This is oursecond roundof investigation into Xe, as quite a bit hastranspiredin the last few months.
That’s what this article is about.
AMD and Nvidia’s cores perform one operation per clock, but Intel’sexecution units(EUs) perform eight.
Apart from Intel’s need to build with eight bricks at a time, their construction techniques are straightforward.
They can throw a few bricks together and make a wall.
One apartment is all you need, so the low-power cards use just one slice.
In Gen11, Intel’s integrated GPUs had oneslicemade of eightsub-slices, which in turn had eightexecution unitseach.
In September, code accidentally uploaded to GitHubleakedthe configurations of DG1, Ponte Vecchio, and one DG2 variant.
This leak is reliable, as its counter-intuitive prediction that Ponte Vecchio will have two slices was proven correct.
Koduri laterrevealedthat Ponte Vecchio has two slices and sixteen CUs.
With two slices that means that Ponte Vecchio has 128 EUs, 1024 cores.
Note, the two-slice configuration may be just for prototypes.
Ponte Vecchio’s basic slice configuration is expected to be used across high-performance and low-power models as well.
“Xe HP … would easily be the largest silicon designed in India and amongst the largest anywhere.”
-Raja Koduri
Last July, Intel accidentally published adriver(thanks!)
that contained three DG2 codenames,iDG2HP128,iDG2HP256, andiDG2HP512.
That’s two, four and eight slices.
Not long after, however, we saw solidevidenceof a three-slice GPU with 1536 cores being developed as well.
DG2 will also be more than just gaming GPUs.
DG1: Low-Power
The low-power segment is defined as just that, 5W through to 50W.
5W to 20W for integrated GPUs, and 20W to 50W for discrete ones.
Intel has already introduced us to the first member of the LP family.
TheDG1 SDVwas prominently displayed at CES 2020, running Destiny 2 and Warframe with RGB and all.
But it’s only dressing up as a gaming card.
But where Xe LP might outshine them is in clock speeds.
That’s the same amount of performance as a GTX 1650.
Just imagine how efficient this processor must be.
The GTX 1650 has slightly fewer TFLOPs and has a 75W TDP: almost four times as much.
But the good stuff doesn’t stop there.
It’s an excellent way to squeeze more performance out of the same silicon.
That’s not the case for any data center GPU, andPonte Vecchioin particular.
Ponte Vecchio is all about the tricks and techniques that maximize efficiency.
Fun Fact:Koduri named Ponte Vecchio after the bridge in Florence because he likes the gelato there.
If it didn’t give you an indication, then I’ll spell it out: double precision.
Doing some back of a napkin math, that’s about 20 TFLOPs at FP64 per 1024 core card.
Second to high precision workloads, is, naturally, ultra-low precision work.
Ponte Vecchio supports INT8, BF16, and the usual FP8 and FP16 for AI neural internet processing.
However, none of that is particularly novel.
It’s unclear if that refers to a full release or an exclusive early launch for the Aurora supercomputer.
Software
Hardware is good and all, but completely useless without adequate software support.
The good news is Intel seems to be doing their best.
Intel isredesigningits lowest level of software, the instruction set architecture (ISA), for modern high-performance applications.
The encoding of almost every instruction field, hardware opcode and register jot down needs to be updated."
At the driver level, Intel has a long way to go but is making progress.
Intel’s consumer-facing software, on the other hand, is superb.
Driver control is pleasantly straightforward.
The Command Center is unique in providing advanced display controls as well.
I personally use it to control my system, despite running Nvidia hardware.
If that’s the case, then we can expect a release in the subsequent months.
The most likely candidate is June.
Last October, Koduri tweeted a not-so-subtlehintin the form of an image of his new numberplate.
It reads “Think Xe” and has a June 2020 date.
But until it’s not time we can’t be anything more than cautiously optimistic.
