Google discloses Ironwood TPUv7 details

The newest TPU promises a big step up in performance from Trillium (TPUv6). We're awaiting more details. We believe the big jump comes from native FP8 support, and the new chip delivers raw throughput on that type similar to Nvidia Blackwell. Blackwell, however, also supports FP4 at twice the rate, and Nvidia is promoting that type for inference. Google also increased HBM capacity and interface speed, but it's in line with Blackwell and AMD's offerings. A key Google advantage has been interconnect, and Ironwood continues to excel here. The interchip interconnect (ICI) can link 9,216 Ironwoods, creating a huge, flat computing plane. The TPU architecture is unusual. Details for v7 have been withheld, but we expect Google continues to employ a few cores, large MAC arrays, and large on-chip memories. Supplementing these are "sparse cores" that handle embeddings and do other supplementary functions.

Other contents

Intel to invest in SambaNova. This is not a repost

Caltech Researchers Take Another Stab at One-Bit AI Models

Nvidia puts $2B into Marvell

BWR 11: GTC and OFC

BWR 11: GTC and OFC

Arm Chips Are Back, and This Time They Mean Business

Arm Chips Are Back, and This Time They Mean Business

Arm is Building Chips

Arm is Building Chips

What the Heck is a Groq?

What the Heck is a Groq?

Byrne-Wheeler Report Discusses AI Deals, Broadcom and Nvidia Earnings

Byrne-Wheeler Report Discusses AI Deals, Broadcom and Nvidia Earnings

Meta Bares MTIA Roadmap, Accelerates NPU Development

Meta Bares MTIA Roadmap, Accelerates NPU Development

s/CPX/LPX/g

s/CPX/LPX/g