Back in November, Intel detailed its supercomputer ambition with Ponte Vecchio—a 7nm GPU built with the Intel Xe architecture. Now, Intel's maiden data centre graphics card has appeared in pre-alpha form, according to a registration on the EEC, way ahead of its 2021 release window.
The Ponte Vecchio RVP (reference validation platform) GPU is likely a far cry from the Xe chip that will one day make its way into the US Department of Energy's Aurora supercomputer in 2021—powered by Intel Xe, OneAPI, and Sapphire Rapids Xeon CPUs. It's not yet possible to say what state Ponte Vecchio is in. All we know is that Intel's reportedly working up to a "V5" version of the would-be GPU, and comes in a familiar form factor as an add-in card.
The EEC listing confirms three variants of a Ponte Vecchio RVP AIC card: GAPV3KI2TC, GAPV4KI2TC, and GAPV5KE2TC. All three are listed as "pre-alpha".
At Ponte Vecchio's core is Intel Xe, the overarching graphics architecture that's set for its first public release with Tiger Lake chips, and probably a discrete GPU, later this year.
Ponte Vecchio differs from these consumer parts in a couple of ways, however. For one, it's built using the full-fat Intel Xe-HPC architecture, as opposed to the Xe-HP and Xe-LP architecture subsets for professional and gaming applications. It specialises in modelling, simulation, and AI workloads.
And while Intel Tiger Lake and friends will are planned for the 10nm process node, Ponte Vecchio will come in 7nm garb. It is also set to incorporate many of Intel's developing technologies, such as Foveros 3D and EMIB packaging tech and the CXL-based Xe Link interconnect.
Intel's evidently set itself lofty goals for Intel Xe, combining many largely untried and untested technologies into a single product and setting itself a brief window to get it done by. But whatever state Ponte Vecchio might be in today, these pre-alpha cards are a sign of movement from the Xe graphics team after a quiet patch in the wake of CES.