Friday, May 7th 2021

Intel Xe-HPG DG2 GPU Specifications Leak, First GPUs are Coming in H2 2021 in Alder Lake-P Laptops

Yesterday, we got information that Intel's upcoming DG2 discrete graphics card is "right around the corner". That means that we are inching closer to the launch of Intel's discrete GPU offerings, and we are going to get another major player in the current GPU market duopoly. Today, however, we are in luck because Igor from Igor's LAB has managed to get ahold of the specifications of Intel's Xe-HPG DG2 graphics card. For starters, it is important to note that DG2 GPU will first come to laptops later this year. More precisely, laptops powered by Alder Lake-P processors will get paired with DG2 discrete GPU in the second half of 2021. The CPU and GPU will connect using the PCIe 4.0 x12 link as shown in the diagram below, where the GPU is paired with the Tiger Lake-H processor. The GPU has its subsystem that handles the IO as well.
Next up, we have the specifications of various SKUs. Intel has prepared as many as five SKUs, each differentiating in Execution Unit (EU) count and memory interface configuration. Starting with the biggest model, the 512 EU, 4096 core design, this SKU will be DG2's biggest model. It will feature 16 GB of GDDR6 VRAM, running at 16 Gbps speeds on a 256-bit memory bus. It has 16 MB of so-called smart cache present inside the GPU. While the desktop variant clock speeds and TGP are unknown, the base clock of mobile design is 1.1 GHz, with a boost speed of 1.8 GHz. The TGP of this GPU is rated at 100 Watts, which is the average for higher-end GPUs in laptops. This SKUs smaller brother is a design with 384 EUs and 3072 cores, up to 12 GB of GDDR6 VRAM on a 192-bit memory bus. This one has a base clock of just 600 MHz, while the boost speed goes up to 1.8 GHz. You can find the rest of the specifications in the table below, as well as to check out other SKUs.
Additionally, Igor's LAB got ahold of the PCB diagram that shows the board layout of DG2 in the laptop along with the CPU. The DG2 GPU is seated on the BGA2660 socket, which is also pictured below with the pin configuration.
Sources: Igor's LAB, via WCCFTech
Add your own comment

20 Comments on Intel Xe-HPG DG2 GPU Specifications Leak, First GPUs are Coming in H2 2021 in Alder Lake-P Laptops

#1
chris.london
Igor’s Lab believes that manufacturing will start in week 43 at the earliest so this doesn’t look like a 2021 launch.
Posted on Reply
#2
Tardian
SK1 specifications look hopeful?
Posted on Reply
#3
londiste
SKUs in the table do not make sense. At least when it comes to EUs, clock speeds and TDP.
Posted on Reply
#4
MDDB
The CPU to GPU link will be PCIe4 x12 : that sounds brilliant, no current GPU benefits from having full 16 lanes, so why implement it? It may help making things more stable and less power hungry, i imagine. How come no other vendor has done it before? Or have they?
Posted on Reply
#5
londiste
MDDBThe CPU to GPU link will be PCIe4 x12 : that sounds brilliant, no current GPU benefits from having full 16 lanes, so why implement it? It may help making things more stable and less power hungry, i imagine. How come no other vendor has done it before? Or have they?
In laptops GPU connected to CPU via PCIe x16 is an exception, not the rule. x8 is common and x4 is not that uncommon.

Also, Tiger Lake-H actually might be the first mobile CPU with PCIe 4.0. Renoir (and Cezanne) have PCIe 3.0.
Of course, both using narrower bus and 3.0 instead of 4.0 helps in laptops by lowering power consumption.
Posted on Reply
#6
ArdWar
MDDBThe CPU to GPU link will be PCIe4 x12 : that sounds brilliant, no current GPU benefits from having full 16 lanes, so why implement it? It may help making things more stable and less power hungry, i imagine. How come no other vendor has done it before? Or have they?
PCIe x12 lane still use x16 mechanical form, I guess it's just not worth the saving.
Posted on Reply
#7
londiste
ArdWarPCIe x12 lane still use x16 mechanical form, I guess it's just not worth the saving.
Pretty sure the use case with x12 is dies on PCB, not a PCIe port. In these cases the wider physical connection definitely is not used.
Posted on Reply
#9
ArdWar
londistePretty sure the use case with x12 is dies on PCB, not a PCIe port. In these cases the wider physical connection definitely is not used.
I mean it's not worth the saving compared to the marketing opportunity cost or whatnot. PCIe 5.0 PHY/Transceiver only consume less than 1W per lane anyway.

