Les performances annoncées dépassent la génération Blackwell actuelle, mais le calendrier réel retarde le déploiement significatif à 2027. D'ici cette date, la concurrence aura elle aussi fait évoluer ses architectures.
OpenAI cherche à briser la dépendance exclusive aux processeurs Nvidia en créant son propre silicium d'inférence. Le contrôle direct du matériel et du matériel mémoire vise à réduire drastiquement les coûts opérationnels et la latence d'exécution de ses modèles à grande échelle.
Présentée par Richard Ho lors de la conférence Hot Chips, la puce Jalapeño surpasse l'actuelle architecture Nvidia Blackwell sur le benchmark InferenceX en débit par kilowatt et en jetons par utilisateur. Développée conjointement avec Broadcom et assistée par les modèles d'OpenAI, cette plateforme matérielle intègre le calcul et la mémoire pour éliminer les goulots d'étranglement réseau lors des phases de pré-remplissage et de gestion du cache KV.
Impact concret
Les directeurs informatiques et responsables d'infrastructure doivent prendre en compte l'émergence de puces propriétaires d'inférence dans leurs projections budgétaires à long terme, tout en maintenant leurs déploiements actuels sur les GPU existants.
Applications concrètes
Traitement à très faible latence des requêtes pour les agents virtuels en temps réel et conservation locale du cache KV pour les modèles gérant de très longs contextes.
At the Hot Chips conference on Tuesday, OpenAI shared a more detailed look at Jalapeño , including the first batch of benchmark results for the new system. Tested on SemiAnalysis’ InferenceX benchmark, Jalapeño registered both more tokens per user and more throughput per kilowatt than the currently available state-of-the-art inference processors. “The bottom line is that the results show a very, very significant performance advance over state of the art,” said Richard Ho, OpenAI’s head of hardware, in a press call. “Jalapeño can serve more AI work per unit of power, while also returning responses more quickly. It’s very efficient to serve a lot of customers, but it can also be very low latency.” Notably, that comparison is against an Nvidia Blackwell system — but by the time Jalapeño reaches full deployment, the competition may have advanced significantly. Ho estimated that Jalapeño would deploy at the end of 2026 “in very small volumes,” with more significant deployment coming in 2027. First announced last October, Jalapeño was developed by OpenAI in close collaboration with Broadcom, with OpenAI’s own models assisting in the development process. The company plans to make Jalapeño a multigenerational platform, allowing AI products, models, chips, and memory all developed in concert. Because of that full-stack approach, OpenAI was able to address specific phases in the inference process that often cause friction during inference processing. In particular, Jalapeño is designed to minimize delays during the prefill and communication phases of processing, which OpenAI says often act as bottlenecks. “We designed Jalapeño to minimize data movement and communication delays,” the company said in a blog post presenting the results. “This means that model state, including the KV cache used while generating a response, can be explicitly placed and kept local while the system activates the right combination of compute, memory, and networking for each inference phase.” Topics AI , Broadcom , OpenAI When you purchase through links in our articles, we may earn a small commission . This doesn’t affect our editorial independence. Russell Brandom AI Editor Russell Brandom has been covering the tech industry since 2012, with a focus on platform policy and emerging technologies. He previously worked at The Verge and Rest of World, and has written for Wired, The Awl and MIT’s Technology Review. He can be reached at russell.brandom@techcrunch.com or on Signal at 412-401-5489. View Bio October 13 – 15 San Francisco In less than 48 hours, your chance to save up to $300 on your tickets will end! REGISTER NOW Most Popular Two years after launch, Walmart’s Flipkart is closing in on India’s quick-commerce leaders Jagmeet Singh Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research Anna Heim Michael Polansky is training an AI model on skin that’s still alive Connie Loizos How AI accounting startup Rillet raised $100M and became a unicorn in 48 hours Dominic-Madori Davis Tesla’s solar roof is dead — here’s what went wrong Tim De Chant Oura faces lawsuit accusing it of misleading consumers about sleep-tracking accuracy Aisha Malik Home batteries are suddenly cheap and everywhere. Here’s why. Tim De Chant