On Friday afternoon, Intel published a letter by Jeff McVeigh, the corporate’s interim GM of their Accelerated Computing Methods and Graphics group (AXG). In it, McVeigh supplied a short replace on the state of Intel’s server GPU product lineups and buyer adoption. However, extra importantly, his letter supplied an replace to Intel’s server GPU roadmap – and it’s a little bit of a bombshell. Briefly, Intel is canceling a number of server GPU merchandise that have been deliberate to be launched over the subsequent 12 months and a half – together with their HPC-class Rialto Bridge GPU – and going all-in on Falcon Shores, whose trajectory has additionally been altered, delaying it till 2025. There’s quite a bit to digest right here, so let’s dive proper in.
Intel’s most up-to-date public server GPU product roadmap, launched again throughout SC22 in Could of 2022, outlined that Intel supposed to launch server GPU merchandise at a reasonably fast clip. This was (partly) to make up for misplaced time from the a number of delays involving Ponte Vecchio, Intel’s first HPC-class GPU that’s now bought beneath the Information Middle GPU Max household. At that occasion, Intel introduced Ponte Vecchio’s successor, Rialto Bridge, which was to be an evolution of Ponte and was set to pattern in mid-2023. Following up on Rialto could be Falcon Shores in 2024, which together with using a good newer model of Intel’s Xe GPU structure, could be Intel’s first mixed CPU+GPU product (XPU) and was slated to be the final word evolution of each Intel’s HPC CPU and GPU strains.
However, citing “a objective of maximizing return on funding for patrons”, Intel has considerably refactored their server GPU roadmap, cancelling their beforehand deliberate 2023/2024 merchandise. Rialto Bridge, introduced lower than a 12 months in the past, has been scrapped. As an alternative, Intel will probably be shifting straight on to Falcon Shores for his or her HPC GPU wants.
These cancellations additionally affect Intel’s Flex line of server GPUs for cloud gaming and media encoding, with Intel axing Lancaster Sound (aka “Subsequent Sound”) in favor of shifting on to its successor, Melville Sound. Melville Sound’s improvement will probably be accelerated, in flip, although Intel hasn’t hooked up a tough date to it. Beforehand, it was anticipated at across the similar time as Falcon Shores.
Intel didn’t present any new visuals for the brand new roadmap, however I’ve gone forward and kitbashed one thing collectively primarily based on their SC22 slide with a purpose to present the place issues stand, and what merchandise have been cancelled
In response to Intel, these product modifications are designed to permit the corporate to align itself to a two-year product launch cadence for server GPUs. Rivals NVIDIA and AMD have been working at an analogous cadence for the previous a number of years, so this may carry Intel’s product cycles roughly in step with the competitors. Which, as Intel places it “matches buyer expectations on new product introductions and permits time to develop their ecosystems.”
Studying between the strains, the implication is that Intel wasn’t assured on their means to promote Rialto and Lancaster to their core buyer base. Whether or not that’s solely due to the fast product launch cycles, or if there was one thing extra to it, is a element that may stay contained in the halls of Intel for now. However regardless, the corporate is basically scrubbing their plans to launch new server GPU merchandise this 12 months with a purpose to give attention to their successors afterward.
And in case there’s any doubt about whether or not this can be a good or dangerous improvement for Intel, I’ll word that this info was launched after 5pm Jap on a Friday. So Intel could be very clearly attempting to bury dangerous information right here by releasing it on the finish of the week.
Falcon Shores Goes from an XPU in 2024 to a GPU in 2025
Apart from cancelling Rialto Bridge and Lancaster Sound, the opposite main replace to return from Intel’s letter is the revelation that Intel has considerably refactored their plans for Falcon Shore. What was (and should but nonetheless be) Intel’s first mixed CPU+GPU product for HPC has now been retasked to function Intel’s next-generation HPC GPU. Which in flip has important repercussions for Intel’s product lineup, and their aggressive positioning.
Falcon Shores was one other 2022 Intel announcement, which was revealed again on the firm’s 2022 Investor Assembly in February. At a excessive degree, Falcon Shores was designed to be Intel’s first XPU product – a chip utilizing a wide range of compute architectures with a purpose to greatest meet the execution wants of a single workload. This might be achieved through the use of separate processor tiles, and whereas in observe this primarily meant placing Xe GPU tiles and x86 CPU tiles collectively on a single chip, the tiled strategy left the door open to including different varieties of accelerators as properly. Intel is already doing tiled server chips with merchandise like Ponte Vecchio and the not too long ago launched Sapphire Rapids XCC, so Falcon Shores was slated to be the subsequent step by putting greater than only one sort of accelerator on a chip.
