Almost two years ago to the day, NVIDIA mentioned his platform for the first time Vera Rubincomposed of CPU Vera and GPU Rubin. It was June 2, 2024, during Jensen Huang’s keynote, at the opening of Computex 2024. On March 16, 2026, the firm finally announced the launch of its Vera CPU. A launch on the other hand only for “giants” like Anthropic, OpenAI, SpaceXAI or even Oracle Cloud Infrastructure for example, who already have Vera CPUs in their huge data centers. On the other hand, for less imposing partners, the meeting with Vera is only scheduled for the second half of 2026.
In anticipation of this, our colleague Michael Larabel from the Phoronix site went to attend a little communication stunt organized by NVIDIA with some hand-picked media. NVIDIA notably presented its dual-socket platform which it likes to show as much as possible. It must be said that it’s actually quite pleasing to the eye, don’t you think…?
The NVIDIA Vera dual-socket platform
The opportunity to give a little refresher on Vera, without necessarily mentioning Rubin, no but…! the NVIDIA Vera is therefore the new processor Arm of the company, intended for agentic AI. It follows Grace with one big difference: the use of hearts Olympuswhose custom architecture was developed by NVIDIA itself. For Grace, NVIDIA used Neoverse V2 cores, an architecture borrowed directly from Arm. With Vera, NVIDIA’s ambitions are therefore great…: show the world its know-how, and strike a major blow in terms of performance. The Olympus cores must prove themselves, and we also note the fact that there are now 88 on a single CPU (176 threads), against 72 hearts for Grace.
Now let’s talk about RAM, because a CPU may be as powerful as you want, if it is not “well supported”, the bottleneck can ruin everything. This is also where Vera is strong since the CPU is associated with up to 1.5 TB of LPDDR5X-9600 in SOCAMM2. With the 1024-bit memory bus, this gives a bandwidth of 1.2 TB/s…! In the case of the two-socket platform illustrated above, the fear could this time be the bottleneck created by the communication to be ensured between the two CPUs. But here too, NVIDIA has the solution: a connection NVLink-C2C Ã 1,8 To/s.
Let’s go back to the small committee event organized by NVIDIA, and to the Phoronix site. Michael Larabel asked NVIDIA on site if he could carry out measurements with the equipment presented. NVIDIA accepted, this time with a single-socket platform on the other hand. And above all, with specific requests concerning the type of benchmarks under Linux that were going to be carried out. Basically, only a few benchmarks in situations likely to correspond to real use of a Vera CPU. As Michael Larabel points out in his article, this is not a sponsored test, because in this case it is the manufacturer who offers the test for money, with the right to impose its conditions. There, the situation was simply “if you want to take measurements, it’s on our conditions otherwise no measurements”. Michael Larabel explains that he quickly determined that he would rather have measures, even without real freedom, than nothing at all.
The single-socket Vera platform benched by Phoronix.
We let you discover via the link at the end of the news all the measurements carried out by our colleagues, but if you just want a quick look at the overall result, here is their most telling graph…
On the few benchmarks carried out, Vera outperforms Grace by 63%, which is far from trivial. The AMD and Intel x86 platforms are also no match, and switching to dual-socket does not change anything for them, weighed down as they are by memory bandwidth concerns.
A final word on consumption. NVIDIA refused to allow Michael Larabel to carry out consumption readings for the configuration he was able to test, which will probably sow doubt in the minds of some. In a post made on the official NVIDIA blog about the Phoronix test, NVIDIA claims that the configuration in question had a TDP of 450 W for the Vera CPU itself, which has always been the officially announced value, while the total consumption of the LPDDRX is estimated at around 30 W. Values which, if they are subsequently confirmed by independent sources, are also very impressive with less than 500 W for CPU + RAM, when in x86 it is often the consumption of the CPU alone, with bars which will add another 100 W to that.
The NVIDIA Vera CPU benched by Phoronix, it’s here!
