We found a benchmark blind spot when using cpp_extension.load() in our own NVFP4 suite on KernelArena. JIT compilation triggers CPU core migration, quietly inflating kernel launch times and deflating benchmark scores. Here's the investigation.
10 patterns we've tracked where LLMs game GPU kernel benchmarks, manipulating timers, returning garbage, caching results, and more, along with the defenses that catch them.
An open platform for benchmarking AI-generated GPU kernels — with initial results from WaferBench NVFP4 on B200 and KernelBench HIP on MI300X.
Get accurate 1:1 kernel mappings across hardware providers. Compare large vLLM traces in seconds with clean prefill vs. decode separation.
Give your AI coding assistant direct access to GPUs. No manual SSH setup, no Docker, or infrastructure management.
Cloud CUDA compilation with PTX/SASS output, PyTorch headers, and VS Code integration. No local CUDA install required.
How a non-kernel-expert achieved 8x speedup on latency-critical CUDA clustering code using profile-guided optimization with Wafer.
How adding profiling tools to our CLI helped an agent break through a theory-based optimization plateau, achieving 11.65x speedup on the Kimi Delta Attention kernel.
We used an AI agent to optimize AMD's topk_sigmoid kernel, achieving a 9x speedup over PyTorch. Here's exactly how our agent did it
How a fused kernel claiming 104x speedup passed our correctness checks while reading garbage memory, and the determinism check that catches it.
LLM-generated kernels are all the rage right now. We used frontier AI models to write HIP kernels for KernelBench and ran them on MI300Xs. Which ones performed the best?
Give your AI coding assistant direct access to GPU documentation, trace analysis, and remote kernel evaluation with the wafer-ai CLI.
The GPU documentation tool that thousands of engineers loved in our IDE extension is now available as a standalone web app.
Profile AMD GPUs directly in VS Code and Cursor. View hardware metrics, roofline analysis, and kernel stats — all without leaving your editor.
Open Chrome trace JSON files directly in your IDE with full Perfetto functionality — timeline, flamegraphs, SQL, and metrics.
Wafer is the GPU development stack that lives inside your editor: profiling (NCU), compiler explorer, and enhanced GPU docs.
As the AI hardware ecosystem rapidly expands, choosing the right accelerator has become increasingly complex. We're excited to introduce Chip Benchmark, an open-source benchmarking suite purpose-built to evaluate the performance of open-weight LLMs across diverse hardware platforms.
Large language models are driving a surge in inference workloads. While the AI community often gravitates toward more well-known GPUs, AMD's MI300X quietly stands out. Equipped with 192 GB of HBM3 and memory bandwidth of 5.3 TB/s, we explore how targeted optimization and quantization can unlock its potential.
