DeepGEMM vs Karpathy Skills

DeepGEMM is DeepSeek's open-source library of highly optimized FP8 General Matrix Multiplication (GEMM) kernels targeting NVIDIA SM90/SM100 GPUs — the H100, H800, and Blackwell class. The headline feature is a lightweight just-in-time (JIT) compiler that eliminates the need for offline CUDA compilation at install time, dramatically lowering the barrier for teams who want raw GPU throughput without complex build pipelines. The library covers FP8 and FP4 dense GEMMs, BF16 accumulation, grouped GEMMs for Mixture-of-Experts architectures with overlapped NVLink communication, and multi-query attention scoring kernels. On H800 hardware DeepGEMM posts up to 1,550 TFLOPS — competitive with hand-tuned vendor libraries — while remaining fully open source under the MIT license. For LLM inference teams running on H100/H800 clusters, DeepGEMM slots directly into inference stacks like vLLM and SGLang. It's especially notable because it came from DeepSeek's internal training infrastructure, meaning it's been battle-tested at the scale that produced some of 2026's most cost-efficient models. This isn't research code — it's production tooling going public.

Karpathy Skills is a CLAUDE.md plugin distilled from Andrej Karpathy's public observations on LLM coding pitfalls. Drop the single file into your project root (or install it as a Claude Code skill) and every Claude Code session starts pre-loaded with the four principles Karpathy identified as most commonly violated: think before writing, prefer simplicity, make only targeted changes, and close loops with explicit verification. The project has accumulated 1,450+ GitHub stars in under two weeks. The implementation is intentionally minimal — it's a structured system prompt, not a framework. Each principle is spelled out with concrete anti-patterns to avoid: no premature generation, no over-engineering simple tasks, no cascading refactors when a surgical fix suffices, no ending a session without verifying the goal was actually met. It's Karpathy's "Software 2.0" thinking applied to the agent workflow meta-layer. What makes this compelling isn't the technology — it's the curation. Karpathy has spent more time thinking about LLM behavior patterns than almost anyone outside the major labs. Packaging that into something installable in 30 seconds lowers the floor for teams who want more reliable agent outputs without extensive prompt engineering work.