Over the past few months, we have seen two out of three 2022 compute GPUs for AI and HPC applications from leading developers — AMD's Instinct MI250X OAM module based on the codenamed Aldebaran dual-tile GPU as well as Nvidia's H100 SXM5 module featuring the GH100 GPU. Time to meet the third contender — Intel's Ponte Vecchio.
Last week Intel confirmed that it had started sampling its Ponte Vecchio compute GPU for artificial intelligence (AI) and high-performance computing (HPC) applications with customers without any elaborations. However, this week Mikael Moreau, an Intel PR manager for Western Europe, posted a picture of Intel's Ponte Vecchio module with a cooling system on Twitter, which confirms that these parts have shipped outside Intel's labs in the U.S.
Intel's Ponte Vecchio is the company's upcoming flagship compute GPU that sports a design with around 47 tiles that contain 100 billion transistors in total. Naturally, these transistors want power, and a Ponte Vecchio OAM compute module will consume about 600W, which requires liquid cooling. We have already seen Ponte Vecchio with a pre-production water block, but the real one certainly looks considerably more compact and only needs one input and output for liquid.
Meanwhile, the Intel representative did not say whether he showcased a regular or XT version of Ponte Vecchio, which offers higher performance at the cost of higher power consumption and extra heat.
Intel's Ponte Vecchio module, with its water block and cooling system, looks rather massive, which is not particularly surprising given its many transistors and how much computing power it promises to offer (PetaFLOPS-class AI performance in Raja Koduri's terminology).
In addition to a Ponte Vecchio sample, Mikael Moreau also said he had Intel's 4th Generation Xeon Scalable 'Sapphire Rapids' processor with onboard HBM memory. Both parts will power the 2 ExaFLOPS Aurora supercomputer that will come online later this year.