News Intel GPUs - we've given up on B770, where's Celestial already

[Aapje]

 

[blckgrffn]

TL;DW man

A verdict?

 

[Aapje]

Pretty good results, but still worse than Nvidia and AMD, so the price has to be lower with equal performance.

 

[DAPUNISHER]

TL;DW man

A verdict?

87% passed his criteria. He also approached it as an average shopper. Which means his troubleshooting was generic stuff you would expect them to do. If he approached it as a buyer that did their homework on ARC and used DXVK, it would have been more like a 95-97% pass rate. Alan Walks 2, Starfield, and a couple of others I don't think there is much helping unless the driver team works more magic. The performance in those is below anything in the price class. I was happy Starfield even got working fully, after the way it started out. I was half convinced it was nerfed on the hardware.

 

[blckgrffn]

87% passed his criteria. He also approached it as an average shopper. Which means his troubleshooting was generic stuff you would expect them to do. If he approached it as a buyer that did their homework on ARC and used DXVK, it would have been more like a 95-97% pass rate. Alan Walks 2, Starfield, and a couple of others I don't think there is much helping unless the driver team works more magic. The performance in those is below anything in the price class. I was happy Starfield even got working fully, after the way it started out. I was half convinced it was nerfed on the hardware.

Nice, that makes sense.

I am optimistic about the sophomore release. Even if the PPA isn't great and all that, if they can make it just *work* better because they aren't fighting the hardware so hard that would be great.

I do fully expect them to put Gen 1 into "best effort" support as fast as possible though and shift everything they can to the new shiny to ensure it gets as much run its first 6-12 months possible. That wouldn't make me mad, it would just make sense.

I was thinking about getting an Intel card for my Plex/biking sim PC but I think I'll try to hold out for this next gen instead. I think it will age better and just be that much better at being power efficient.

 

[DavidC1]

For comparison AD107 is 24SMs/3072 Shaders on a ~160mm² die on 4N

20 Xe² is 160 EUs/2560 Shaders at 270mm² on N4P

Terrible PPA if so, yes.

It's not bad because Lunarlake is basically 50% faster at the same clocks and shader count. Obviously making one shader perform 50% faster has a cost. Clocks are supposed to be at least 25% higher. 30% gets to 4070 Super performance.

If they fix ReBar, and the compatibility and performance at 1080p and lower settings improve so there isn't that big gap from 4K, it'll go a long way on making it more attractive.

 

[DavidC1]

How can it be 400mm2 for 32xe2 cores. What about the shiping manifests 56/64xe2 versions that would put them above reticle limit.
8 cores of Xe2_lpg die area is ~40.7mm2 Media is ~13.5mm2 display ~7mm2 Assuming 1 display engine and 1 media engine 20Xe2 core should be: 122,25mm2 on N3E ? and around 153,625mm2 on N4 [N5->N3E is 1.3x chip density N5-N4 is 1.06x]
Im missing L2$ controls, 2 gddr phys or 1 gddr phys (192/256bit) and l2$
Still 270mm2 for 20xe2 cores and 400mm2 for 32xe2 seems ridiculous too me.

After some calculations i would put this at
32Xe2: 205-275mm2
64Xe2: 410-550mm2

If 32xe2 is 400mm2 and 4070non-super competitor bmg is DOA... 4070ti is 294.5mm2

N3E is 1.6x logic density and 1.3x "chip density" but that's an arbitrary number with SRAM cells. That would make your calculations at 195mm2.

They probably had 56 Xe core version that was quite different. But Lunarlake shows that it performs unexpectedly better - based on high level specs, the 32 Xe core Battlemage performing like 4070S would look "stellar" per shader(not really because it took lots of extra transistors to do so, but still).

This is like RDNA3 but opposite, where it seems to have crappy Flops but performs much higher. RDNA3 didn't perform like it's high Flop rating because the dual issue didn't take much transistors, while Intel spent hidden resources on making the seemingly same shader count perform far better.