On the time, Falcon Shores was slated for launch in 2024, to be constructed on an “angstrom period course of”. As deliberate, Falcon Shores would have been Intel’s energy play for the HPC trade, permitting them to ship a mixed GPU+CPU product primarily based on their newest architectures and constructed on their cutting-edge 20A/18A course of node, givng them an edge within the HPC market and placing them forward of conventional GPU-based accelerators. And whereas Falcon Shores might but reside as much as these objectives, it received’t be doing so in 2024. And even 2025, for that matter.
As an alternative, Falcon Shores has been retasked to function an HPC GPU half. Intel’s temporary letter doesn’t go into the technical modifications, however the implications are that reasonably than mixing CPU and GPU tiles on a single chip, Intel goes to commit itself to constructing a product out of simply GPU tiles, decreasing what was to be intel’s XPU right into a extra easy GPU.
With the cancellation of Rialto Bridge, it goes with out saying that Intel wants a brand new HPC-class GPU product in the event that they need to stay available in the market. And whereas there’s at the moment no motive to suppose that Falcon Shores received’t be product, the truth that Intel has delayed it till 2025 shouldn’t be a promising signal. Ponte Vecchio is already an older design than its age lets on – it was initially supposed to launch in 2021 and compete with the HPC GPUs of that period – so by not having something newer to supply till 2025, Intel is basically ceding the HPC GPU marketplace for the subsequent two years, outdoors of their supercomputer wins. Intel’s long-term plans nonetheless name for the corporate to take a large chunk of the extremely worthwhile server GPU market – to swipe a chunk of what’s largely been NVIDIA’s pie – so these developments basically put these plans on maintain for an additional two years.
And what might show worse, this places Intel even farther behind its rivals in delivery an XPU/APU-like product. As confirmed by Serve The Home, whereas Intel hasn’t given up on delivery a Falcon Shores XPU, the prioritization of GPUs signifies that such a product wouldn’t launch in 2025. Meaning an Intel server-class XPU is now a 2026 product at greatest – and longer nonetheless at worst.
In the meantime AMD will probably be delivery an analogous server APU later this 12 months with the Instinct MI300. That half will make use of a number of chiplets/tiles on a single chip with a purpose to provide HPC-class CPU and GPU efficiency on a single chip, conducting a long-awaited objective for AMD. And whereas NVIDIA is a bit farther behind in integration with their Grace Hopper superchip – it’s basically two tightly-coupled chips on a single board, reasonably than being chiplets on a single chip – it’s nonetheless forward of the place Intel is at this time. And, worryingly for Intel, the next-gen successor to that half is sort of sure to launch earlier than any potential Falcon Shores XPU hits the market.
In different phrases, by pushing again their server GPU schedule, all of Intel’s merchandise that relied on it are actually equally delayed. That leaves Intel’s server CPUs to carry the road on their very own for the subsequent few years.
A Historical past of Server GPU Cancellations as Intel Retains Making an attempt
Finally, the cancellation of Rialto Bridge makes it the newest in what has turn out to be a surprisingly lengthy line of cancelled GPUs and GPU-like accelerators over at Intel. The corporate has, in a single type or one other, been trying to interrupt into the accelerator market because the late ‘00s, with tasks resembling Larrabee. And now, 15 years later, Intel nonetheless isn’t as far alongside because it needs to be, and remains to be cancelling chips to get there.
As the newest in Intel server GPU casualties, Rialto Bridge is becoming a member of not solely Larrabee, however Intel’s ill-fated Xeon Phi lineup as properly. And even then, Rialto Bridge can’t declare to be the primary Xe structure half to get canned – that honor goes to Xe-HP, which was cancelled in favor of using Intel’s mixed compute and graphics Xe-HPG silicon.
The silver lining to this example, no less than, is that Intel isn’t cancelling a complete structure right here. The Xe structure has firmly established itself as Intel’s in-house GPU structure, and the truth that it’s delivery in the whole lot from SoCs to video playing cards to HPC accelerators underscores its viability and significance at Intel. The cancellation of Rialto Bridge coupled with the delay of Falcon Shores remains to be a major setback for Intel, however on the finish of the day it’s simply shelving one iteration of the broader Xe structure for an additional.
Within the meantime, Intel will preserve working to broaden their GPU choices, and to interrupt into the HPC GPU market in a much bigger manner. It’s a market that is just too large and too worthwhile for Intel to disregard – least it eats much more into their server CPU gross sales – in order that they should preserve at it till they discover the success they search.