Intel Scraps Rialto Bridge GPU, Next Server GPU Will Be Falcon Shores In 2025
friday afternoon, Intel published Jeff McVeigh’s letter, interim GM of the company’s Accelerated Computing Systems and Graphics Group (AXG). In it, McVeigh provided a brief update on Intel’s server GPU product lineup and customer adoption. But more importantly, his letter provided an update on his Intel’s server GPU roadmap. In short, Intel is going all-in on Falcon Shores, canceling multiple server GPU offerings that were due to be released over the next year and a half, including his Rialto Bridge GPU in the HPC class. That is until 2025. There’s a lot to digest here, so let’s dive right in.
Intel’s latest public server GPU product roadmap released at SC22 in May 2022 outlined that Intel intended to release server GPU products fairly quickly. This was (partly) to make up for lost time from multiple delays associated with Ponte Vecchio, Intel’s first HPC-class GPU currently on sale under the Datacenter GPU Max family. At that event, Intel announced the Rialto Bridge, the successor to the Ponte Vecchio. This is an evolution of Ponte, with samples scheduled for mid-2023. Following Rialto is 2024’s Falcon Shores, which will feature an even newer version of Intel’s Xe GPU architecture as well as being Intel’s first integrated CPU + GPU product (XPU), the ultimate in both of Intel’s HPC CPUs. It was planned to be an evolution of and the GPU line.
However, citing “the goal of maximizing customer return on investment”, Intel has significantly refactored its server GPU roadmap and canceled previously planned 2023/2024 offerings. Unveiled less than a year later, the Rialto Bridge was scrapped. Instead, Intel plans to go straight to Falcon Shores for his HPC GPU needs.
These cancellations also affect Intel’s Flex line of server GPUs for cloud gaming and media encoding, with Intel ditching Lancaster Sound (aka “Next Sound”) in favor of its successor, Melville Sound. Development of Melville Sound will accelerate, but Intel hasn’t given a firm date.
Intel didn’t provide new visuals for its new roadmap, but kitbashed something based on the SC22 slides to show where things stand and which products have been cancelled.
According to Intel, these product changes are designed to help the company align with its two-year product release cycle for server GPUs. His rivals NVIDIA and AMD have operated in a similar cadence over the past few years, which means that Intel’s product cycle will more or less coincide with that of its competitors. As Intel puts it, this “meets customer expectations for new product introductions and gives us time to develop our ecosystem.”
Reading between the lines means Intel wasn’t confident in its ability to sell Rialto and Lancaster to its core customer base.whether it exclusively Details will stay in Intel’s halls for now if there is a rapid product release cycle or if there is anything more. But in any case, the company has essentially dropped plans to release a new server GPU product this year in order to focus on its successor.
If you’re wondering if this is a good or bad development for Intel, please note that this information was released Friday after 5pm ET. So Intel is clearly trying to fill the bad news here by releasing it over the weekend.
Falcon Shores will move from 2024 XPU to 2025 GPU
In addition to the Rialto Bridge and Lancaster Sound cancellations, another major update to come from Intel’s letter is that Intel has significantly refactored its plans for Falcon Shore. What was (and may still be) Intel’s first combined CPU + GPU offering for HPC has been reassigned to serve as Intel’s next-generation HPC GPU. This will have a significant impact on Intel’s product lineup and competitiveness.
Falcon Shores is another Intel announcement for 2022, revealed at the company’s 2022 investor conference in February. Broadly speaking, Falcon Shores was designed to be Intel’s first of his XPU products, chips that use a variety of computing architectures to best meet the execution needs of a single workload. This is achieved by using separate processor tiles. In practice, this mostly meant bundling his Xe GPU tiles and x86 CPU tiles onto his single chip, but the tiled approach leaves avenues for adding other types of accelerators. was also open. Intel has already implemented tile his server his chips in products such as Ponte Vecchio and the recently launched Sapphire Rapids XCC, so Falcon Shores will be able to put multiple types of accelerators on one chip. , was going to be the next step.
At the time, Falcon Shores was slated for a 2024 release that would be built on the “Angstrom Age Process”. As planned, Falcon Shores is Intel’s power play in the HPC industry, allowing them to offer products that combine GPUs and CPUs built on top of the latest 20A/18A process nodes, based on their latest architecture. , leading the HPC market and ahead of traditional GPU-based accelerators. Falcon Shores may still meet these goals, but not in 2024.
Instead, the Falcon Shores have been reassigned to serve as HPC GPU parts. Intel’s brief letter doesn’t mention any technical changes, but the implication is that Intel will build products out of GPU tiles only, rather than mixing CPU and GPU tiles in a single chip. It means to focus on and reduce what should be. Convert Intel’s XPU to a simpler GPU.
It goes without saying that with the demise of the Rialto Bridge, Intel will need a new HPC-class GPU offering if it wants to remain in the market. There’s no reason to think that Falcon Shores isn’t a good product now, but the fact that Intel delayed it until 2025 doesn’t bode well. , originally intended to launch in 2021 and compete with HPC GPUs of that era. So by not offering anything new until 2025, Intel is essentially moving away from his HPC GPU market. His next two years were anything but supercomputer triumphs. Intel’s long-term plans still call for the company to capture a sizeable portion of the highly profitable server GPU market.
To make matters worse, Intel has fallen further behind its competitors in shipping products like the XPU/APU.As confirmed by serve the home, Intel hasn’t given up on shipping Falcon Shores XPUs, but its GPU prioritization means such a product won’t launch in 2025. That means his XPUs in the Intel server class are now 2026 products at best, and longer. at the worst case.
AMD, on the other hand, is expected to ship a server APU similar to the Instinct MI300 later this year. For its part, it employs multiple chiplets/tiles in a single chip to deliver HPC-class CPU and GPU performance in a single chip, achieving AMD’s long-awaited goal. NVIDIA is also a bit behind in integrating with the Grace Hopper superchip (which is basically two tightly coupled chips on one board instead of chiplets on one chip) More advanced than current Intel. And, worryingly for Intel, it’s almost certain that the next-gen successor for that part will launch before the Falcon Shores XPU hits the market.
In other words, delaying the server’s GPU schedule caused all Intel products that depended on it to delay as well. This will keep Intel’s server CPUs in a line of their own for years to come.
The history of server GPU cancellation Intel continues to challenge
Ultimately, the Rialto Bridge cancellation is the latest in a surprisingly long line of canceled GPUs and GPU-like accelerators at Intel. The company has been trying to tap into the accelerator market in one form or another since the late 2000s with projects such as his Larrabee. And now, 15 years later, Intel is still not as far along as it would like to be, and is still canceling chips to get there.
As the latest Intel server GPU victim, the Rialto Bridge joins not only Larrabee, but Intel’s ill-fated Xeon Phi lineup. And even then, the Rialto Bridge cannot claim to be the first canned Xe architecture component. That honor goes to Xe-HP. Xe-HP has been canceled in favor of Intel’s use of mixed computing and graphics Xe-HPG silicon.
At least the silver lining in this situation is that Intel isn’t canceling an entire architecture here. The Xe architecture is firmly established as Intel’s in-house GPU architecture, and the fact that it is found in everything from SoCs to video cards to HPC accelerators underscores its viability and importance at Intel. I’m here. The cancellation of Rialto Bridge and the delay of Falcon Shores are still significant setbacks for Intel, but all it does is shelve one iteration of his broader Xe architecture for another.
In the meantime, Intel will continue its efforts to expand its GPU offerings and enter the HPC GPU market in a bigger way. This is a market too big and too profitable for Intel to ignore. At the very least, it eats up more server CPU sales.
sauce: intel
