Google announced TorchTPU, an open-source PyTorch backend that allows developers to run PyTorch applications natively on Google's Cloud TPUs with minimal code changes. The move, announced at the PyTorch Conference on October 20, 2026, follows Amazon's similar effort and aims to reduce developers' reliance on Nvidia's AI hardware and software ecosystem.
OpenAI presented benchmarks at Hot Chips conference showing its 700W Jalapeño ASIC, co-developed with Broadcom, delivers up to 1.9x throughput per kilowatt and 3.6x lower latency compared to Nvidia's 1,400W flagship GPU. The chip is designed for fast AI inference at scale.
d-Matrix presented Raptor at Hot Chips 2026, featuring a TSMC 4nm compute die bonded face-to-face at 36-micron pitch directly onto custom DRAM, achieving 100 terabytes per second memory bandwidth while using one-sixth the energy per bit of HBM3.
SK hynix and SanDisk unveiled the first technical specification for High Bandwidth Flash (HBF), a new memory technology designed to bridge the performance gap between HBM and SSDs in AI accelerators, offering up to 3 TB/s bandwidth and terabyte-scale capacity. The specification was announced at the FMS 2026 conference, six months after the consortium's formation, with support from Google and Tenstorrent.