TSMC's 3nm update: N3P in production, N3X on track

The last and final FinFET node from TSMC lives on.

www.anandtech.com

 

[Kepler_L2]

[v2] drm/xe/uapi: Expose SIMD16 EU mask in topology query - Patchwork

patchwork.freedesktop.org

"PVC, Xe2 and later platforms have 16-wide EUs." So if Xe3/Celestial is still SIMD16 I wonder what Tom Petersen meant with following NVIDIA uarch.

 

[gaav87]

[v2] drm/xe/uapi: Expose SIMD16 EU mask in topology query - Patchwork

patchwork.freedesktop.org

"PVC, Xe2 and later platforms have 16-wide EUs." So if Xe3/Celestial is still SIMD16 I wonder what Tom Petersen meant with following NVIDIA uarch.

Its already known since computex that xe2 has native simd16.
And i speculated in march it will have simd16 and i was 100% spot on the spec's.
Each EU in ACM had 8 SIMD lanes so 8 FP32 operations per cycle each lane can handle one FP32 operation.
1 eu 8 fp32 (add) / 16 fp32 (mad)
Now they have native16 that fully supports fp32 i think that what he meant by fully supporting NVIDIA uarch
fp32 (add) 16 and fp32 (mad) 32

 

[gaav87]

BMG top sku performance speculation from everything i know:
32xe2 cores 256bit 16GB 20gbps 640GB/s >>400mm2 ~21-23TFLOPs FP32
4070s>32xe2>4070 - gaming performance

24xe2 cores 192bit 12GB 20/19.5gbps ~540GB/s ~270mm2 ~16.5Tflops FP32
4070>24xe2>3070ti around 7700xt - gaming performance


Comfirmation of max 8 slices for xe2 =32xe2 cores is max
This is a dmesg0.txt from intel bmg-1 with 8 cores/2slices. (vcs2 and vcs0 active)

 

[DavidC1]

Its already known since computex that xe2 has native simd16.
And i speculated in march it will have simd16 and i was 100% spot on the spec's.

Kepler is talking about Xe3, not Xe2.

Also it has been known as SIMD16 way before March.

 

[511]

BMG top sku performance speculation from everything i know:
32xe2 cores 256bit 16GB 20gbps 640GB/s >>400mm2 ~21-23TFLOPs FP32
4070s>32xe2>4070 - gaming performance

24xe2 cores 192bit 12GB 20/19.5gbps ~540GB/s ~270mm2 ~16.5Tflops FP32
4070>24xe2>3070ti around 7700xt - gaming performance


Comfirmation of max 8 slices for xe2 =32xe2 cores is max
This is a dmesg0.txt from intel bmg-1 with 8 cores/2slices. (vcs2 and vcs0 active)

View attachment 102870

4070S with 4060ti price is fine I guess if they deliver on it

 

[gaav87]

Kepler is talking about Xe3, not Xe2.

Also it has been known as SIMD16 way before March.

And i explained my opinion why Tom mentioned nvidia. Also do you have any source of simd16 way before march im interested ?

4070S with 4060ti price is fine I guess if they deliver on it

Judging by tsmc prices and bionic twitts 349-399$ 4070s perf is doable

 

[DavidC1]

And i explained my opinion why Tom mentioned nvidia. Also do you have any source of simd16 way before march im interested ?

Flops also doesn't correspond to performance as Battlemage will run 50% faster than Alchemist at the same clock.

Whatever "deficiency" Alchemist seems to have compared to Nvidia in terms of Flops will be the opposite in Battlemage generation, where it looks like it's absolutely number 1 in that regard.

25TFlops Battlemage should be competitive with 35TFlops 4070S. Of course, again relying only Flops is misleading as it'll take 400mm2 die on N5 to do so.

 

[511]

Flops also doesn't correspond to performance as Battlemage will run 50% faster than Alchemist at the same clock.

Whatever "deficiency" Alchemist seems to have compared to Nvidia in terms of Flops will be the opposite in Battlemage generation, where it looks like it's absolutely number 1 in that regard.

25TFlops Battlemage should be competitive with 35TFlops 4070S. Of course, again relying only Flops is misleading as it'll take 400mm2 die on N5 to do so.

