Jeff McVeigh, Intel vice president and head of the Super Compute Group, said that the company is moving to a two-year plan for releasing data center GPUs and that Rialto Bridge series accelerators will not be released at all. This will make Falcon Shores accelerators the new Max graphics series for the data center, but will enter production in 2025, a year later than previously planned.
The high-end CPU/GPU-focused Falcon Shores XPUs were originally scheduled to launch in 2024, but have now been pushed back to 2025 and are debuting as graphics components. Intel also dropped the release of Lancaster Sound graphics as part of the Flex series – they were intended for less intensive tasks like encoding media files. It will be replaced by a new generation of Melville Sound products.
The company found that the new schedule meets customer expectations for data center graphics and is broadly in line with those of competing manufacturers, including NVIDIA. The decisions came after a reorganization of AXG’s graphics division to separate gaming and data center products. Intel also stated that it will now focus more on the software direction – updates for the Max and Flex series accelerators will be released more frequently and they will include performance improvements, new features and support for more operating systems.
Intel Falcon Shores XPU chips are direct competitors to NVIDIA Grace Hopper Superchips and AMD Instinct MI300 Data Center APUs. Products from NVIDIA and AMD combine CPU and GPU cores with HBM memory, a new type of architecture that offers significant advantages in high-performance computing and is difficult to integrate with existing components. However, year-old Falcon Shores products will come out as GPUs first, and the company hasn’t said when the CPU version will arrive — Intel will lag behind NVIDIA and AMD by more than a year, offering Xeon and AMD processors in the HPC segment Ponte Vecchio graphics.
Falcon Shores is a product based on a heterogeneous architecture that aims to achieve a set of goals: a five-fold increase in performance per watt, a five-fold increase in compute density in an x86 socket, and a five-fold increase in memory and bandwidth in server chips. Intel’s roadmap for high-end CPUs and GPUs aligns with Falcon Shores’ roadmap – these chips will fill both roles. The company will be able to offer customers central, graphics and hybrid processors according to their needs, which are expected to be produced according to Intel 20A standards, although the manufacturer has not specified this point.
Intel Rialto Bridge accelerators should follow the existing Ponte Vecchio this year, but with the former eliminated, the latter are the only thing the company has to match the NVIDIA Hopper H100. And this may seem unwise at a time of massive development of large AI voice models like ChatGPT, which are growing in popularity and attracting multi-billion dollar investments. However, the Rialto Bridge could hardly keep up with the extremely powerful NVIDIA Hopper H100: The existing Ponte Vecchio offer 128 Xe cores with a peak power of 600 W; while Rialto Bridge should offer up to 160 Xe cores, up to 30% performance boost and up to 800W peak power. After the main characteristics, we are talking about not too significant evolutionary changes.
Intel also said that work on the Xeon series chips is progressing according to plan and technological nodes are being updated. However, the processors of the Sierra Forest series, which are optimized for hyperscalers, will not appear until 2024, i.e. a year later than AMD Bergamo, so the blue camp will also lag behind the competitor in this discipline.