Intel Gave Up on Its Gaudi Chips. We Used AI to Make Them Outperform NVIDIA H100. – Standard Kernel Co.
Standard Kernel researchers optimized Intel Gaudi2 software for Qwen3.5-27B inference, achieving a 60% higher decode throughput and 8% higher cached-prefill throughput than an NVIDIA H100 80GB PCIe on a long-context benchmark. Because Gaudi2 hardware costs significantly less to rent, the optimized software configuration delivered 9.4x more output tokens per dollar than the H100. The authors argue that comprehensive software optimization can make alternative AI accelerators highly competitive without altering the underlying silicon. Open article →