Intel ARC GPUs, Xe Architecture for dGPUs [2022-]

Intel engineers have been spotted working on Xe3 (codenamed Celestial) enablement in Linux’s grand archive of all mail that affects kernel development. The latest update to the drm-xe-next pull request, dated October 10, 2024, was flagged by Phoronix after it showed several references to Xe3. This probably means that driver development for this GPU architecture has begun for Linux, even though discrete Xe2 Battlemage GPUs are yet to hit the market.
 
amd needs to get these chips in more laptops. lunar lake launched later and has easily twice as many laptops available. also the majority of the ai 9 laptops are also paired with a dgpu, (nvidia of course because although amd allegdly makes mobile graphics they dont sell them they just like filling up a warehouse with them). but why pair your most powerful igpu with a discrete card if that apu is supply constrained?
 
Nice little video from Tom. I wonder what specific compatibility issues they ran into with Alchemist’s SIMD8. Game engines are probably tuned for Nvidia’s and AMDs SIMD32 but I don’t think graphics APIs support cross thread communication. Maybe SIMD8 isn’t great for optimal cache and memory bandwidth usage.

 
There’s a pretty rich wave intrinsic API in SM6.x these days for cross wave communication, including quad-level, available in graphics shader stages. Or did you more mean exposing something like group shared memory to graphics, for communication across waves in the same group?
 
There’s a pretty rich wave intrinsic API in SM6.x these days for cross wave communication, including quad-level, available in graphics shader stages. Or did you more mean exposing something like group shared memory to graphics, for communication across waves in the same group?

I meant communication across lanes in a wave. I know it’s been a big deal for compute workloads but good to hear it’s supported by graphics APIs as well.
 
Should I do my once in a 5 years pop up here and say that I am following these with great interest? :)

Though not that it means more than more competition is better and I have always been on underdog's side.

EDIT: oh dear, only 14 months since my last post... I am becoming addicted it seems.
 
B580 really needs to beat A770 soundly to even be in the same conversation with upcoming $300 cards from AMD and Nvidia. If they can pull that off it will be a massive jump in efficiency from Alchemist to Battlemage.
 
Maybe this is the forum where I can admit that I do have another reason for following these. Mostly atmosphere has here been nice and so on and so forth (And I am waiting Davros to show up in this thread to proof me otherwise. ;) )

In autumn 2022, Siru Innovations was bought to be part of Intel Graphics division. As far as I know and I have been able to search, they are still part of it and very much of in hardware graphics design.
Other oldies do remember my history and do know the how it connects to people who found Siru Innovations after leaving the Qualcomm.


So, maybe.... And hopefully this would end over quater-century long period of time for me. Not that I am in position of buying any of this in my life right now, but it would be nice to have a closure this history... Even if it would be based on asumptions based on rumors and time-line likelihoods, etc. :)

tbh, I don't really even care that much how good it is, as long as it is buyable. :D
 
If someone did not know, Launch date for something (most likely B570 and B580 based on BMG-G21?) will most likely be tomorrow. Release would be 12th of this month.

Found few rather believable leaked intel documents from last july describing even package details and pin counts for BMG-G31 and there has been R&D purpose shipping documents for G31 as well, so I would expect it to follow bit later. (Not a year later though. Not being taped out is not believable. If it is that late, it is more likely to be cancelled.)

It will be interesting to see, how Intel handles the launch if the G31 is "soon to be follow". (The classic how to announce new product without making statement from the future which eats the market share of this launch.)

EDIT:
In my books these are looking as follows:
  • BMG-10
    • Cancelled somewhere in late 2023 / early 2024 or was redeveloped to G31.
  • BMG-G21
    • B570 variant
      • 18 Xe2 cores (?)
      • 160 Bit memory bus
      • 10GB Memory
    • B580 variant
      • 20 Xe2 cores
      • 192Bit memory bus
      • 12GB memory
  • BMG-G31
    • 32 Xe2 cores
    • 256Bit memory bus
    • 16GB memory

EDIT2: Corrections and self adjust to nowadays style of launch before release... forgive me, it has been almost two decades since posts like this. :)
 
Last edited:
If someone did not know, Launch date for something (most likely B570 and B580 based on BMG-G21?) will most likely be tomorrow. Release would be 12th of this month.

Found few rather believable leaked intel documents from last july describing even package details and pin counts for BMG-G31 and there has been R&D purpose shipping documents for G31 as well, so I would expect it to follow bit later. (Not a year later though. Not being taped out is not believable. If it is that late, it is more likely to be cancelled.)

It will be interesting to see, how Intel handles the launch if the G31 is "soon to be follow". (The classic how to announce new product without making statement from the future which eats the market share of this launch.)

EDIT:
In my books these are looking as follows:
  • BMG-10
    • Cancelled somewhere in late 2023 / early 2024 or was redeveloped to G31.
  • BMG-G21
    • B570 variant
      • 18 Xe2 cores (?)
      • 160 Bit memory bus
      • 10GB Memory
    • B580 variant
      • 20 Xe2 cores
      • 192Bit memory bus
      • 12GB memory
  • BMG-G31
    • 32 Xe2 cores
    • 256Bit memory bus
    • 16GB memory

EDIT2: Corrections and self adjust to nowadays style of launch before release... forgive me, it has been almost two decades since posts like this. :)
it seems like you are right. The presentation of the new mid-range Intel GPUs is going to be tomorrow afaik. Did you work at Intel or in a similar architecture if that's not asking much?

woY1aho.jpeg
 
No. I am just hardware 3D geek from first decade of 21st century. I did have connections back then but they have more or less vanished. :)

There was a moment where I knew three different upcoming gfx chip specs. Unfortunately my luck was not up to my skill search and filter information as only one of them ever was released and even that came 10 months late with sub par clocks compared what the target was.

Back then the business was probably one tenth what it is now, so leaks were more scarse and making cross blind checks needed lot of connections, but while there is now more leaks, the amount of misinformation has raised more than the actual info itself.
 
Last edited:
Some additional details on the GPU:-

Intel BMG-G21
- TSMC N5
- 272 mm²
- 19,6 Billion Transistors
- 72,1 MTr/mm²

This is pretty poor density. AD107 as per TPU is 18.9B transistors with a 159mm2 die size. And this barely beats AD107. Intel are definitely not making much of a margin on this, esp with the additional 4 GB VRAM. No direct comparison to AMD since they don't have mainstream N5 parts but N33 is 204mm2 on N6 and is fairly close in performance.
 
Some additional details on the GPU:-

Intel BMG-G21
- TSMC N5
- 272 mm²
- 19,6 Billion Transistors
- 72,1 MTr/mm²

This is pretty poor density. AD107 as per TPU is 18.9B transistors with a 159mm2 die size. And this barely beats AD107. Intel are definitely not making much of a margin on this, esp with the additional 4 GB VRAM. No direct comparison to AMD since they don't have mainstream N5 parts but N33 is 204mm2 on N6 and is fairly close in performance.

Yeah it’s pretty bad. Intel has more control overhead than Nvidia and AMD. SIMD16 instead of SIMD32. But not sure that explains such a huge difference in density.
 
Back
Top