Intel's Xe HP GPU Chiplet Pictured - 3696mm^2 in Size!!!

[Djole123]

So, as you know, Intel is trying to finally get in to the competitive GPU market, and they're going to give it a shot with the Intel Xe GPU platform.

Raja Koduri has teased the new GPU as the "baap of all", which gives us hope that this new GPU is going to be a monster in power.

 

But now, we finally get to see the new beast, and it looks just as beastly as we hoped:

 

 

Being 3696mm^2 in size, we expect the new GPUs to be absolute powerhouses, just like the Xeon Phi chips it will superceed.

 

What do you guys think of this?

 

Source: https://wccftech.com/intels-father-of-all-xe-hp-gpu-pictured-massive-3696-mm-package/

[BobVonBob]

That is a very, very large chip that I am very, very disappointed is under a heat spreader instead of using bare die cooling like AMD and Nvidia.

[porina]

14 minutes ago, BobVonBob said:

That is a very, very large chip that I am very, very disappointed is under a heat spreader instead of using bare die cooling like AMD and Nvidia.

Look at Threadripper or Rome for something similar in package. In a quick search, the package of TR4 is about 4411mm^2, so TR4 is bigger than this. Given the size and package, this is going to be a server part, maybe high end workstation part. You're not going to find this anywhere near a consumer GPU.

 

I've dug out a couple of older slides that might help with context. Note the part of this thread may not necessarily be the ones I linked below, but you get the idea.

 

Spoiler

https://www.anandtech.com/show/15120/intels-2021-exascale-vision-in-aurora-two-sapphire-rapids-cpus-with-six-ponte-vecchio-gpus

 

https://www.anandtech.com/show/15188/analyzing-intels-discrete-xe-hpc-graphics-disclosure-ponte-vecchio/3

 

 

 

[porina]

1 minute ago, Tedny said:

this isn't GPU, just moded CPU,right?! 

We don't have the details yet, but in the consumer iteration it will do GPU things a consumer would expect a GPU to do. Like gaming. They want a scaleable architecture, so you could see this as Intel's GPU chiplets. 

 

22 minutes ago, Benji said:

Hmm... If it will supercede the Intel Xeon Phi, is it really a GPU? Because the Xeon Phi were apparently once intended to be used as some sort of "software-rendering GPU" (if I got that right), yet they were only released as extra computational processors.

Xeon Phi, at least in the later iterations, were essentially an array of Atom x86 cores each with a big FPU strapped on it. It is a potential monster at eating more complex FP64 workloads, but not graphical data as a gamer would see a GPU. 

 

Xe is described by Intel as a GPU, so we can assume it is more recognisable as something like AMD or nvidia might bring out for heavier professional workloads. They are already laying down the groundwork for consumer level applications of it, by ensuring the same software stack can scale up and down the whole chain. 

[NumLock21]

Imagine a GPU, where you can just swap it out like a CPU, that will be a neat idea. Some really old video cards back then, it does not let you swap the GPU, but it does let you add more ram.

[schwellmo92]

Misleading title, that’s the package size.

[williamcll]

But how big are the dies in that heat spreader?

[leadeater]

24 minutes ago, huilun02 said:

With the unit being of this size i dont see how it can be priced competitively against AMD or Nvidia

Can already hear idiots saying this is not for consumers

Yes with Intel they want to hype it up like its for consumers, and when it turns out to be dogshit price/perf they suddenly say its not for consumers

This packaged as it is definitely never was for consumer, maybe the chiplets inside of it and looks like 4 to me, but this never. If it was a BGA package maybe, but not this LGA package. Definitely for server systems and highly custom boards.

[Soppro]

Intel's foray into GPUs is going to be interesting .. I hope their mid-2020 schedule isn't significantly disrupted by the current situation 

[RejZoR]

22 hours ago, Benji said:

Hmm... If it will supercede the Intel Xeon Phi, is it really a GPU? Because the Xeon Phi were apparently once intended to be used as some sort of "software-rendering GPU" (if I got that right), yet they were only released as extra computational processors.

I would definitely be interested in seeing the full specs and seeing what it is capable of.

That was original LarraBee. Xeon Phi was low clock, high core count x86 chip meant for highly parallel general computation.

[NumLock21]

AMD is bigger

Spoiler

 

[Dylanc1500]

2 hours ago, NumLock21 said:

AMD is bigger

  Hide contents

 

Come on now, size doesn't matter, it's how you use it that matters.

[Maticks]

the waterblock will look awesome  just thinking of aesthetics. haha

This doesn't look consumer though.

[porina]

On 5/2/2020 at 6:14 AM, huilun02 said:

With the unit being of this size i dont see how it can be priced competitively against AMD or Nvidia

Can already hear idiots saying this is not for consumers

Yes with Intel they want to hype it up like its for consumers, and when it turns out to be dogshit price/perf they suddenly say its not for consumers

These are starting at the top, most likely because of a multiple reasons including that's where the bigger money is, and dealing with a few large customers is easier than millions of individuals. They have not indicated much about consumer level parts yet, other than they are planned. 

