Computer

NVIDIA Grace: the 144-core processor to outperform Intel and AMD

Undoubtedly, what NVIDIA proposes is an option halfway between Intel, AMD and even Apple, an option as intelligent as it is revolutionary that promises to surpass its rivals and possibly even the tables when the two main players in the industry launch their bets in this 2022. EPYC and Xeon are in the crosshairs of Huang and this “new” concept could revolutionize the AI ​​and DL industry.

NVIDIA Grace, a system more than a processor in itself

The first thing that we must be clear about is that we are talking about a SoC practically to use and not about a processor as such, since being technical and specific is more faithful to reality. NVIDIA Grace is the brand’s blueprint for ARM and is therefore based on the N2 Perseus platform which is also the first to support the ARM V9 architecture.

It should be noted that we are talking about two SoCs differentiated by their name and their characteristics: Grace Hopper and Grace CPU SUPERCHIPwhere the first is the most interesting as it includes CPU and GPU on the same PCB, while the second is two CPUs joined with the same methods and designed for a more specific market in complex tasks.

This at the same time implies more details and added technologies, since as such NVIDIA specified that the reduction of the lithographic node together with the aforementioned manages to add CXL 2.0, CCIX 2.0, DDR5 and PCIe 5.0 like its rivals on desktop and also HBM3. This without forgetting technologies such as C2C NVLink Chip-to-Chip with up to 900GB/s and memory coherence.

NVIDIA-Grace-Hopper

All this is very good and represents a giant leap in the industry, well ahead of what Intel and AMD currently have for the same sector, but if NVIDIA is based on neo-verse and add a whopping 72 cores in a dual configuration that make the aforementioned count of 144 Cores, as we have a bomb that is difficult to digest for its rivals.

Specifically, the SoC is made up of these two ARM v9 chips from 72 cores attached to a Hopper GPU connected as we said by C2C and its bandwidth, but the most impressive thing is not this in itself, but the fact that its maximum consumption is encrypted at 500 watts in total for the two CPUs, a technological milestone if we take into account that they integrate SVE scalable vector instructions.

NVIDIA-Grace-SuperChip

But there is more, since being a SoC it needs RAM memory, which arrives as a soldered LPDDR5X and at its maximum capacity per battery, that is, 64GBso given the module count (16, eight for each chip) we have a whopping 512GB in total where thanks to the C2C and the improvements of the V9 architecture NVIDIA affirms that the internal bandwidth has increased 30 times nothing less.

A SoC with 600 GB of memory in total

NVIDIA does not add here the impressive 396 MB CPU cachebut it does specify that the bandwidth that can be achieved internally is 1TB/sdouble what Intel and AMD currently make on their EPYCs and Xeons.

As for the GPU itself, there is not much data and it makes sense, since it is reserved so as not to give clues about desktop architectures as such.

The presentation data of the GH100 is not very specific either, quite the contrary, since it was revealed that it is manufactured at 4 nm, it will have 144 SM distributed in 8 GPC with 48MB of L2 and little else, so it is a mystery that will be revealed in a few months.

What NVIDIA did leave in the running are the 80GB of HBM3 memory, a truly impressive figure that sets a new benchmark in the sector. As for its performance, there is no data shown by the company, but there are statements as such. NVIDIA claims that its Grace is 1.5 times faster than the EPYC Rome 7742 in SPECrate_2017_int_base, where it also achieves an efficiency of 250 watts per chip versus AMD’s 280 watts.

NVDIIA-HGX-H100

If we add to this the fact that NVIDIA claims that Grace is compatible with RTX, IA, HPC and Omniverse, it is certain that Intel and AMD have a serious competitor that will arrive in the future. first half of 2023. On the other hand, and due to the specified time GAP, NVIDIA has commented that there will be a technical document from Grace later, where we hope to know more details about its GPU.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *