According to Toms Hardware Chipzilla is winding them down just a couple of years after they first hit the scene, all because Intel's got its eye on the next big thing – Falcon Shores and the new Gaudi 2 and 3 AI accelerators.
Intel's not yet ditching Ponte Vecchio completely, but they're not pushing them anymore. They'll keep making them, but only for those who already have them. New customers will have to see what AMD and Nvidia offer.
The HPC and AI game is mega competitive these days, what with the AI craze and Ponte Vecchio's old news now, especially with AMD's shiny new MI300 and Nvidia's B200 Blackwell on the block.
Ponte Vecchio first popped up in 2022, and it was the biggest GPU Intel's ever made. Packed with over 100 billion transistors across 47 tiles, made on five different process nodes. Depending on how you set it up, a Ponte Vecchio GPU could have up to eight compute tiles and four HBMe2 stacks, all based on Intel's beefy Xe HPC architecture – that's even more hardcore than the standard Xe stuff in their Arc Alchemist GPUs.
Ponte Vecchio hasn't exactly set the world on fire. Take the Aurora Supercomputer – it's not exactly leading the pack, guzzling more juice and delivering less oomph than AMD's Frontier.
The latest Top500 list has Aurora in second place with 1,012 petaflops and using 38,698 kW, while the older Frontier's still ahead with 1,206 petaflops and only 22,786 kW. Sure, Aurora's got its moments with certain tasks, but it was meant to be hitting 2 exaflops – it's only halfway there.
Ponte Vecchio's looking a bit long in the tooth next to AMD's MI300 and Nvidia's B200. And
Falcon Shores was meant to be out in 2024, mixing it up with CPU and GPU cores to take on AMD's MI300 APU and Nvidia's Grace Hopper Superchip. However, Intel pushed it back to 2025 and scale it down to just a GPU.
Kicking off Ponte Vecchio's retirement before Falcon Shores rocks up means Intel can throw more at getting Falcon Shores out the door quicker. It might help them catch up a bit with AMD and Nvidia, but they're still playing catch-up, since the other two are already cooking up the next gen of HPC hardware.