The “Ponte Vecchio” and “Arctic Sound” data center GPUs are being phased out by Intel.
It appears that Intel has begun a gradual period of deprecation for its “Arctic Sound” and “Ponte Vecchio” Data Center GPU Flex series. Data Center GPU Flex Series and Data Center GPU Max Series are deprecated in version 1.3.3 of the Intel XPU Manager, a free and open-source utility for monitoring and controlling Intel data center GPUs, according to the changelog. It is now recommended that owners of these GPUs avoid updating to the most recent version and instead continue with the older version 1.2.42 for optimal feature support. This is occurring for a number of reasons, and one of the theories is that Intel is getting ready for the “Jaguar Shores” AI accelerator shift, despite the fact that its Data Center GPU Flex/Max series has just been on the market for two years.
Ponte Vecchio is a huge chip with around 100 billion transistors and a size of 1,280 mm². It was constructed using Intel’s 10 process on the Xe-HPC microarchitecture. It has 1,024 tensor cores, 16,384 shading units, and a maximum TDP of 600 W per card. The design used to be attractive for specific HPC and AI applications because of the Max 1100, 1350, and 1550 SKUs, which provide 48 to 128 GB of memory. But because of its size, it is expensive, uses a lot of electricity, and presents integration problems. The much smaller Intel “Arctic Sound,” which is based on the Generation 12.5 GPU microarchitecture and is produced on the Intel 10 node, has a TDP of over 500 W and measures around 190 mm². It has about 8 billion transistors, 8,192 shading units, and 128 ROPs. With the GPU debuting in 2021, the Arctic Sound 1T, 2T, and M SKUs are small 16 GB designs targeted for midrange data center jobs.
For a few obvious reasons, these designs received limited application in enterprise settings and never gained a sizable market share. Established accelerators and well-established software stacks already controlled the market that Ponte Vecchio and Arctic Sound joined. A strong ecosystem, tooling, and driver stability are just as critical to most consumers as raw silicon. Given the expenses of migration and support, integrators and cloud providers faced extra difficulties as a result of Intel’s need to simultaneously develop hardware and software ecosystems. Nonetheless, Intel accomplished a noteworthy engineering feat by successfully integrating these GPUs into the Aurora supercomputer. The chips’ ability to function at the largest scale and satisfy the needs of a demanding client was shown by this deployment. Nevertheless, suppliers were deterred from adopting an Intel-only strategy by Intel’s frequent changes in accelerator strategy. More sophisticated silicon, HBM4, and an established OneAPI software stack will be features of future designs, including the forthcoming Jaguar Shores rack-scale platform.
