Intel's DG2 (512 version) GPU die is approx. 396mm^2, larger than RTX 3070 die

[g67575]

The approximately half the size 190mm^2 die apparently has 8 billion transistors, so the 512 Execution Unit variant probably has about 16 billion transistors. Source here:
https://www.techpowerup.com/gpu-specs/intel-dg2.g918

The RTX 3070 die has 17.4 billion transistors, so might still end up being slightly ahead in performance (unless the DG2 has high csizeore + memory clocks) But, I think the DG2 512 might be a bit cheaper, and offer more GDDR6 VRAM (12 or 16GBs), which might could make it more appealing.

Since the die is slightly larger, you might expect more transistors to be crammed in than the GA104 die, but I suppose this mostly depends on if Intel is using TSMC's 6nm EUV fabrication process, or standard TSMC 7nm (still unconfirmed, so far 6nm has seen very little adoption in products).

As far as I know, Intel still plans to release DG2 graphics cards by the end of the year.

The die size in the title is based on estimations people have made by pixel counting, based on the photos of the die, like the one below:
https://hexus.net/tech/news/graphics/147899-intel-shows-photo-xe-hpg-dg2-gpu-512-eus/

 

[Grim5]

I hope you mean billion because if not DG2 is going to be very slow

 

[g67575]

Corrected, thanks.

 

[Purgatory]

.....clearly not even real numbers and just a placeholder... wasn't it clear by the details stated ?

 

[g67575]

Yeah, maybe it was just an estimate. Still, no indication yet Intel will be using 6nm EUV for the die.

 

[Purgatory]

Corrected, thanks.

It does state 8,000 million so would be 8 Billion, it's a placeholder. Also too much missing info on there and just will get filled out when known.


Edit: I need to go to specsavers too it seems. Techpowerup normally do this they fill out some fields and wait for the rest to come in, they like to use it as a clickbait placeholder.

 

[g67575]

EDIT - The DG1 appears to have been fabbed on the 10nm SuperFin process, according to these websites:
https://www.guru3d.com/news-story/i...-ryzen-platform-and-high-end-intel-mobos.html

https://www.tomshardware.com/uk/news/intel-dg1-gpu-teardown-and-first-failed-benchmarks

Seems possible that Intel would choose 10nm SuperFin / Superfin Enhanced for DG2.

The calculation above assumes a transistor density of 42.1 million per mm^2.

So, for the smaller DG2 42.1 million x 190mm^2 = 7.999 billion transistors.

For the fullsize (512 Execution Unit) DG2, 42.1 million x 396mm^2 = 16.671 billion transistors.

If the transistor density is correct, it would still mean Intel could cram in almost as many transistors as GA104 (RTX 3070 die) with 10nm.

With Samsung's 8nm process, Nvidia achieved a transistor density of 44.4M - 45.1M / mm² for their Ampere series, so it makes sense that transistor density for a 10nm GPU die would be a bit lower.

So, I think how powerful it ends up being, will largely depend on the fab. process Intel is using for DG2.

 

[g67575]

This seems to be the biggest clue, that the DG2 may actually fabbed on Intel's latest 10nm process:

https://www.tomshardware.com/uk/news/intel-xe-hp-hpc-graphics

in August 2020, Intel released details about a 512 Execution Unit data center GPU called the Xe HP, that was designed for Intel's 10nm process node.

 

[g67575]

More info here about the mobile DG2 version, which was scraped from Intel's website:
https://videocardz.com/newz/intel-c...ics2-dg2-will-offer-up-to-512-execution-units

The mobile variant has upto 512 EUs and comes in 4,6, and 16GB VRAM versions.

The card is rated to use upto 150w (TGP) for the high end model, which is higher than the TGPs of the RTX 3070 Mobile version (115-125W) and Radeon RX 6800M (145w), and about the same as the RTX 3080 Mobile (115-150W) sources here:

https://www.reddit.com/r/hardware/comments/m4ophe/nvidia_geforce_rtx_30_mobile_series_performance/
https://www.notebookcheck.net/AMD-R...aptop-GPU-at-a-145-W-TGP-target.541446.0.html

It looks like a less power efficient design, compared to Nvidia's and AMD's mobile GPUs, although the 8GB version has a lower TGP of 120w.

 

[g67575]

I wonder if Intel plans to use GlobalFoundries to fabricate 7nm GPU dies for XE HPG (DG2)?

I'd bet this would get them plenty of production capacity (arguably, the main concern for Intel, considering the massive demand for GPUs), assuming that Intel and GlobalFoundries managed to get the 7nm node working. I suppose if Intel buys GlobalFoundries, GlobalFoundries 7nm might be used in future Intel CPUs also (since GlobalFoundries' 7nm fabrication process is likely ahead of Intel's 7nm)

Details of a possible deal between Intel and GlobalFoundries, here:
https://www.techpowerup.com/284611/intel-in-talks-to-purchase-globalfoundries-for-usd-30-billion

$30 billion is a massive deal, whatever they have planned must be substantial. We know GlobalFoundries have produced 12nm GPU dies for AMD a few years ago, so I think it's plausible.

GlobalFoundries were working on 7nm in 2018, and were apparently using EUV lithography also:
https://spectrum.ieee.org/nanoclast...es/globalfoundries-halts-7nm-chip-development

We haven't heard any confirmation from Intel or TSMC that 7nm (TSMC) would be used for XE HPG GPUs, which is why I find this claim dubious, considering they have previously been competitors.