Outside of tech nerds and Intel themselves no one is going to measure I got 400mm2 N4 die size for so cheap it is so cool 🤣

 

[KompuKare]

Outside of tech nerds and Intel themselves no one is going to measure I got 400mm2 N4 die size for so cheap it is so cool 🤣

It mainly matters because of any expectation that Intel may be disruptive with pricing.

And the lingering fear that Intel could do an Intel: kill anything which isn't x86 if they don't get a return quickly enough no matter how much they've spent up until that point.

Despite the huge corporate welfare largesse governments are handing Intel for fabs, the rest of the company doesn't look that healthy - more so if they have to pay for a major Raptor Lake recall.

 

[511]

It mainly matters because of any expectation that Intel may be disruptive with pricing.

And the lingering fear that Intel could do an Intel: kill anything which isn't x86 if they don't get a return quickly enough no matter how much they've spent up until that point.

Despite the huge corporate welfare largesse governments are handing Intel for fabs, the rest of the company doesn't look that healthy - more so if they have to pay for a major Raptor Lake recall.

Btw N4 is 4-5 year old node at this point it cost like 13-14k for N5/N4 wafer and around for 9-10K for N7/N6 wafer

 

[DavidC1]

Btw N4 is 4-5 year old node at this point it cost like 13-14k for N5/N4 wafer and around for 9-10K for N7/N6 wafer

Assuming ~160x 400mm2 dies, it's $63 per die which is very expensive for a $250 product which includes 8 VRAM chips, and a large and complex circuit board with all the connectors, housing and at least two fans. Remember, vendors like Newegg needs to make money too, so Intel aren't selling them for $250. Probably $150-180 at the most.

Even half that is expensive. What's needed is going UP the stack. If they could up it to $399 in Battlemage generation, 400mm2 starts being acceptable.

Despite the huge corporate welfare largesse governments are handing Intel for fabs, the rest of the company doesn't look that healthy - more so if they have to pay for a major Raptor Lake recall.

Gelsinger isn't a beancounter like the previous 3 CEOs and one thing he said he wants to achieve is GPUs. It looks like it'll be longer than expected for success, with real changes coming with Celestial.

 

[KompuKare]

Assuming ~160x 400mm2 dies, it's $63 per die which is very expensive for a $250 product which includes 8 VRAM chips, and a large and complex circuit board with all the connectors, housing and at least two fans. Remember, vendors like Newegg needs to make money too, so Intel aren't selling them for $250. Probably $150-180 at the most.

Even half that is expensive. What's needed is going UP the stack. If they could up it to $399 in Battlemage generation, 400mm2 starts being acceptable.

Gelsinger isn't a beancounter like the previous 3 CEOs and one thing he said he wants to achieve is GPUs. It looks like it'll be longer than expected for success, with real changes coming with Celestial.

Taking https://isine.com/resources/die-yield-calculator/ and a square 20x20mm die for exactly 400mm², 300mm wafers, and a defect rate of 0.07 I get 100 good dies per wafer:

which at least makes the cost calculation easy: $14k / 100 = $140 per good die.
Defects and wafer prices are the big variables but that with a square die like there at most 131 candidates should a given.

400mm² is a bit bigger than GA103 which Nvidia can use for RTX 4080. If Intel can reach 4070 speed with that, it seems better than last gen, I think?

 

[DavidC1]

which at least makes the cost calculation easy: $14k / 100 = $140 per good die.
Defects and wafer prices are the big variables but that with a square die like there at most 131 candidates should a given.

So it's a complete disaster and loss maker at $10K and N6 process, because $100 is raw cost of the die. It probably costs them at least $150 to make if not $200.

$140 is doable if they can sell it for $399. It won't be a lot of money because they'll realistically get maybe $330 at most, but not like Alchemist. Early value estimates before A770 launched had the selling price closer to $599.

At this point, it needs to reach 4070S levels. Difference in $50 will be a matter of whether they are making a bit or being breakeven.

