Intel discontinues first-generation Max GPUs to prioritize next-generation Gaudi and Falcon Shores architectures.

Intel is sunsetting its first-generation Max Series GPU, previously code-named Ponte Vecchio, to focus on selling its Gaudi AI accelerator chips and developing its next-gen Falcon Shores GPU, according to a Tuesday report by industry publication Serve The Home and a source familiar with the matter who confirmed key details to CRN. An Intel spokesperson confirmed on Wednesday that the company is focusing on selling its Gaudi chips and developing Falcon Shores, which is expected to launch in 2025 and will combine Gaudi technology with the Xe GPU architecture that was used in Ponte Vecchio. The company is gearing up to launch Gaudi 3 later this year as a more powerful and efficient solution over Gaudi 2, which debuted in 2023.

Intel stated it is "focused on fulfilling the rapid expansion of the Intel Gaudi AI accelerator—capitalizing on its proven performance edge and competitive pricing" to meet growing demand for enterprise AI. The spokesperson noted that the Max GPU remains available on Intel Tiber Developer Cloud (recently renamed from Intel Developer Cloud) and that "Intel Xeon remains the host CPU choice for HPC solutions along with our Max GPU available on Intel Developer Cloud." The company indicated that this approach "will pave the path for developer and ecosystem readiness for Falcon Shores, our next-generation GPU for AI and HPC which will leverage the Xe IP architecture that is foundational to the Max Series GPU."

The Max Series GPU, launched as Ponte Vecchio at the beginning of last year, arrived after roughly two years of delays caused in part by issues with Intel's 7-nanometer manufacturing process that emerged in 2020. This delay also affected the delivery of the Aurora exascale supercomputer to the U.S. Department of Energy's Argonne National Laboratory. According to Serve The Home, Intel has stopped selling Ponte Vecchio into new server clusters and is only fulfilling orders for existing clusters, including for Intel Tiber Developer Cloud. After getting outfitted last year with more than 10,000 server blades of Max Series GPUs and CPUs from Intel, Aurora recently surpassed 1 exaflop of performance for the LINPACK benchmark, as confirmed this week by Top500, the independent organization that tracks the world's fastest supercomputers.

Intel originally planned to release a GPU successor to Ponte Vecchio called Rialto Bridge with customers in mid-2023, but announced in March of that year that it had decided to cancel Rialto Bridge to focus on Falcon Shores for a 2025 launch. The company initially planned a 2024 release window for Falcon Shores, which was conceived as a "flexible chiplet-based architecture" that would include versions with CPU cores in addition to GPU cores. In May 2023, Intel announced it would no longer include a CPU-core variant and would incorporate technology from its Gaudi accelerator chips into Falcon Shores as part of a decision to converge the Gaudi chip and Max Series GPU road maps, making Falcon Shores the successor to both Ponte Vecchio and Gaudi 3.

At the Intel Vision event in March, the company claimed that the upcoming Gaudi 3 "can best Nvidia's powerful H100 GPU for training large language models and offer similar or, in some cases, better performance than the rival's memory-rich H200 for large language model inferencing." The chip is set to debut in the third quarter, with support planned from Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro. During its latest earnings call in April, Intel CEO Pat Gelsinger said the company expects to make more than $500 million in revenue from Gaudi 3 sales in the second half of the year. Two months earlier, Gelsinger said the company had built a sales pipeline of more than $2 billion for its accelerator chips, which included Gaudi 2 and 3 as well as the Max and Flex Series GPUs. These figures are significantly smaller compared to Nvidia's $47.5 billion in data center revenue last year, though AMD expects to make $4 billion from data center GPUs this year thanks to its recently launched Instinct MI300 accelerator chips, which CEO Lisa Su called its "fastest-ramping product" ever.