News Intel GPUs - Battlemage dGPUs officially launched

[gdansk]

I find it interesting that their improved drivers for Xe-Hp are indicated as having improvements to the Xe-LP products, and that the XeSS tech can use the DP4a units on the Xe-LP to also successfully upscale. I hope that someone takes the time to see if the 80/96eu parts in Tiger Lake benefit from these improvements enough to put some distance between them and Vega8 in the Cezanne. It'll also be interesting to see how Cezanne and TigerLake work at 900p -> 1080p upscaling an higher detail settings.

And Vega could do FSR too, for a more balanced comparison, right?

 

[uzzi38]

Linked image is a summary of supported GPUs courtesy of @stebler (I'm on my phone so I can't check if this is the actual account or if it was capital S, sorry if it's wrong).

Just going to add in that Series X can also support where the PS5 can't, unless Intel do another version that works with FP16.

 

[mikk]

https://twitter.com/x/status/1428406906528350222

Impressive track record from Moore's Law Is Dead lately. He even predicted the 45 TFLOPs FP64/FP32 from Ponte Vecchio months ago (he told the target is 46 TFLOPs).

 

[mikk]

die size calculation from tomshardware:

We've drawn lines to show how large the chips are, and based on our calculations; it looks like the larger Xe HPG die will be around 24x16mm (~384mm^2), give or take 5–10% in each dimension. We counted the dies on the wafer as well, and there appear to be 144 whole dies, which would also correlate to a die size of around 384mm^2.

Intel Arc Alchemist: Release Date, Specs, Everything We Know

Can Intel's Arc Alchemist compete with AMD and Nvidia GPUs?

www.tomshardware.com

 

[KompuKare]

Anyone have any idea how much TSMC charges for 6nm wafers?
Nearly 400mm² should give around 100 fully functioning dies per wafer. $10k would make each such die costs around $100.

 

[uzzi38]

Anyone have any idea how much TSMC charges for 6nm wafers?
Nearly 400mm² should give around 100 fully functioning dies per wafer. $10k would make each such die costs around $100.

Not exact numbers, but $/Tr is flat vs N7. Slightly costlier, but difference is more or less made up by smaller dies (and thus more dies per wafer).

 

[KompuKare]

Not exact numbers, but $/Tr is flat vs N7. Slightly costlier, but difference is more or less made up by smaller dies (and thus more dies per wafer).

I guess the simplicity of nearly exactly 100 good dies is that $10k makes them $100 each, $15k makes them $150 each, etc.
Might give us an idea on how low Intel could price if they want to be disruptive.

 

[mikk]

https://twitter.com/x/status/1428213564809125894

Doesn't sound like DG2 is approaching RTX 3060/Ti only, I'm definitely expecting at least RTX 3070 level of performance.

 

[jpiniero]

Doesn't sound like DG2 is approaching RTX 3060/Ti only, I'm definitely expecting at least RTX 3070 level of performance.

From the quote they are talking about PV. At least on paper it does sound like it could be very competitive with Hopper.

 

[Mopetar]

I don't think there's too much doubt that Intel could make a decent compute focused GPU, but there's always been some doubt regarding gaming since that requires a lot more software support and that's going to take time.

 

[DownTheSky]

I can't believe nobody posted this. Intel DG2-512 vs Navi 22 and GA104

https://twitter.com/x/status/1429096201350049794

 

[jpiniero]

I don't think there's too much doubt that Intel could make a decent compute focused GPU, but there's always been some doubt regarding gaming since that requires a lot more software support and that's going to take time.

Yeah. As above, this should still be a 700-800 dollar mining card at the very least.

 

[Tup3x]

I can't believe nobody posted this. Intel DG2-512 vs Navi 22 and GA104

https://twitter.com/x/status/1429096201350049794

View attachment 49133

Lots of ROPs and TMUs compared to Navi 22. Definitely 1440p+ cars.

 

[andermans]

Lots of ROPs and TMUs compared to Navi 22. Definitely 1440p+ cars.

The TMU per FP32 unit is actually the same as the Navi22 card, and the Ampere card has half that (consequence of doubling the number of FP32 Units per SM?).

 

[IntelUser2000]

For Ampere it's not much as it halved the TMU/ROP, but the other way around, by doubling FP32 without increasing others.

 

[Stuka87]

Lots of ROPs and TMUs compared to Navi 22. Definitely 1440p+ cars.

We have no way of actually knowing that as we do not know how these ROPs or TMU perform. We already know that this is a bit different between nVidia and AMD.

Overall, the numbers are cool to see, and give us a better idea as to what to expect. But at the same time, we also dont really know what to expect as we lack something to compare them to.

 

[LightningZ71]

And Vega could do FSR too, for a more balanced comparison, right?

Well, after some mild research, it appears that DP4A is a requirement of SM 6.4
Vega 8 in Renoir, Lucienne, and Cezanne all support SM 6.4 (they were supposedly derived from the existing 7nm design for Radeon7/VEGA 20)

Since Tiger Lake and Vega 8 will both be using the DP4A fallback mode, it's going to be interesting to see how their performance compares using the same mode.

 

[jpiniero]

Intel confirmed that PV is 1:1 DP. So 45 TF of FP64.

 

[itsmydamnation]

Intel confirmed that PV is 1:1 DP. So 45 TF of FP64.

then why is it only 45 of FP 32..... that seems silly of them

 

[uzzi38]

Intel confirmed that PV is 1:1 DP. So 45 TF of FP64.

Good. If it was 1:2 it would have been absolutely pathetic in the one area it was supposed to be good at - HPC.

 

[xpea]

The GPU shortage is expected to persist well into next year. Unfortunately, as Intel is using TSMC, don’t expect retail availability to improve much.

yeah these will be sold out immediately. A small chunk of TSMC N6 capacity is totally unable to meet the demand...

 

[RTX2080]

DG2 ALCHEMIST SKU2

43×37.5mm
12GB RAM
192BIT
TDP 80-120w

source:https://tieba.baidu.com/p/7512355628

 

[Hitman928]

DG2 ALCHEMIST SKU2

43×37.5mm
12GB RAM
192BIT
TDP 80-120w

source:https://tieba.baidu.com/p/7512355628

That die size has to be wrong, right?

 

[RTX2080]

That die size has to be wrong, right?

The leaker also being surprised by the size, seems close to 400mm2, that's what I could tell......

 

[Hitman928]

The leaker also being surprised by the size, seems close to 400mm2, that's what I could tell......

The rumored die size is bigger than the reticle limit, so the chances of it being accurate are pretty much 0%. Obviously makes the rest of the info hard to trust as well.