Update (03/06/23): Intel confirmed that the first Falcon Shores product would be a GPU and would not integrate CPU chiplets on-package. The story has been updated to clarify that point.

Intel CEO Pat Gelsinger is taking a no-holds-barred approach to cutting costs as he whips the company back into financial shape.

Intel has already exited seven businesses, and recently made wholesale graphics processors changes by axing products and changing its enterprise GPU roadmap.

Intel has scrapped a supercomputer GPU codenamed Rialto Bridge, which was advertised as the successor to its current Max Series GPU codenamed Ponte Vecchio.

“Rialto Bridge, which was intended to provide incremental improvements over our current architecture, will be discontinued,” said Jeff McVeigh, corporate vice president and interim general manager of the Accelerated Computing Systems and Graphics group at Intel, in a blog entry on Intel’s website.

The GPU was considered important to Intel’s expansion into the high-performance computing and AI markets. The cuts demonstrate Gelsinger is taking no prisoners as he tries to reverse Intel’s fortunes following a 60% decline in profits and 20% decline in revenue last year.

The carefully worded blog entry credited to McVeigh warms up readers with Intel’s progress in GPUs, which is a new business. It also sows confusion on Intel’s supercomputing product plans, and leaves many questions unanswered on the company’s chiplet strategy.

HPCwire has reached out to Intel with a request to interview McVeigh to clarify and provide more insight on the product changes. But here is what we know.

Intel has set a two-year release cycle for its GPUs, a change from the plans to release new datacenter GPUs on a faster cadence. Rialto Bridge was announced last year, and the first units were expected to ship in 2024.

“Building on the momentum of the Max Series GPU, our next product in the Max Series family will be the GPU architecture codenamed Falcon Shores,” McVeigh wrote.

The Falcon Shores GPU will be released in 2025, and its “flexible chiplet-based architecture will address the exponential growth of computing needs for HPC and AI,” McVeigh wrote.

Falcon Shores was originally due next year, but Intel is biding its time for a more revolutionary performance jump. The chip represents an inflection point in the way Intel builds and manufactures chips.

Rialto Bridge – built on an older manufacturing process and chip constructs – would have provided only incremental performance improvements over Ponte Vecchio.

While Intel had previously projected Falcon Shores as the first in a new family of chips that would integrate GPU and CPU cores using new chiplet packaging technologies, for now, it’s only a GPU, and there’s no CPU component to it in 2025.

Intel executives have said Falcon Shores will deliver five times or more improvements in performance-per-watt, compute density, memory capacity and bandwidth improvement. The chiplet design of the XPU will also allow Intel to tack on additional accelerators for applications like artificial intelligence, which is expected to play a significant role in supercomputing.

The 2025 release timing for the chip will come at just around the time Intel is expected to complete a multi-year restructuring effort to become a manufacturing-first company.

Gelsinger is undoing broken processes that caused Intel to lose its manufacturing leadership. Intel has cut product lines and laid off employees with the aim to save $10 billion by 2025.

The restructuring is designed to help Intel regain manufacturing leadership by jumping four manufacturing upgrades over the next five years.

The jump straight to Falcon Shores in 2025 is in line with Intel’s desire to get a chipmaking and design edge over its rivals with the “Angstrom-era” manufacturing processes.

The chipmaker planned a larger Falcon Shores family of products where non-x86 CPUs from Arm and RISC-V designs could be incorporated alongside other chiplets.

Intel scrapped its investments in RISC-V projects earlier this year as part of its cost-cutting measures, but the chipmaker is expected to maintain support for non-x86 architecture in its chips.

The company will make chips for Qualcomm – which designs Arm-based processors – on the Angstrom era processes, which use new transistor designs and power-delivery mechanisms.

“The first Falcon Shores products are going to be all Intel IP. But future incarnations of it – we envision opportunities where they are custom IP, where you could replace the x86 chiplets with a different CPU architecture,” McVeigh previously told HPCwire at the Supercomputing Conference in 2022.

Intel surprised analysts by reporting unexpectedly high consumer and datacenter GPU revenue in the most recent earnings report. The earnings were more than meaningful, considering the chipmaker just entered the GPU market, said Dean McCarron, principal analyst at Mercury Research.

But the first quarter for the datacenter GPU market was “otherwise really, really ugly looking” as customers tightened their purse strings in a weak purchasing environment, McCarron said.

He assessed that the GPU market may turn in the second half of this year, but much of that depended on the economic environment. The widening impact of a weak macroeconomic environment may have made it easier for Intel to cancel Rialto Bridge.

AI is already having a profound impact on hardware decisions made by the top cloud companies. Microsoft and Google launched new search engines with ChatGPT-style ability to provide human-like answers to questions, which need high-performance chips. Facebook and Microsoft are arming up on high-performance GPUs.

The Rialto Bridge cancellation is bad news for customers that selected the GPU for their next HPC installation. One of those customers was Barcelona Supercomputing Centre, which planned to use the GPU in its upcoming MareNostrum 5 supercomputer.

The supercomputer, which is now being constructed, was to use Intel x86 CPUs and GPUs in one partition, and Nvidia’s GPUs and Arm-based CPUs in the other. It is unclear how BSC will substitute the Rialto Bridge GPUs.

Intel will have to start working with customers on GPU products like Falcon Shores now as the purchasing, testing, implementation, and scaling can span years, McCarron said.

Intel’s Ponte Vecchio is appearing in the supercomputer called Aurora, which has undergone multiple delays as Intel’s CPU and GPU cancellations and delivery dates slipped. The supercomputer, which is being installed at the Argonne National Laboratory, is expected to deliver more than two exaflops of (peak) performance.

Intel’s delay means one less competitor in the datacenter market for Nvidia and AMD, which dominate the space. AMD’s supercomputing APU called MI300 will be in another two-exaflop supercomputer called El Capitan.

Intel also chopped its upcoming datacenter GPU called Lancaster Sound, which was a successor to the Flex Series GPU (codenamed Arctic Sound) that started shipping last year. The chipmaker is also moving to a two-year upgrade cycle for the Flex GPUs, and the successor is now Melville Sound.

The cancellation of Lancaster Sound, which would have provided incremental improvements, helps dedicate more resources to the development of Melville Sound, “which will be a significant architectural leap from the current generation in terms of performance, features and the workloads it will enable,” McVeigh wrote.

The fine-tuning of Intel’s GPU roadmap does not affect the CPU roadmap, which includes Emerald Rapids, which will succeed Sapphire Rapids and is on track to ship later this year.

“Sierra Forest and Granite Rapids will build on that momentum in 2024 and are already showing good early process and product health,” Gelsinger said in a financial earnings call late last month.

Intel also has a set of additional AI accelerators with the AI training chips like Gaudi3, which is due this year, and its FPGAs. The chipmaker is also building incremental AI-acceleration units – mostly capable of inferencing – directly into its server chips. But GPUs will remain a mainstay for the company’s AI ambitions.

“When you view a broader GPU [market], we are gaining momentum with Ponte Vecchio in the marketplace now. For HPC and AI use cases … you’ll see us putting more emphasis on our GPU offerings over time,” Gelsinger said in an earnings call last month.