Blonskr, founder of TapeOut, is back in school. V2 and the contract audit are on the way. English-speaking developers, researchers, and users should have their own TG group. He has class. The English community is under community care for now — which, this week, means me. Special assignment: check his homework every day. Coding only. Come in. Help us keep him honest. 👉 #TapeOut# 🦾

Hot Chips. Day 2. Another hot day 🔥🌶️ Monday they sold factories. Tuesday OpenAI walked in with one. Nine months. RTL to tapeout. ChatGPT is already running on it. 1,459 tokens a second. GB200 did 535. At half the watts. NVIDIA named the factory OS. 200 versus 7,200. Cerebras said 30x. Google stuffed 9,600 chips in a pod. I called this the Agentic CPU in March. They stopped selling chips. They sell factories.

people in software know so little about how chips get made example: emulators are room-sized fpga machines from Cadence/Synopsys that run the chip much faster than software sim, at real-ish speeds and boot actual operating systems so you can find more bugs before tapeout is there good content about this? should i make content about the end to end chip making process? people are so excited about alphachip etc. (cool work!) for ai floorplanning but it’s a relatively small fraction of the full workflow. a lot more to do

Hot, hot chips, hot hot chip Met Mr jalapeño Richard Ho OpenAI just printed Jalapeño. 216 gigs. 15.4 TB/s. 700 watts. Inference. NVIDIA still trains the factory. They keep buying it. Nine months, RTL to tapeout. More at BEP Research

OpenAI just revealed the first results from Jalapeño, its first custom AI inference chip, and the gains are huge. Across big models including DeepSeek R1 670B and Kimi K2.5 1T, it delivers 1.5-1.9x more AI work per watt, 1.7–3.6x lower end to end latency, and up to 4.1x higher performance for highly interactive workloads versus comparison systems. AI helped design the chip, enabling OpenAI to go from initial design to tapeout in only 9 months. AI generated implementations for selected workloads were already 1.5–1.8x faster than human expert written versions. AI designs better chips → better chips run AI faster and cheaper → stronger AI designs even better chips → repeat