[SpaceGhostC2C]

On 5/1/2020 at 10:42 PM, BobVonBob said:

That is a very, very large chip that I am very, very disappointed is under a heat spreader instead of using bare die cooling like AMD and Nvidia.

Seeing the triangle in the corner and the pin pads, I must ask: do AMD or Nvidia sell socketed GPUs you manually install and remove from PCBs?

 

On 5/1/2020 at 10:46 PM, Benji said:

Hmm... If it will supercede the Intel Xeon Phi, is it really a GPU? Because the Xeon Phi were apparently once intended to be used as some sort of "software-rendering GPU" (if I got that right), yet they were only released as extra computational processors.

Xeon Phi are x86 processors, meant as an alternative to GPGPU for parallelization , since being x86 like the CPUs simplifies coding significantly, compared to using something like CUDA hooked here and there to the main code running on the CPU.

[3rrant]

1 hour ago, SpaceGhostC2C said:

Seeing the triangle in the corner and the pin pads, I must ask: do AMD or Nvidia sell socketed GPUs you manually install and remove from PCBs?

It is possible to make socketable GPUs, but this would create a lot of problems like fixed interfaces, pins count etc. In the past there have been GPUs that had something similiar to this for upgradable RAM, but a fully "plug-and-play" socketable GPUs never existed (that I know of at least) and isn't likely to in the future.

 

The triangle could be used to indicate the position to place on the PCB if the IHS doesn't have any recognizable naming on top of it (if it is blank like in the image, how do you know which side is which?). It really looks like a CPU honestly. See answer below.

[leadeater]

26 minutes ago, 3rrant said:

The triangle could be used to indicate the position to place on the PCB if the IHS doesn't have any recognizable naming on top of it (if it is blank like in the image, how do you know which side is which?). It really looks like a CPU honestly.

This is an LGA socket package, but not really a "GPU". Intel designs most of the server boards on the market that aren't HPE/Dell/Lenovo so they have the sway to create OEM/ODM spec designs that people will actually use so it's not going to be some weird crazy left fielding thing. This is actually less weird and custom than Nvidia's SMX2 GPUs

 

 

 

I won't be at all surprised if there is HBM under that IHS and all those pin contacts are just power, PCIe and multi-socket interconnects.

[RejZoR]

"Chiplet"

 

The "chiplet" is the size of Texas lol

[dalekphalm]

4 hours ago, RejZoR said:

"Chiplet"

 

The "chiplet" is the size of Texas lol

I’m under the impression that they are calling it a chiplet because it is likely multiple chips linked and scaled, similar to something like AMD’s CCX’s.

 

I could be wrong.

[RejZoR]

Just now, dalekphalm said:

I’m under the impression that they are calling it a chiplet because it is likely multiple chips linked and scaled, similar to something like AMD’s CCX’s.

 

I could be wrong.

Yeah, but that's not a chiplet. Chiplet is individual module that is a building block of the entire processor. One CCX on AMD's processors is a chiplet. 4x CCX makes a whole processor. Unless Intel is literally going the route of 3dfx if you remember it. There, each VSA100 chip was a chiplet so to speak. And they just stacked whatever number they needed on the PCB. But given these size of these, anyone would have problems stacking more than 1 on any kind of PCB, be it motherboard or graphic card format.

 

Looking at the backside of the entire thing, it looks like there are at least 4 units of something on that PCB. And i think those are the chiplets. Not the whole thing as it's presented in the title.

[porina]

2 hours ago, RejZoR said:

Yeah, but that's not a chiplet. Chiplet is individual module that is a building block of the entire processor. One CCX on AMD's processors is a chiplet. 4x CCX makes a whole processor.

One AMD chiplet (CCD) is up to 8 cores, in two CCX. At a minimum you need to pair one of those chiplets with the IOD to make a CPU package. You can have two CCD for up to 16 cores, and 4 or 8 CCDs in threadripper for up to 64 cores.

 

What intel is doing is the above taken up a notch. I don't know if there is a formal definition of what it takes to be considered a chiplet, but just one "GPU" chip from Intel isn't likely to be too useful in itself. Like AMD pairing their chiplet with IOD to form a CPU, Intel will be mixing and matching different types of chiplet as they need to make the overall product.

[SpaceGhostC2C]

11 hours ago, 3rrant said:

It is possible to make socketable GPUs, but this would create a lot of problems like fixed interfaces, pins count etc.

I don't àrticularly care for socketable GPUs oe way or another, I was just pointing out that what AMD or NVidia put o a graphics cards is a different animal than this, so the use of a heatspreader wasn't something to be particularly worried about, but something we are used to see in CPUs, probably for similar reasons.

 

[RejZoR]

14 hours ago, SpaceGhostC2C said:

I don't àrticularly care for socketable GPUs oe way or another, I was just pointing out that what AMD or NVidia put o a graphics cards is a different animal than this, so the use of a heatspreader wasn't something to be particularly worried about, but something we are used to see in CPUs, probably for similar reasons.

 

At this point Intel could solder CPU's on motherboards for all I care since they swap stupid socket with each release anyway... What's even the point?