 

[Aapje]

Taking https://isine.com/resources/die-yield-calculator/ and a square 20x20mm die for exactly 400mm², 300mm wafers, and a defect rate of 0.07 I get 100 good dies per wafer.

which at least makes the cost calculation easy: $14k / 100 = $140 per good die.
Defects and wafer prices are the big variables but that with a square die like there at most 131 candidates should a given.

Most of the dies with defects are salvageable, in no small part because the chips are designed for it. So it seems more realistic to assume that they get something like 129/130 sellable dies, which puts the price at about $108 per die.

As long as selling those bad dies loses them less than $108, it's already worth it to sell those (OK, aside from cannibalisation, so in reality the loss needs to be smaller). Of course, if they do lose money on these, they need to make up for it with more profits on the good die cards, to make a profit overall.

 

[511]

So it's a complete disaster and loss maker at $10K and N6 process, because $100 is raw cost of the die. It probably costs them at least $150 to make if not $200.

$140 is doable if they can sell it for $399. It won't be a lot of money because they'll realistically get maybe $330 at most, but not like Alchemist. Early value estimates before A770 launched had the selling price closer to $599.

At this point, it needs to reach 4070S levels. Difference in $50 will be a matter of whether they are making a bit or being breakeven.

If it was fabbed at intel they could easily swallow the cost as N4 includes 50-60% tsmc margins as well.This is why gelsiner is betting on fabs all he needs to do is keep fabs full be it 70% intel and 30% external they are planning to bring Celestial on 18A they bought TSMC capacity to not loose market share on a side note i don't think after intel Trinity they had a good leadership maybe craig Barrett he invented Copy Exactly which everyone copied looking at how Lisa Su Changed AMDs Fate can Gelsinger save Intels fate we will see when CLF and GNR launches

 

[DavidC1]

bring Celestial on 18A they bought TSMC capacity to not loose market share on a side note i don't think after intel Trinity they had a good leadership maybe craig Barrett he invented Copy Exactly which everyone copied looking at how Lisa Su Changed AMDs Fate can Gelsinger save Intels fate we will see when CLF and GNR launches

Celestial is on N3. It will be at least Druid until it gets to Intel process. The problem is, with a 2 year cycle that is 2028. It won't be until 18P that Intel claims will have a density lead, so their process isn't the best fit for GPUs.

I do think he's best chance, but a lot is riding on his shoulders right now. Won't be easy.

Craig Barrett, while he is responsible for Intel's process leadership through Copy Exactly, it was also during his term when they brought the clockspeed focus that led to their downfall. He reminds me of a more sane Brian Kraznich. Fab guys don't seem to be the best CEO, because CEO requires looking at the bigger picture and having more far outlook, while running a factory is about pure discipline and execution which isn't so fitting.

 

[511]

Celestial is on N3. It will be at least Druid until it gets to Intel process. The problem is, with a 2 year cycle that is 2028. It won't be until 18P that Intel claims will have a density lead, so their process isn't the best fit for GPUs.

I do think he's best chance, but a lot is riding on his shoulders right now. Won't be easy.

Craig Barrett, while he is responsible for Intel's process leadership through Copy Exactly, it was also during his term when they brought the clockspeed focus that led to their downfall. He reminds me of a more sane Brian Kraznich. Fab guys don't seem to be the best CEO, because CEO requires looking at the bigger picture and having more far outlook, while running a factory is about pure discipline and execution which isn't so fitting.

You sure about N3E/P for celestial ? Looks like druid it is than Brian Kranzich and Paul otellni has to be the worst Intel CEOs It was paul who declined apple for their Fabs look at them now also the stagnation i guess Bob swan fired due to TSMC Capacity he bought but he made the right decision at the time it will take like H2 25 to see wether they succeed or not in FAB space

 

[Ghostsonplanets]

You sure about N3E/P for celestial ?

The PTL Celestial iGP 12 Xe³ iGP tile is fabbed on N3E. So it make sense that the dGP silicon is also fabbed there.

Granted, Celestial is also fabbed on Intel 3. So we might have a surprise on low-end dGP silicon given the IP is there already.