A star engineer, Raja Koduri is one of the Intel executives who tends to reveal development progress of upcoming products via social media. A couple of months ago, he announced that the first GPU based on Intel's Intel's Xe HPG architecture had powered on, and this week, he seemingly teased the bring-up process of Intel's upcoming entrant into the best gaming graphics card race.
From 2012 to 2021 - same Intel Folsom lab, many of the same engineers with more grey hair , I was at Apple back then, getting hands on with pre-production crystalwell, 9 years later playing with a GPU that’s >20x faster! pic.twitter.com/RgmRJuhOXwMarch 12, 2021
"From 2012 to 2021 — same Intel Folsom lab, many of the same engineers with more grey hair," Raja Koduri, general manager of Architecture, Graphics, and Software at Intel, wrote in a Twitter post Thursday. "I was at Apple back then, getting hands on with pre-production Crystalwell, nine years later playing with a GPU that is >20x faster!"
The two pictures Koduri included show him running some tests on two development systems: one based on Intel's Haswell with Iris Pro 5200 graphics equipped, a 64MB eDRAM package (Crystalwell) from 2012 and another powered by an upcoming Intel Xe. He didn't specify the GPU as Xe HPG specifically, but the running of 3DMark, as well as the video outputs point toward the gaming-focused GPU.
The second image partly reveals the Xe bring-up board. Such boards are designed to provide maximum flexibility in terms of GPU configuration and power delivery, so while they have display outputs, they do not look like graphics cards at all. That said, it's not surprising to see the Xe HPG development board come with a cooling system that looks like it belongs with a server CPU.
It's also interesting to see Koduri says that the Xe GPU was over 20 times faster than Intel's Iris Pro 5200 from 2013. Of course, an upcoming discrete graphics processor should be an order of magnitude faster than an eight-year-old integrated GPU. And ">20x faster" could mean a range of things (not that we would expect Koduri to share performance numbers at this stage).
For some comparison, Intel's Iris Pro 5200 with 40 execution units (EUs) and 128MB of eDRAM scores 1,426 graphics points in 3DMark FireStrike. By contrast, the latest Intel Iris Xe G7 integrated GPU with 96 EUs scores between 5,800 and 5,900 graphics points, making it over four times faster than its ancestor in said benchmark.
Modern discrete graphics cards, such as Nvidia's GeForce RTX 3060 Ti, score around 31,000 graphics points in 3DMark FireStrike, so they are, indeed, over 20 times faster than Intel's Iris Pro 5200. Meanwhile, Nvidia's top-of-the-range GeForce RTX 3090 scores between 52,000 and 53,000 graphics points in 3DMark Firestrike (37 times faster than the Iris Pro 5200).
We can't draw any firm conclusions on Intel Xe HPG's performance based on Koduri's tweet. But if by over 20 times faster, the exec meant something close to 20 times faster, then we can expect the GPU to compete against products like the GeForce RTX 3060 Ti. We'll have to wait for much more information to see.
Last year I would have laughed at the idea of replacing my 1070 with a Xe GPU. After months of trying, and failing, to acquire a new Nvidia or AMD GPU, I'd be thrilled to have a Xe.
Not true. Remember, Intel isn't beholden to anyone else for fab capacity and has a worldwide network of advanced fab facilities. Also, our D1 facility is the biggest cleanroom in the world by a good amount and can crank out silicon like you wouldn't believe. Now that Pat is back and has taken some steps to right the ship, everyone else is on notice.
Also, don't forget to put a cover sheet on your TPS reports.
They could use their node only for the workstation GPUs where the lowest power draw makes the most sense and use their own nodes for the mass of their production.
They are just about to release a backported CPU, it would make sense for their GPU to be designed to be node agnostic.
Also we already saw something like that with nvidia GPUs if I'm not mistaken were they have GPUs made from TSMC and samsung.
Oh you sweet, summer child.
Intel claims to have Xe-HPC in the lab. The process for the 41 chiplets are labeled in one of Raja's twitter threads. According to the labels, the compute chiplets being used are Intel's 7nm process.