Jalapeño Benchmarks: OpenAI’s Inference Chip Outruns Blackwell on Paper, Ships at Scale in 2027
OpenAI's new inference chip, codenamed Jalapeño, shows promising performance on paper, outperforming Nvidia's Blackwell in benchmarks for tokens per user and throughput per kilowatt. However, the chip is not expected to ship in significant quantities until 2027, raising questions about its competitiveness against future hardware. OpenAI's head of hardware, Richard Ho, highlighted the chip's efficiency and low latency, attributing its design to minimizing data movement and local KV cache retention.
- Entities
- Topic: AI
- Topic: Inference Chips
- Topic: Hardware
- Company: OpenAI
- Company: Broadcom
- Person: Richard Ho
- Company: NVIDIA
This article was reported and published by egamers.io. feed.gg links to it as part of a story cluster — full text, images and rights remain with the publisher.