Intel’s Raja Koduri Teases Even Larger Xe GPU Silicon
by Ryan Smith on June 25, 2020 10:00 AM EST- Posted in
- GPUs
- Intel
- Deep Learning
- Xe
- Xe-HP
Absent from the discrete GPU space for over 20 years, this year Intel is set to see the first fruits from their labors to re-enter that market. The company has been developing their new Xe family of GPUs for a few years now, and the first products are finally set to arrive in the coming months with the Xe-LP-based DG1 discrete GPU, as well as Tiger Lake’s integrated GPU, kicking off the Xe GPU era for Intel.
But those first Xe-LP products are just the tip of a much larger iceberg. Intending to develop a comprehensive top-to-bottom GPU product stack, Intel is also working on GPUs optimized for the high-power discrete market (Xe-HP), as well as the high-performance computing market (Xe-HPC).
That high end of the market, in turn, is arguably the most important of the three segments for Intel, as well as being the riskiest. The server-class GPUs will be responsible for broadening Intel’s lucrative server business beyond CPUs, along with fending off NVIDIA and other GPU/accelerator rivals, who in the last few years have ridden the deep learning wave to booming profits and market shares that increasingly threaten Intel’s traditional market dominance. The server market is also the riskiest market, due to the high-stakes nature of the hardware: the only thing bigger than the profits are the chips, and thus the costs to enter the market. So under the watchful eye of Raja Koduri, Intel’s GPU guru, the company is gearing up to stage a major assault into the GPU space.
That brings us to the matter of this week’s teaser. One of the benefits of being a (relatively) upstart rival in the GPU business is that Intel doesn’t have any current-generation products that they need to protect; without the risk of Osborning themselves, they’re free to talk about their upcoming products even well before they ship. So, as a bit of a savvy social media ham, Koduri has been posting occasional photos of Intel's Xe GPUs, as Intel brings them up in their labs.
BFP - big ‘fabulous’ package😀 pic.twitter.com/e0mwov1Ch1
— Raja Koduri (@Rajaontheedge) June 25, 2020
Today’s teaser from Koduri shows off a tray with three different Xe chips of different sizes. While detailed information about the Xe family is still limited, Intel has previously commented that the Xe-HPC-based Ponte Vecchio would be taking a chiplet route for the GPU, using multiple chiplets to build larger and more powerful designs. So while Koduri's tweets don't make it clear what specific GPUs we're looking at – if they're all part of the Xe-HP family or a mix of different families – the photo is an interesting hint that Intel may be looking at a wider use of chiplets, as the larger chip sizes roughly correlate to 1x2 and 2x2 configurations of the smallest chip.
And with presumably multiple chiplets under the hood, the resulting chips are quite sizable. With a helpful AA battery in the photo for reference, we can see that the smaller packages are around 50mm wide, while the largest package is easily approaching 85mm on a side. (For refence, an Intel desktop CPU is around 37.5mm x 37.5mm).
Finally, in a separate tweet, Koduri quickly talks about performance: “And..they let me hold peta ops in my palm(almost:)!” Koduri doesn’t go into any detail about the numeric format involved – an important qualifier when talking about compute throughput on GPUs that can process lower-precision formats at higher rates – but we’ll be generous and assume INT8 operations. INT8 has become a fairly popular format for deep learning inference, as the integer format offers great performance for neural nets that don’t need high precision. NVIDIA’s A100 accelerator, for reference, tops out at 0.624 PetaOPs for regular tensor operations, or 1.248 PetaOps for a sparse matrix.
And that is the latest on Xe. With the higher-end discrete parts likely not shipping until later in 2021, this is likely not going to be the last word from Intel and Koduri on their first modern family of discrete GPUs.
Update: A previous version of the article called the large chip Ponte Vecchio, Intel's Xe-HPC flagship. We have since come to understand that the silicon we're seeing is likely not Ponte Vecchio, making it likely to be something Xe-HP based
Source: Intel/Raja Koduri
37 Comments
View All Comments
IntelUser2000 - Thursday, June 25, 2020 - link
Those are packages, not dies.The smallest one is about the size of the socket on your desktop i7 motherboard.
jbrukardt - Thursday, June 25, 2020 - link
Its chiplets, so not nearly as scary as the big 800 mm2 phi daysJayNor - Friday, June 26, 2020 - link
The tag on the board in the lab says DG1.yeeeeman - Friday, June 26, 2020 - link
It is not one architecture. There are three, Xe-LP, for lower power graphics usages, Xe-HP for gaming and workstations and Xe-HPC which is for GPGPU, like CDNA and Ampere A100.extide - Friday, June 26, 2020 - link
Yeah but it's really not just literally one architecture. It's one base architecture tuned to three different markets -- which is pretty much exactly how nvidia does it. That could mean different ratios/balances of different units in the HPC/HP/LP versions.I just hope Intel knows they need to make a splash here and freaking brings it. I am tired of this nvidia monopoly we've had for a while. (On that matter I hope AMD knocks NAVI2 cards out of the park as well!)
domboy - Thursday, June 25, 2020 - link
"BFP - big 'fabulous' package". Riiight.... I think we all know what they really say off social media, and it's not 'fabulous'... ;)Deicidium369 - Thursday, June 25, 2020 - link
The F in SpaceX's BFR was Falcon, right?quadra - Thursday, June 25, 2020 - link
I saw the code name and chuckled. Sure, I love the beautiful Ponte Vecchio over the River Arno in Florence, Italy, especially as a student of art history.But I was thinking of Intel’s code names: Sandy Bridge, Ivy Bridge...do you know what the literal translation of Ponte Vecchio is from the Italian?
Old Bridge.
:)
Rookierookie - Sunday, June 28, 2020 - link
Now you know where all that R&D budget went.rrinker - Thursday, June 25, 2020 - link
Two references to Osborning in as many days!