With companies, tech execs, and outlets such as yours truly telling everyone over and over again—sometimes very literally—that it's difficult to get your hands on memory right now and will be for some time, it can be a bit jarring to think about what the AI bigwigs are getting up to. I picture that scene where The Joker slides down a pile of cash, except it's DRAM. Anyway, case-in-point is a new AI-focused GPU from Intel that's rocking more than just a little memory.
Intel's new GPU D-Series, previously codenamed 'Crescent Island', crams up to 480 GB of LPDDR5X memory alongside its processor:
"Built on Intel's latest Xe3P architecture, the GPU combines up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute within a 350-watt power envelope. This balance of memory capacity, compute performance and power efficiency reflects Intel's approach to optimizing infrastructure for real-world inference workloads."
The GPU, Intel says, is "designed to simplify deployment within existing data center environments while supporting application portability across Intel GPU platforms."
Intel's previous attempts at AI GPUs, codenamed Gaudi, failed to compete with Nvidia to any significant degree, which led to the cancellation of its follow-up 'Falcon Shores' GPUs. This time around, though, Intel is more directly targeting AI inference workloads rather than training, too.
Intel's VP of AI Products and GTM Anil Nanduri explains: "Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost. With Intel GPU D-series, we're focused on making high-capacity inference practical to deploy and economical to scale."
For context on the memory front, Intel's previous Gaudi 3 chip features 128 GB of unified HBM. Obviously this isn't apples-to-apples, as there's also a switch in the kind of memory it's using, from HBM to slower but more efficient LPDDR5X memory. "Target applications," Intel says, "range from multi-tenant AI services to robotics and other latency-sensitive workloads."
Whatever the case, it's a lot more than we see in our gaming GPUs, and perhaps worryingly is closer to the kind of memory we see in our consumer devices—worrying, that is, because it signals yet more production lines that could have been tailored to us being tailored to AI use cases.
At least we might be getting more memory in the RTX 5080 next year. Except oh, wait, if that happens it'll be tailored for AI, too, and almost certainly priced accordingly. I'll just wait for all this to blow over.