256 cores and 2 nanometers – AMD puts its monster Epyc Venice server into production – Korben

256 CPU cores on a single chip: This is what AMD’s new Epyc Venice offers, the sixth generation of its processor for servers, mass production of which has just started at Taiwanese founder TSMC.

Remarkable technical detail, it is the very first processor intended for high performance computing (HPC, these giant machines which run climate simulations, AI models or quantum chemistry calculations) to be engraved on TSMC’s 2 nanometer manufacturing node, called N2.

A little perspective to situate the beast. AMD’s largest consumer desktop today has 16 cores and 32 threads, while Venice has 256 cores and probably more than 500 threads on a single socket, on a single motherboard, in a single server. You take the most powerful processor in your PC, and you multiply it by sixteen. This is what AMD squeezes into a single machine.

On the figures side, the manufacturer announces a gain of more than 70% in overall performance compared to the previous generation (the Epyc Turin, which capped at 192 cores), a thread density up by 30%, and above all a memory bandwidth which more than doubles, from 614 GB/s to 1.6 TB/s per socket.

The connection between the CPU and the GPU is also multiplied by two. For data centers running large AI models, where the bottleneck often comes from the speed at which the computing chips are powered, that’s a big leap.

The transition to 2 nanometer engraving is an important step. In practice, “2nm” no longer has much to do with a real physical measurement, it has become a commercial name to designate a new generation of manufacturing processes at TSMC.

But behind this, we are talking about an increase in finesse which makes it possible to fit more transistors per square millimeter and to improve the performance to electricity consumption ratio.

Apple has secured a good part of the founder’s initial capacity on this node, and AMD is among the first to be served behind. Intel, for its part, has only announced its competing P-core equivalent (the big cores for servers) for 2027 at the earliest.

AMD also confirmed that the next generation, called Verano, is already on track and that Venice production will eventually move partly to the TSMC factory in Arizona, just to diversify the supply chain in the face of geopolitical tensions around Taiwan (and to please Trump).

Suddenly, the AMD vs Intel fight on the server market is taking a big turn. Intel spent two years trying to catch up on Zen cores, without really succeeding. With Venice, AMD still leaves it a year and a half to run behind.

In short, for anyone who thought that the race for server cores was starting to run out of steam, well no.

Source :
Tom’s Hardware

This article may contain AI-generated images – I take great care with each article, however, if you spot a dumpling, let me know!