Hot Chips: @OpenAI's Jalapeño AI chip is much more impressive than I expected. The first real benchmarks are out, and the Gen 1 inference chip beats Nvidia's GB200/GB300 Blackwell systems across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T: • 1.5-1.9x more performance per watt • 1.7-3.6x lower end-to-end latency • 2.1-4.1x higher performance in highly interactive workloads On DeepSeek R1, Jalapeño reaches 700 tok/s/user vs 169 on GB300. A 128-chip rack delivers 1.7 EFLOPS of MXFP4 compute with 27.5 TB of HBM4. Maybe even more interesting: OpenAI used its own AI to design and program the chip. Codex + GPT-Astra brought three new models to high performance in just two months, while AI-generated kernels beat expert-written kernels by 1.5-1.8x on selected blocks. And this is only Gen 1. Gen 2 is already deep in development, with Gen 3 taking shape.

Hot Chips: @OpenAI's Jalapeño AI chip is much more impressive than I expected. The first real benchmarks are out, and the Gen 1 inference chip beats Nvidia's GB200/GB300 Blackwell systems across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T: • 1.5-1.9x more performance per watt • 1.7-3.6x lower end-to-end latency • 2.1-4.1x higher performance in highly interactive workloads On DeepSeek R1, Jalapeño reaches 700 tok/s/user vs 169 on GB300. A 128-chip rack delivers 1.7 EFLOPS of MXFP4 compute with 27.5 TB of HBM4. Maybe even more interesting: OpenAI used its own AI to design and program the chip. Codex + GPT-Astra brought three new models to high performance in just two months, while AI-generated kernels beat expert-written kernels by 1.5-1.8x on selected blocks. And this is only Gen 1. Gen 2 is already deep in development, with Gen 3 taking shape.

Source:https://x.com/zerohedge/status/2092307516482802124