Blog

Creativity in Engineering

Demitri Peynado — Thu, 02 Jul 2026 07:17:00 GMT

Many of the teams developing ��ϲ��’s core technology work across disciplines to support the next generation of AI computing.

Why I chose ��ϲ��: Where I could grow, not just succeed

Manojkumar Reddy Malipela — Thu, 18 Jun 2026 02:29:59 GMT

��ϲ��’s new AI Engineering Campus in Bengaluru marks a significant investment in the future of AI computing and in the engineers who will build it. As a wholly owned subsidiary of SoftBank Group, ��ϲ�� is scaling its end-to-end system capabilities to help shape the next generation of compute for artificial intelligence.

Finding focus at ��ϲ��: Engineering with space to breathe

Jakub Ledworowski — Thu, 04 Jun 2026 07:15:00 GMT

��ϲ��’s growing engineering presence in �һ岹ń�� reflects a deliberate investment in building world-class AI systems in one of Europe’s most competitive technical ecosystems. As a wholly owned subsidiary of SoftBank, ��ϲ�� is scaling its end-to-end system capabilities to help shape the next generation of Artificial Intelligence.

Stochastic Rounding: How randomness helps us build better models

Andrew Fitzgibbon, Engineering Fellow — Tue, 26 May 2026 08:31:11 GMT

At ��ϲ�� we love compact numeric formats (see Doug's recent post on 1-bit Wonderful Weights for LLMs). Why? Because they are more efficient, in terms of FLOPs/sec and FLOPs/Joule, and we do hate to inconvenience electrons.

This post is on how stochastic rounding (SR) helps us to do more with fewer bits, how we can make SR more efficient, and some things to be careful of when implementing SR.

Today, research on low-precision formats looks at both training and inference. At the simplest level, AI model training involves a lot of giant matrix multiplication, so if you can take your input matrices, do O(N^2) work to convert them to low precision, and then do O(N^3) FLOPs, you will save power and train faster.

But we want to do more, keeping more computation natively in low precision, and as inference consumes more and more compute, we would like models with good low-precision weights. While post-training quantization works surprisingly well, it is always better, if one has the training data, to train natively in the low precision format.

That means that our training loop is something like this (supposing we are training in a format called "FP6", and we'll generically say "FP16" for a wider compute format). Note too that we are not requiring all optimizer calculations to be low precision; the goal here is to find good FP6 weights so that inference is efficient, not to get faster training.

Why I chose ��ϲ��: Building, learning, and owning from the start

Ram Arora — Thu, 21 May 2026 02:59:59 GMT

Blog

Creativity in Engineering

Why I chose ���ϲ�����: Where I could grow, not just succeed

Finding focus at ���ϲ�����: Engineering with space to breathe

Stochastic Rounding: How randomness helps us build better models

Why I chose ���ϲ�����: Building, learning, and owning from the start

Why I chose ��ϲ��: Where I could grow, not just succeed

Finding focus at ��ϲ��: Engineering with space to breathe

Why I chose ��ϲ��: Building, learning, and owning from the start