When Intel announced its Xe family of graphics processors in 2020, it introduced as many as four microarchitectures, which eventually translated into major delays, cancellation of Xe-HP datacenter GPUs and issues with drivers. The company now says it has learned its lesson; with its Xe 2 ‘Battlemage’ family it will offer fewer microarchitectures. Still, there will be Xe2-LPG and Xe2-HPG microarchitectures for different kinds of GPUs.
“There is a Xe and there is a Xe 2 and in that Xe 2 generation there is a Xe-LPG and there is a HPG (…) and there a slight variations (…) which is our big learning,” said Tom Peterson, an Intel Fellow, in an interview with Hardwareluxx. “The idea was we needed to optimize for each segment and build separate chips and do separate verifications. And I think now the real learning is we would be better off concentrating our focus and really thinking of it like a really solidly, hard IP business.”
On its Battlemage generation of GPUs, Intel will stick to Xe2-LPG and Xe2-HPG microarchitectures. It is unclear whether Xe2-HPC is still planned, but it officially Intel’s next-generation GPU for HPC codenamed Realto Bridge is based on ‘Enhanced Xe-HPC’ cores, not on Xe2-HPC cores.
“Now as we go forward in our roadmap, we realized this is a very, very expensive – the QA process and the segmentation. The Thinking was we needed to differentiate our IP and customize it per each segment,” said Peterson. “[…] We are going to just have one thing and it goes everywhere unmodified. That’s more the strategy we are looking at going forward. And that’s because, that’s really the only way to get IP reused to really work.”
Intel’s original Xe family included four microarchitectures: Xe-LP for integrated and low-end standalone GPUs, Xe-HPG for discrete desktop graphics cards, Xe-HP for cloud datacenters and Xe-HPC for high-performance computing. Developing vastly different GPU microarchitectures has its advantages when it comes to performance and die sizes. For example, a slightly smaller iGPU that lacks features like ray tracing translates into hundreds of millions of dollars of cost savings as Intel sells boatload of client CPUs. Meanwhile, redesigned Xe-HPC cores translate into performance advantages.
But to deliver on its promise, Intel needed to design, verify, and produce as many as nine different GPU variations, which is a pretty enormous number for a single product family even for a giant like Intel.
Eventually, the company had to axe its Xe-HP because it decided that the datacenter vertical could be addressed by Xe-HPG and Xe-HPC GPUs instead. But the company lost precious time developing Xe-HP hardware and software stack. Furthermore, because Xe-LP and Xe-HPG are so different, the company had to tune integrated and standalone GPUs separately and then develop different drivers for built-in and discrete graphics processors. As a result, while Xe-LP iGPUs arrived in time in 2020, Xe-HPG and Xe-HPC GPUs were at least a year late.
While the company says it has learnt its lessons with the first generation of its Xe graphics processors, it remains to be seen whether its Xe2 Battlemage GPUs will come out in time to compete against the best graphics cards from AMD and Nvidia.
Source link
