Intel axes GPUs and delays others for HPC and datacenters • The Register

Intel has used a quiet Friday announcement to disclose delays and deletions to its GPU and high-performance computing roadmap.

Intel’s document opens with cheery chat about GPUs. However nicely beneath the fold are a number of humbling items of reports for Chipzilla. The primary two are the dumping of the Rialto Bridge era of GPUs and suspending the Falcon Shores GPU structure – to 2025.

As The Register reported in Might 2022, when Intel supplied an replace to its HPC roadmap on the Worldwide Supercomputing Convention, Intel deliberate to start out sampling Rialto Bridge in mid-2023. On the time, Rialto Bridge was billed as a worthy successor to Ponte Vecchio GPU, and able to delivering a 30 % efficiency increase. Falcon Shores – a platform with x86 CPU cores and Xe GPU cores in a single package deal – was billed as an “XPU” and talked about as prone to debut in 2024.

Intel’s newest announcement reframes Rialto Bridge as a supply of mere “incremental enhancements” – which has subsequently been discontinued.

Falcon Shores, in the meantime, is now “Focused for introduction in 2025.” Intel argued that rescheduling – and a transfer to a two-year launch cadence for the HPC-grade “Max” GPU vary – “matches buyer expectations on new product introductions and permits time to develop their ecosystems.”

The put up additionally mentions discontinuation of the datacenter-grade Lancaster Sound GPU structure used within the present “Flex” GPU vary – once more on grounds it supplied solely incremental efficiency will increase.

Doing so “permits us to speed up improvement on Melville Sound, which will probably be a big architectural leap from the present era by way of efficiency, options and the workloads it’s going to allow.”

No date for Melville Sound’s debut is obtainable, however the put up implies that the adjustments outlined will end in a two-yearly launch cadence for the Flex vary.

Intel’s information is not all glum on the GPU entrance: the chip store has promised “steady updates for our Max Collection and Flex Collection merchandise, with efficiency enhancements, new options, expanded working methods help and new use circumstances to broaden the advantages of those merchandise.”

Element is, once more, not out there. And with Intel’s business-focused GPU roadmap having been elongated, clients will in all probability want to attend not less than a yr earlier than they get to see if “steady supply” provides something worthy.

Assuming steady supply even works for these merchandise, given Intel’s remarks hat HPC clients want longer lead instances to organize their environments for this form of {hardware}.

There is no solution to sugarcoat Intel’s standing as a GPU laggard: Ponte Vecchio was years late and the Arc vary of standalone client GPUs debuted solely final yr.

Nvidia had standalone product available in the market for each punters and professionals for greater than a decade, whereas Intel centered on built-in graphics.

Intel began a good distance behind on GPUs, and will probably be even additional behind when – or if – Falcon Shores debuts in 2024.

And if the troubled big manages to ship the structure in that yr, it’s going to nonetheless have missed deadlines it set itself – including to an extended listing of merchandise that ended up arriving late and in configurations that rivals had already surpassed.

Intel CEO Pat Gelsinger has been within the job for a little bit over two years. One of many first methods he personally announced was “Built-in System Manufacturing 2.0” (IDM 2.0) – a plan to have foundries apart from Intel itself make some parts of its merchandise, with Chipzilla doing closing integration and packaging.

Falcon Shores will, the once-dominant big promised, be a product of IDM 2.0 at work. By the point it debuts, Gelsinger can have had not less than 4 years within the job (fingers crossed) – making the GPU very a lot a product he’ll must personal.

Hardly anybody, nevertheless, appears to have needed to personal Intel’s GPU roadmap replace. Whereas Chipzilla’s earlier HPC GPU roadmap was delivered at an enormous worldwide gabfest, this replace was squeezed out with out fanfare on a Friday – a time when even members of the press is probably not at their sharpest as minds flip to household and spare time activities.

Political journalists have a phrase for that form of habits: they name it “taking out the trash.” If Intel cannot ship on this roadmap, will probably be in peril of delivering a steaming pile of irrelevance. ®