Intel Makes Cuts To Data Center GPU Road Map, Delays Falcon Shores To 2025
With Rialto Bridge canceled, the next GPU in the Max Series will now be Falcon Shores, a new “flexible chiplet-based architecture” which Intel said will launch in 2025 after originally promising a 2024 release window last year. Versions of Falcon Shores that include CPU cores will arrive after 2025.
Intel has made cuts to its data center GPU road map and delayed the introduction of its ambitious Falcon Shores processor for high-performance computing by one year.
In a late Friday announcement, the Santa Clara, California-based company said it has canceled Rialto Bridge, the successor to its first-generation Max Series GPU for HPC and AI applications, which launched in January after years of anticipation and previously carried the code name Ponte Vecchio.
[Related: Intel Vows To Restore Staff Salaries In Fall As EMEA Head Resigns]
Intel previously expected to start sampling Rialto Bridge with customers in mid-2023.
With Rialto Bridge canceled, the next GPU in the Max Series will now be Falcon Shores, a new “flexible chiplet-based architecture” which Intel said will launch in 2025 after originally promising a 2024 release window last year. Versions of Falcon Shores that include CPU cores will arrive after 2025.
Intel said it has also discontinued development of Lancaster Sound, the successor to its first-generation Flex Series GPU, which launched last year and is geared towards video transcoding, cloud gaming, virtual desktop infrastructure and AI inference workloads.
This means the Flex Series, like the Max Series, will skip a generation and move to Melville Sound, a GPU that will have a “significant improvement” in performance, features, and covered workloads compared to the first generation, previously known as Arctic Sound-M. This contrasts with the “incremental improvement” Lancaster Sound would have provided.
Lancaster Sound was previously slated in launch in 2023 or later. While Intel did not provide a time frame for Melville Sound, the GPU is likely to launch in 2024 since the company said it will release data center GPUs every two years from now on.
Jeff McVeigh, interim general manager of Intel’s Accelerating Computing Systems and Graphics group, said the company decided to move to a two-year cadence of new data center GPUs to match customer expectations and allow more time for ecosystems to build around the products.
This, he added, will maximize the “return on investment for customers.”
McVeigh said Intel is also “increasing” its focus on the software ecosystem for the Max Series and Flex Series GPUs. This will translate into “continuous updates” for both product lines as well as “performance improvements, new features, expanded operating systems support and new use cases.”
“We have simplified our road map with the goal of doing fewer things better and are rapidly rolling out products to our customers,” he said.
McVeigh didn’t say whether the cancellation of the Rialto Bridge and Lancaster Sound GPUs was influenced by Intel’s plan to cut costs by up to $23 billion through 2023.
Intel Reels Back Expectations For Falcon Shores
Intel has previously said that Falcon Shores would bring together x86 CPU, Xe GPU cores and memory into one chip package that can fit into a single server socket, which is why the company first dubbed it an “XPU.” The company had also said that Falcon Shores would be offered in various configurations of CPU and GPU cores, with some versions even containing only CPU cores or GPU cores.
However, while Intel’s vision for Falcon Shores as a “flexible chiplet-based architecture” has not changed, the company clarified to server news outlet ServeTheHome after Friday’s announcement that Falcon Shores will be available in a GPU-only configuration in 2025 and that variants with CPU cores will come later.
The company said it also still plans to develop configurations of Falcon Shores that contain technologies designed by other companies, but those iterations will arrive after 2025 too.
“We are working on variants for this architecture supporting AI, HPC and the convergence of these markets,” McVeigh wrote.
AMD And Nvidia To Intro CPU-GPU Hybrid Chips This Year
When Intel first announced Falcon Shores in May of 2022, the company said the processor would provide five times greater performance-per-watt, memory capacity and memory capacity than platforms that were available at the time, likely meaning Nvidia A100 GPUs paired with third-generation Intel Xeon Scalable or third-generation AMD EPYC CPUs.
The chipmaker also promised that Falcon Shores, one that is likely equipped with CPU cores, would feature five times more compute density in an x86 socket than the best CPU available at the time.
Rivals AMD and Nvidia have been planning to release their own CPU-GPU hybrid chips for HPC and AI applications. These chips are now set to beat Intel’s Falcon Shores XPU to the market by three years because both competitors plan to debut them this year.
AMD said in January that its Instinct MI300 chip, which combines CDNA 3 GPU cores, Zen 4 CPU cores and memory in a single integrated design, is set to arrive in the second half of this year.
Meanwhile, Nvidia plans to release the Grace Hopper Superchip, which combines a Hopper-based GPU, Arm-based CPU and memory on a single package, in the first half of this year.
Partner Says Road Map Changes Create Uncertainty
Alexey Stolyar, CTO of International Computer Concepts, a Northbrook, Ill.-based Intel system integration partner, said he’s disappointed but not surprised by Intel’s GPU road map changes, given the deep cuts the company has been making and its struggles with delivering products on time.
He said Intel has had issues in recent years following through on road map promises with its server CPUs, most recently the fourth-generation Xeon Scalable processors, code-named Sapphire Rapids which launched in January after roughly a year of delays.
“They can’t release timely products and stick to what they say,” he told CRN.
With Intel now cancelling two data center GPUs that were due this year or next, the company is creating uncertainty for future products on the road map, according to Stolyar.
“Even if they have a great product, customers are going to be like, ‘Is it credible? Is it going to be around for the next generation?’” he said.
The other challenge Intel faces is convincing developers to move away from CUDA, Nvidia’s parallel computing platform that has allowed it to dominate the accelerated computing space, Stolyar added.
“I think they’re not going to be able to do as much as Nvidia does — and quickly. So I think they need to pick specific areas where they want to be successful and focus on that,” he said.