When Intel introduced its Xe household of graphics processors in 2020, it launched as many as 4 microarchitectures, which finally translated into main delays, cancellation of Xe-HP datacenter GPUs and points with drivers. The corporate now says it has discovered its lesson; with its Xe 2 ‘Battlemage’ household it can supply fewer microarchitectures. Nonetheless, there shall be Xe2-LPG and Xe2-HPG microarchitectures for various sorts of GPUs.
“There’s a Xe and there’s a Xe 2 and in that Xe 2 technology there’s a Xe-LPG and there’s a HPG (…) and there a slight variations (…) which is our massive studying,” mentioned Tom Peterson, an Intel Fellow, in an interview with Hardwareluxx. “The concept was we would have liked to optimize for every section and construct separate chips and do separate verifications. And I believe now the true studying is we’d be higher off concentrating our focus and actually pondering of it like a extremely solidly, laborious IP enterprise.”
On its Battlemage technology of GPUs, Intel will keep on with Xe2-LPG and Xe2-HPG microarchitectures. It’s unclear whether or not Xe2-HPC continues to be deliberate, however it formally Intel’s next-generation GPU for HPC codenamed Realto Bridge is predicated on ‘Enhanced Xe-HPC’ cores, not on Xe2-HPC cores.
“Now as we go ahead in our roadmap, we realized it is a very, very costly – the QA course of and the segmentation. The Considering was we would have liked to distinguish our IP and customise it per every section,” mentioned Peterson. “[…] We’re going to simply have one factor and it goes in all places unmodified. That is extra the technique we’re taking a look at going ahead. And that is as a result of, that is actually the one strategy to get IP reused to essentially work.”
Intel’s unique Xe household included 4 microarchitectures: Xe-LP for built-in and low-end standalone GPUs, Xe-HPG for discrete desktop graphics playing cards, Xe-HP for cloud datacenters and Xe-HPC for high-performance computing. Creating vastly completely different GPU microarchitectures has its benefits relating to efficiency and die sizes. For instance, a barely smaller iGPU that lacks options like ray tracing interprets into a whole lot of thousands and thousands of {dollars} of price financial savings as Intel sells boatload of consumer CPUs. In the meantime, redesigned Xe-HPC cores translate into efficiency benefits.
However to ship on its promise, Intel wanted to design, confirm, and produce as many as 9 completely different GPU variations, which is a fairly monumental quantity for a single product household even for a large like Intel.
Finally, the corporate needed to axe its Xe-HP as a result of it determined that the datacenter vertical might be addressed by Xe-HPG and Xe-HPC GPUs as an alternative. However the firm misplaced valuable time growing Xe-HP {hardware} and software program stack. Moreover, as a result of Xe-LP and Xe-HPG are so completely different, the corporate needed to tune built-in and standalone GPUs individually after which develop completely different drivers for built-in and discrete graphics processors. Consequently, whereas Xe-LP iGPUs arrived in time in 2020, Xe-HPG and Xe-HPC GPUs had been not less than a yr late.
Whereas the corporate says it has learnt its classes with the primary technology of its Xe graphics processors, it stays to be seen whether or not its Xe2 Battlemage GPUs will come out in time to compete towards the finest graphics playing cards from AMD and Nvidia.