% over a cloud VM; kernel gains stack
- Bhakta warns roughly 80% of autoresearch attempts are bad
Related
Developments·Paper
GPT 5.6 Sol agents optimize Kimi K3-like model from 65 to 406 tok/s
A swarm of GPT 5.6 Sol agents spent over 40 hours optimizing a Kimi K3-like model, boosting inference speed from 65 to 406 tok/s. The agents discovered operator fusions, transformed the execution graph, and developed custom WebGPU kernels for Kimi Delta Attention, Multi-Head Latent Attention, and Mixture of Experts, reducing GPU dispatches from 331 to 22 per token.
A swarm of GPT 5.6 Sol agents spent over 40 hours optimizing a Kimi K3-like model, boosting inference speed from 65 to 406 tok/s. The agents discovered operator fusions, transformed the execution graph, and developed custom WebGPU kernels for Kimi Delta Attention, Multi-Head Latent Attention, and Mixture of Experts, reducing GPU dispatches from 331 to 22 per token.
59 engagement1 sourceredditFri, Jul 24, 2026, 04:41 PM
Developments·Paper
CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
A new paper introduces CUDA Agent, a large-scale agentic reinforcement learning approach for generating high-performance CUDA kernels. The method aims to improve LLMs' intrinsic CUDA optimization ability, which currently lags behind compiler-based systems like torch.compile.
A new paper introduces CUDA Agent, a large-scale agentic reinforcement learning approach for generating high-performance CUDA kernels. The method aims to improve LLMs' intrinsic CUDA optimization ability, which currently lags behind compiler-based systems like torch.compile.
2 engagement1 sourceredditTue, Aug 18, 2026, 06:33 AM
Developments·Paper
Harness Engineering for LLM-Driven GPU Kernel Generation
A paper presents a harness-centered system for LLM-driven GPU kernel optimization in the MLSys 2026 FlashInfer AI Kernel Generation Contest on NVIDIA Blackwell B200 GPUs. The system separates an evaluation harness from a profile-backed optimization controller to enforce compilation, correctness, and timing. This approach addresses the practical challenge of reliably constraining, validating, profiling, and selecting LLM-generated GPU kernels.
A paper presents a harness-centered system for LLM-driven GPU kernel optimization in the MLSys 2026 FlashInfer AI Kernel Generation Contest on NVIDIA Blackwell B200 GPUs. The system separates an evaluation harness from a profile-backed optimization controller to enforce compilation, correctness, and timing. This approach addresses the practical challenge of reliably constraining, validating, profiling, and selecting LLM-generated GPU kernels.
0 engagement1 sourcearxivMon, Jul 20, 2026, 02:14 PM
Releases·Tool Release
Together AI's ParallelKittens primitives target multi-GPU communication bottleneck
A YouTube talk by Simran Arora, who leads frontier performance research at Together AI, presents ParallelKittens, a small set of primitives for writing fast multi-GPU kernels. It addresses the widening gap between GPU compute and interconnect speeds, where standard PyTorch and NCCL baselines fall below 50% of communication-aware roofline.
A YouTube talk by Simran Arora, who leads frontier performance research at Together AI, presents ParallelKittens, a small set of primitives for writing fast multi-GPU kernels. It addresses the widening gap between GPU compute and interconnect speeds, where standard PyTorch and NCCL baselines fall below 50% of communication-aware roofline.
25 engagement1 sourceyoutubeThu, Aug 27, 2026, 05:00 PM