One of the issues Intel has faced so far is not having a powerful GPU architecture, which has caused it to lose large supercomputer contracts to AMD, which can win contracts. more easily thanks to its combined CPU and GPU technology. With NVIDIA buying ARM and leaving Intel out of its designs, the only solution that exists for Pat Gelsinger is to choose to develop its own GPU.
Intel is advancing its Ponte Vecchio GPU
Some Ponte Vecchio eye candy pic.twitter.com/fCG2rZrozI
– Raja Koduri (aRajaontheedge) May 5, 2021
Through his Twitter account, Intel’s chief graphic architect Raja Koduri once again referred to the Ponte Vecchio through an animation where we can see how his most ambitious design to date is organized by different tiles or chips., The GPU for HPC or high performance. calculation of Ponte Vecchio.
Ponte Vecchio is based on a design of Intel Xe-HPC simetric GPU backs composed of tchiplets islands, which to communicate with each other uses the latest Intel packaging technologies such as Foveros and EMIB. And is that we are faced with a design that has a complexity of one hundred billion transistors distributed over its 47 tiles.
SuperMUG-NG: Ponte Vecchio beyond Aurora
Thanks to Germ an support Hardwareluxx, we learned that at Leibniz Supercomputing Center in Munich, Germany, Intel was awarded a contract for the development of the second iteration of the supercomputer. SuperMUC-NG for 2022
SuperMUC-NG processors will be Intel “Saphire Rapids”, the fourth generation Intel Xeon Scalable which will share the core of Golden Cove with Alder Lake and is expected to use up to a 1 PB in storage Via a DAOS (Distributed Asynchronous Object Storage) where SSDs based on Intel Optane persistent memory will be used.
In principle, it wouldn’t be a surprise for supercomputers with Intel processors, but in the chosen GPU there is a big change, because instead of going for an NVIDIA Tesla GPU this time, they went for the Intel Ponte Vecchio GPU. Which is an important victory for Intel, since the current iteration of SuperMUC-NG uses precisely NVIDIA GPUs, which shows us that the Ponte Vecchio GPUs outperform the NVIDIA GA100, which is NVIDIA’s GPU for HPC at this day.
We have to keep in mind that NVIDIA releases a new architecture every two years and we know from its most recent roadmap that the next architecture could appear in 2022. On the same date Intel will launch the Ponte Vecchio . It is not known if the next generation of NVIDIA was considered for the second iteration of SuperMUC-NG, but it cannot be denied that winning such a contract is not easy. Especially if your rival is the almighty NVIDIA which dominates this market with an iron fist.