In the middle of Super Computing 21, Intel has made a small presentation to drop new details about its Sapphire Rapids-SP processors, as well as its graphics card for Xe-HPC, also known as Ponte Vecchio. Both are technologies derived from the development of your PC hardware. For example Sapphire Rapids uses Golden Cove cores that are nothing other than the same P-Cores used in the Intel Core 12, as well as other additions such as the DSM unit.
As for Ponte Vecchio, we can say the same, since its architecture is an HPC version of the one used in the Inter ARC Alchemist that will reach the gaming market in 2022. Let us see, therefore, what news Intel has presented in Super Computing 21.
Intel at Super Computing 21: what’s new
Although they have not presented any new architecture, they have given new details about their Sapphire Rapids architecture, which is the first Intel CPU that will not be built by a single chip, but by several combined with each other. In this case we have 4 different blocks or tiles mounted on an interposer that use the company’s own silicon bridge technology to communicate, baptized as EMIB, with a total of 10 interconnections.
This information is not new, nor is the version with memory HBM2e, which adds 4 more bridges to communicate with the 4 batteries of said memory in TSV also placed in the same interposer to 64 GB of this type of RAM. The utility of making use of this type of memory previously assigned to GPUs for the HPC market? The inclusion of AMX units, designed to accelerate operations with matrices widely used in deep learning, which make use of large volumes of data.
The new information? Intel has revealed the size of the entire package, which will be 4446 mm2 for models without memory HBM2e and 5700 mm2 for which includes said memory. Intel doesn’t really have much left to reveal about its Sapphire Rapids architecture.
New details about Ponte Vecchio
Intel’s icing on the cake for its presentation at Super Computing 21 is its GPU for HPC Ponte Vecchio, with which Intel has followed in the footsteps of NVIDIA by keeping the units for rendering that AMD has removed in the case of its CDNAs. This means that these GPUs will not only be used for scientific computing, but also for gaming in the cloud and even for rendering very complex scenes based on Ray Tracing. And we are not referring only to games, but modest animation studios are going to speed up the work a lot. It is very clear that Intel intends to go toe-to-toe against Jensen Huang’s in all markets.
In the packaging sphere, the GPU is highly complex, as it combines chips from various Intel and TSMC manufacturing nodes and intercommunicates them through technologies such as EMIB and Foveros. All this to place 128 Xe Cores placed in 8 different GPUs that are intercommunicated with each other. As for the cache of this monstrosity we have 64MB L1 cache in two separate piles, as well as 408 MB L2 cache. The enormous complexity of this GPU is due to the fact that it is not a piece of hardware designed for the domestic market, where Intel will also combine chips manufactured by itself with those created in TSMC.
It is the first 3DIC GPU in history, which uses Foveros technology to interconnect its 8 GPUs with each other, as well as with the first and second level caches. Something that so far neither NVIDIA nor AMD have been able to do, so in Ponte Vecchio at Intel they have used all their strength as a foundry.
The details that Intel has revealed in its Super Computing 21? Substrate size is 4857 mm2, each of the two Base die is 650 mm2 and the chips that go above measure up to 41 mm2As you can see, this is not the classic GPU for gaming.