News Intel GPUs - Intel launches A580

Page 69 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Jul 27, 2020
16,338
10,348
106
Maybe they are just being cautious, after having superior products like Optane failing in the market. Their GPUs aren't the absolute best otherwise they would be in reviewer's hands by now. They will test the waters with this initial run and improve their GPUs generation by generation, like they did going from Comet Lake to Alder Lake. Give them time.
 

DrMrLordX

Lifer
Apr 27, 2000
21,637
10,855
136
Falcon Shore sounds like a HPC APU, built with tiles.

The product AMD never saw fit to release. Perhaps we're about to find out why not.

Doesn't Intel have their own fabs, though? They should be able to pump out much more product than this if they really wanted to.

On which process? Intel already had plans to produce Ponte Vecchio with 10ESF/Intel 7 and 7nm/Intel 4, but they scrapped that entirely in favor of TSMC nodes (N6 and N5 if I recall correctly). Intel is already putting Alder Lake, Raptor Lake, Sapphire Rapids, and Emerald Rapids on 10ESF with unknown yields (and an unknown amount of 10nm+ fab capacity is still finishing out production of IceLake-SP, assuming that's still in production). All indicators hint that Intel either didn't like the results they were getting for Ponte Vecchio tiles using their own nodes or that they just weren't going to have enough capacity to cover both dGPUs and CPUs. Possibly both.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
In fine print, Intel says $700 million of the $1 plus-billion is essentially royalties. This means majority of the 4 million units are coming from the smaller 128 EU dies.

I assume they are baking in the possibility that the high end Alchemist might not be that competitive. And I mean by perf/$.

@NTMBK Falcon Shores sounds like the productized version of MLID talking about a Sapphire Rapids offshoot that puts in an Xe tile in place of one compute tile.

When it comes to these CPUs, you cannot say one CPU is one process anymore.

On which process? Intel already had plans to produce Ponte Vecchio with 10ESF/Intel 7 and 7nm/Intel 4, but they scrapped that entirely in favor of TSMC nodes (N6 and N5 if I recall correctly). Intel is already putting Alder Lake, Raptor Lake, Sapphire Rapids, and Emerald Rapids on 10ESF with unknown yields (and an unknown amount of 10nm+ fab capacity is still finishing out production of IceLake-SP, assuming that's still in production).

Ponte Vecchio is using Intel 7, Intel 10nm SF, TSMC N5 and N7. Supposedly they might move the compute tile to Intel 4 later on.

Alderlake reaches 5.5GHz in the KS variant. The yields are pretty good on that node now.
 
Last edited:

mikk

Diamond Member
May 15, 2012
4,141
2,154
136
I don't know if it has a meaning, in the Intel slide Alchemist is placed in the middle of 2022 whereas Battlemage in the middle of 2023/2024. Just saying because some people like Charlie claimed Battlemage is coming early 2023.
 
  • Like
Reactions: IntelUser2000

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
So DG2 running at 2.4GHz is spotted in Geekbench. Gets 87K score. The score is around RTX 2060. I guess the clocks are more of a news as the previous DG2 scores were barely outperforming the Iris Xe G7.

Wait what? They're using 10ESF and 10SF? Why 10SF?

SF is for the base tile.
 
Last edited:

Tup3x

Senior member
Dec 31, 2016
965
951
136
Sub test scores were all over the place so it impossible to make any conclusions.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
It doesn't matter much. Majority of Intel shipments this year is going to be the 128EU die. 4 million units worth $300 million revenue after you take into account the accounting shenanigans.

To me this is a possible admission that high end won't sell lots because it won't be good as people are expecting. The expectations are quite high.
 

Frenetic Pony

Senior member
May 1, 2012
218
179
116
It doesn't matter much. Majority of Intel shipments this year is going to be the 128EU die. 4 million units worth $300 million revenue after you take into account the accounting shenanigans.

To me this is a possible admission that high end won't sell lots because it won't be good as people are expecting. The expectations are quite high.

I think it's more "we don't have dies". Initially I think they could sell a million or more upper end dies if they had a decent price, with supply and demand as it is. But if they don't have dies, we'll it doesn't matter how much demand there is.
 

Dayman1225

Golden Member
Aug 14, 2017
1,152
974
146
(and an unknown amount of 10nm+ fab capacity is still finishing out production of IceLake-SP, assuming that's still in production)

Not sure why you says this. Icelake SP will be in production for years and Icelake SP will be higher volume than SPR until atleast the end of 2023.
 

DrMrLordX

Lifer
Apr 27, 2000
21,637
10,855
136
Not sure why you says this. Icelake SP will be in production for years and Icelake SP will be higher volume than SPR until atleast the end of 2023.

If Intel can ramp production of Sapphire Rapids then there might not be much need for IceLake-SP outside of the usual long-term contracts. Assuming Intel chooses to saturate their entire Xeon Scalable lineup exclusively with Sapphire Rapids.

But that's a big "if'.
 

Dayman1225

Golden Member
Aug 14, 2017
1,152
974
146
If Intel can ramp production of Sapphire Rapids then there might not be much need for IceLake-SP outside of the usual long-term contracts. Assuming Intel chooses to saturate their entire Xeon Scalable lineup exclusively with Sapphire Rapids.

But that's a big "if'.
Intel should have no issue ramping SPR but it’s volumes won’t surpass ICL volume due to DDR5 volumes. DDR4 is cheaper and far more readily available. It’s the same story for Milan and Genoa. But this is the wrong thread for that discussion
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I think it's more "we don't have dies". Initially I think they could sell a million or more upper end dies if they had a decent price, with supply and demand as it is. But if they don't have dies, we'll it doesn't matter how much demand there is.

The bottleneck is said to be elsewhere, not TSMC. Like with my GPU repair and electronics as a hobby when I check supply, lot more than usual are out of stock. They are used to build the boards for the GPUs. Also the industry was talking about substrate shortages little while ago.

3070 Ti is only 20% faster than 3060 Ti so while they might have targetted the former, silicon quality can easily swing it that much. This is a matter of whether they can reliably reach certain frequencies.
 
Last edited:

Saylick

Diamond Member
Sep 10, 2012
3,171
6,404
136
63 total tiles and 16, a little over a quarter of the total, are thermal tiles? It looks like there's a thermal tile over every EMIB connection. Rather interesting... It looks a lot less impressive now that I know that a good portion of that middle "compute" area is unpowered silicon.

1645506748764.png
 
  • Like
Reactions: moinmoin

DrMrLordX

Lifer
Apr 27, 2000
21,637
10,855
136
@Dayman1225

So it's all 10ESF, N7 and N5? That's interesting. No 7nm/Intel 4 to be seen, as expected.

(also I don't know where or who posted that 10SF would be involved but apparently it isn't. I have no idea why Intel would have used 10SF).
 

jpiniero

Lifer
Oct 1, 2010
14,618
5,227
136
It doesn't matter much. Majority of Intel shipments this year is going to be the 128EU die. 4 million units worth $300 million revenue after you take into account the accounting shenanigans.

$75 is rather optimistic for the 128 EU chip+GDDR6 if they are excluding mining from their price calculations.