Intel has introduced a new GPU, codenamed 'Crescent Island' and now designated as the D-Series, that features up to 480 GB of LPDDR5X memory. This release comes at a time when consumer RAM is already scarce, potentially exacerbating supply concerns for the broader market.
The new GPU is built on Intel's Xe3P architecture and combines its substantial memory capacity with 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute, all within a 350-watt power envelope. Intel states this configuration is optimized for real-world inference workloads, emphasizing performance, memory capacity, and power efficiency.
Intel aims for the D-Series GPU to simplify deployment in existing data center environments and support portability across its GPU platforms. This strategic shift targets AI inference, a crucial aspect of AI deployment, rather than AI training, a move that follows past challenges with its Gaudi and 'Falcon Shores' GPU lines which did not significantly challenge Nvidia's market position.
Anil Nanduri, Intel's VP of AI Products and GTM, highlighted the focus on inference, stating, "Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost." He added that the Intel GPU D-series is designed to make high-capacity inference practical and economical to scale.
For comparison, Intel's previous Gaudi 3 accelerator featured 128 GB of unified HBM. While the D-Series uses LPDDR5X memory, which is generally slower than HBM but more power-efficient, its 480 GB capacity is a significant increase. Intel indicates that target applications include multi-tenant AI services, robotics, and other latency-sensitive workloads.
The substantial memory on Intel's new GPU for AI inference raises concerns about its impact on the availability of memory for consumer products, including gaming PCs. Reports suggest that even next-generation consumer GPUs may be earmarked for AI applications, potentially impacting pricing and availability for gamers.