Pearl Research

Pearl builds in the open. Explore our whitepaper, foundational research, engineering notes from the serving stack, and the open problems we’re still working to solve.

AI Research

Optimizing Grouped GEMM for MoE on Blackwell
How we squeezed up to 15% more performance out of MoE's core kernel.
Read the article
Hawkeye: Reproducing GPU-Level Non-Determinism
Reproduces the exact bit-level matrix multiplication an NVIDIA GPU ran by re-executing it on a CPU with no precision loss, so an auditor can check any training or inference run against a deterministic reference at essentially no overhead.
Read on arXiv
A Fused Blockwise Hadamard Quantization Kernel in CuTe DSL
How Pearl quantizes activations on the inference hot path: a fused CuTe DSL kernel that folds the blockwise Hadamard transform into INT quantization, benchmarked against vLLM and running in production behind models like Gemma 4. Much of it was written by our in-house agentic kernel tool.
Read on X
400x Faster Hashing on GPUs
Pearl mining has to hash large tensors that already live on the GPU, so hashing on the host would add CPU-GPU round trips to every matrix multiplication. This article walks through building a cryptographically secure BLAKE3 kernel that optimizes a single digest on-device with no host I/O, instead of the usual GPU approach of just computing many hashes at once. It covers the design choices from picking the hash to the kernel architecture: chunk compression fully unrolled in registers, fused with the Merkle tree reduction, down to which PTX instructions move the data.
Read on X

The Pearl Protocol

Pearl Research · Pearl Research Labs