OpenAI Unveils Jalapeño Inference Chip as Custom AI Silicon Expands at Hot Chips 2026
Редакция ЦОД:Главное
Hot Chips 2026 highlighted a widening push by cloud and AI companies to develop proprietary accelerators alongside the established CPU and GPU suppliers. OpenAI publicly introduced Jalapeño, its first in-house chip for large-language-model inference, while Google detailed separate eighth-generation TPUs for training and inference.
- OpenAI presented Jalapeño as its first proprietary chip for large-language-model inference.
- Broadcom supported Jalapeño’s silicon implementation, networking and system connectivity.
- OpenAI claims 1.5x–1.9x higher peak throughput per watt and 1.7x–3.6x lower end-to-end latency for specified models versus comparison systems.
- Google split its eighth-generation TPU into TPU 8t for training and TPU 8i for inference.
- Custom accelerators could increase the need for platform-specific power, cooling and network design in AI clusters.
Hot Chips 2026 underscored the broadening field of AI-accelerator development beyond Intel, AMD and NVIDIA. Google, Microsoft, Meta, OpenAI, Cerebras and SambaNova were among the companies presenting or discussing proprietary AI silicon.
OpenAI introduced Jalapeño as its first internally developed chip, optimised for large-language-model inference. OpenAI designed the architecture, while Broadcom contributed to silicon implementation, networking and system connectivity.
For GPT-OSS 120B, DeepSeek R1 and Kimi K2.5, OpenAI claims Jalapeño can deliver 1.5x to 1.9x higher peak throughput per watt than comparison systems, alongside a 1.7x to 3.6x reduction in end-to-end latency. These are vendor-stated preliminary results; the available source does not disclose the comparison methodology, system configurations or test conditions.
Google also outlined an eighth-generation TPU architecture split into dedicated variants: TPU 8t for training and TPU 8i for inference. The separation reflects a wider move to optimise accelerator platforms around distinct workload requirements rather than use a single general-purpose AI configuration.
For data-center infrastructure, this trend could alter the power, cooling and network requirements of large AI deployments. If specialised platforms improve performance per watt or reduce latency in production workloads, operators may place greater emphasis on platform-specific integration of IT racks, power delivery, cooling and fabric design. The available information, however, provides no data on deployed volumes, server-level power draw, capital expenditure or associated data-center capacity additions.
The shift toward workload-specific ASICs and accelerators could reshape the design basis for large AI clusters. As hardware is tuned more closely to model and inference profiles, data-center operators may need to co-design rack power, cooling, networking and server procurement around specific platforms rather than standardised GPU configurations.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.