100 billion transistors: Intel’s Ponte Vecchio GPU will go head-to-head with Nvidia, AMD

IT House May 9 news, there is no doubt that AMD, Nvidia, Intel and more are looking forward to the future of artificial intelligence, machine learning and supercomputers.

In November last year, AMD released the new Instinct MI250 / MI250X GPU, which uses 2.5D EFB bridge technology, the industry’s first multi-Die integrated package (MCM), and integrates two cores inside. And in March of this year, Nvidia showed off the new GH100 GPU with their H100 SXM5 high-performance computing module. Now, it’s finally Intel’s turn, and the Ponte Vecchio GPUs will help them implement more algorithms using graphics technology over the next decade for building data centers and machine learning systems.

(AMD Instinct MI250X GPU)

(NVIDIA GH100 GPU)

Intel has pledged to use 100 billion transistors in the new Ponte Vecchio GPU, which will bring the company into more mainstream artificial intelligence and high-performance computing applications.

(Intel Ponte Vecchio GPU)

Last week, Intel confirmed on Twitter that they have begun testing and spot-checking Ponte Vecchio GPUs. Its new compute cards will enable the company to give consumers more development and functionality through AI and HPC applications. Intel’s PR manager Mikael Moreau revealed a water-cooling module specially designed for the Ponte Vecchio GPU,This means that Intel has already started delivering parts to partners in the United States.

Intel’s next-generation Ponte Vecchio GPUs contain up to 47 chips (including 16 X^e-HPC architecture computing chips, 8 Rambo cache chips, 2 Xe base chips, 11 EMIB link chips, 2 Xe Link I/O chips and 8 HBM chips), the whole set of chips contains a staggering 100 billion transistors , such a huge size will consume a lot of power. Ponte Vecchio’s 47 chips will be manufactured using different node processes, including TSMC’s 7nm or 5nm process, as well as Intel’s 7nm and 10nm processes, and will be integrated with Foveros 3D packaging through EMIB.

IT House learned that Ponte Vecchio adopts OAM module, which is an open hardware computing accelerator form and interconnection structure. There are five layers, from bottom to top are the bottom plate, PCB circuit board, top plate, water cooling device and fixed cover plate. The new OAM computing module requires 600W of power. Since the operating power consumption of the system is so terrible, it needs to obtain a powerful cooling system. Intel has chosen to investigate liquid cooling for the new compute cards, but the newly developed system has the smallest footprint compared to traditional large cooling systems. It’s not clear from the Twitter thread whether we’re seeing a standard Ponte Vecchio GPU or an advanced XT version of the next-gen GPU.

The Intel Ponte Vecchio looks pretty big, almost the size of a tablet. However, given its transistor count and power consumption, such a large size is not too surprising. Ponte Vecchio is expected to deliver “PetaFLOPS-level AI performance.” (Peta: used for the unit of measurement, representing 10 to the 15th power, representing a quadrillion)