Very few of the gaming gears are "right sized" anyway, and we *do* bitch about even 1% performance difference.
Posted on Reply
#10
Zareek
I wonder why mobile first, is the architecture not competitive at higher TDPs? Perhaps it doesn't scale well. I hope it can be competitive and comes out this year. The sooner, the better this market is insane! They probably won't produce enough volume to make a difference anyway.
Posted on Reply
#11
londiste
ZareekI wonder why mobile first, is the architecture not competitive at higher TDPs? Perhaps it doesn't scale well.
Would suspect frequencies, not TDPs. Either does not clock too well or efficiency goes really bad. Mobile first probably has to do with synergy they are able to build, a laptop platform instead of a discrete GPU (plus better profit margins).

Besides, this is all rumors at this point anyway, they might be same far with dedicated cards but we just do not know about it yet :)
Posted on Reply
#12
1d10t
Kinda weird.
- all share same TDP despite different EU count, frequencies doesn't seem friendly.
- boost behavior SKU3 look really off, I mean 3 times base clock seem "magical".
- as someone already mentioned, PCIe 4.0 x 12 , what it like 8 GPU + 4 something-something PCH in between?
Posted on Reply
#14
Zareek
londisteWould suspect frequencies, not TDPs. Either does not clock too well or efficiency goes really bad. Mobile first probably has to do with synergy they are able to build, a laptop platform instead of a discrete GPU (plus better profit margins).

Besides, this is all rumors at this point anyway, they might be same far with dedicated cards but we just do not know about it yet :)
Yes but do you not agree that adding frequency and/or adding EUs will ramp TDPs. We know desktop cards are hitting more than 100Watt TDPs these days.

You can nitpick how I expressed it til' the cows come home. It's pretty obvious I'm wondering how well it scales up and if that is an issue. We are used to GPU vendors starting with or near halo products and working their way down the stack not the other way around. It seems like they should be beyond rumors by now, at this rate they will be obsolete a year before they launch.
Posted on Reply
#15
mechtech
I imagine as soon as these hit the street Nvidia and AMD will be dissecting them and be looking for patent infringement if any.
Posted on Reply
#16
londiste
mechtechI imagine as soon as these hit the street Nvidia and AMD will be dissecting them and be looking for patent infringement if any.
All of them have quite wide cross-licensing agreements. Just look up all the lawsuits these three have had against each other :)
Posted on Reply
#17
Xajel
MDDBThe CPU to GPU link will be PCIe4 x12 : that sounds brilliant, no current GPU benefits from having full 16 lanes, so why implement it? It may help making things more stable and less power hungry, i imagine. How come no other vendor has done it before? Or have they?
I guess this is because the CPU has only x16 PCIe 4.0 lanes.. and they need x4 lanes for the NVMe...

Just a guess, it could be more or less but I guess you get the idea.
Posted on Reply
#18
voltage
If these come with DDR5 then I'll buy, if not, forget it. Enough already with DDR4!
Posted on Reply
#20
ScaLibBDP
Exascale Processing Capability of Intel Xe-series GPUs is evaluated in a Video Technical Report on Youtube.
Take a look if interested:

Exascale Processing Capability of Intel Xe-series GPUs
Posted on Reply
Add your own comment
Copyright © 2004-2021 www.techpowerup.com. All rights reserved.
All trademarks used are properties of their respective owners.