Google Translate is turning 20! . There are 20 fun facts and tips in the thread below. Translate is one of my favorite Google products because it brings us all closer together! I've been involved with a couple of things over the years. The first was our deployment of the
@jeffdean
-
Cloud Next Conference: AI Infrastructure Discussion with Industry Experts
By
–
The video of my conversation with Amin Vahdat, @gilbert
, and @djrosent at Cloud Next last week is now up. https://
youtu.be/BpnJYJmbXcM?si
=vUY3hI_aDX8K6gco
… Thanks for a great conversation! -
Decoupled DiLoCo Paper Now Available on Arxiv
By
–
The Arxiv for the new Decoupled DiLoCo paper is now up:
-
Large-Scale Asynchronous Training Techniques for Neural Networks
By
–
It's worth pointing out that we have been pushing on large-scale training and asynchronous techniques for the last ~14 years. Here's our NeurIPS 2012 paper where we demonstrated that this approach could be used to train very large neural networks (for the time: 30X larger than
-
Google TPU 8i Co-Designed for Low Latency Inference
By
–
TPU 8i is co-designed with our Gemini research team to support low latency inference. Among the attributes that support this are large amounts of on-chip SRAM, enabling more computations to be done on chip without having to go to HBM for weights or KVCache state as often. The
-
TPU 8t: 3X FP4 Performance Boost Over Ironwood
By
–
First, let's talk about TPU 8t, which is designed for large-scale training and inference throughput. The pod size is increased slightly to 9600 chips, and provides ~3X the FP4 performance per pod vs. Ironwood (8t has 121 exaflops/pod vs. 42.5 exaflops/pod for Ironwood). In
-
Google announces eighth-generation TPU chips for agentic era
By
–
I had a good time discussing yesterday's Google TPU v8t and v8i announcement at Cloud Next with Amin Vahdat along with @AcquiredFM hosts @gilbert and @djrosent
. The blog post announcement has lots of details about these new chips: https://
blog.google/innovation-and
-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
… Here's a thread of -
Decoupled DiLoCo Training System Enables Resilient Large Scale AI
By
–
It's been a delight to provide small amounts of advice and suggestions to people working on the Decoupled DiLoCo training system. This approach enables graceful handling of failures in large scale training jobs, by allowing (N-1) / N units to proceed when one fails.
— Jeff Dean (@JeffDean) 23 avril 2026
Thread ⬇️ https://t.co/z97PgtNBuuIt's been a delight to provide small amounts of advice and suggestions to people working on the Decoupled DiLoCo training system. This approach enables graceful handling of failures in large scale training jobs, by allowing (N-1) / N units to proceed when one fails. Thread
-
Modern Computer Architecture: Quantitative Evaluation and Design
By
–
Yes, content is still highly relevant today (quantitative evaluation, modern computer architecture aspects like cactus, branch predictors, TLBs, different types of computer arithmetic, warehouse computer, vector processors, and more). Table of contents below.
-
Ricardo Receives ACM Award Recognition
By
–
In case it's not clear, you can click on the picture in the first image to learn more about Ricardo and the award: Twitter swallowed my actual link text to https://
awards.acm.org/about/2025-bar
roso
… to turn it into the picture, I guess.