Nvidia took advantage of the GTC to give news of Grace, its processor for supercomputers and other data centers. Announced in 2021 and planned for 2023, this CPU will be equipped with 77 Arm cores and can be used as a duo (144 cores) or accompanied by a Hopper GPU.
Remember, just a year ago, Nvidia announced work on Grace, a processor for supercomputers and data centers. The company then clarified that Grace is “designed to meet the computing needs of the world’s most advanced applications, including natural language processing, recommendation systems and AI supercomputers, which analyze huge datasets requiring both ultra fast and massive memory capacity“. No technical details were provided yet.
Obviously, the subject could not be evaded by Jensen Huang during the 2022 edition of the GTC. The manager first of all wanted to be reassuring about the smooth development of the chip. No delay is to be expected, as it is still expected for the first half of 2023. In terms of technical characteristics, we understand that the processor has 77 Arm cores, that it is accompanied by 396 MB of cache memory and that it embeds an LPDDR5X memory controller.
Sold as a module, this processor will be offered in two forms. The first is called Grace Hopper and consists of combining a Grace processor with a Hopper GPU. Simple on paper, but it mainly relies on the NVLink-C2C communication link which operates at 900 GB/s or more (unspecified) to connect the two chips. The second form is called Grace Superchip and combines two processors for a total of 144 cores. Again, the communication between the two chips goes through an NVLink-C2C link, a somewhat magical technology that allows you to communicate a bit of everything and nothing: CPU, GPU, DPU, SoC and so on. This communication can be done between two chips via the PCB or an interposer as well as between two dies – like Apple’s UltraFusion process, which operates for its part at 2.5 TB / s.
Solutions based on the Grace processor will therefore be offered to Nvidia customers in the second half of 2023, particularly within HPE (Hewlett Packard Enterprise) supercomputers. Because, we recall, this processor has, unlike the M1 Ultra chip, no vocation to be found in conventional computers, but in servers intended for AI in particular. What come to compete with solutions from Amd (Epyc) and Intel (Xeon) and thus offer turnkey server solutions entirely built around Nvidia technologies.
Still, we are now dreaming of developing a processor – or more precisely a SoC – Nvidia for consumer computers. The market seems ready: Apple has revitalized its Macs with its M1 series and Microsoft now offers everything you need for Arm processors. And given the poor performance offered by Qualcomm solutions, the arrival of a new player for PCs would be energizing for this sector.