News Intel GPUs - Falcon Shores cancelled

[igor_kavinski]

i may be imagining this, but did gamers nexus imply that intel gpus benefit from a memory allocation process that requires intel boards/cpu to work? (cant remember the name)

Rebar. No, it also works with AMD Smart Memory Access. Don't worry. Intel isn't that stupid.

 

[MrTeal]

i may be imagining this, but did gamers nexus imply that intel gpus benefit from a memory allocation process that requires intel boards/cpu to work? (cant remember the name)

Rebar likely. Yes they do, but it's not Intel specific.

 

[igor_kavinski]

It was the same when i leaked malaysia shipment for G21(b580). 3-4weeks later there was another shipment to the AIB's. litteraly 1week later intel annouced B580. So i guess CES is battlemage ?

How is that huge? Are those full FOUR containers of G31 GPUs???

 

[LightningZ71]

Clever and useful idea for the unused PCIe lanes: https://videocardz.com/newz/maxsun-adds-two-m-2-ssd-slots-to-arc-b580-graphics-card

Hope more card makers jump on it, especially with PCIe 5.0 x8 being more than enough for current GPUs.

Hmm. I think that most boards won't have an issue driving lanes 9-12 for an m.2 slot as most boards will have a version with a bios that supports splitting the x16 into 2 x86 slots. Very few actually have support for x8/X4/X4.

Interesting concept though. I've always been fond of AMD's past forest into GPUs that support SSD storage with a pair of the FirePro cards. If they could spare four pairs of lanes, they could drive quad optane drives to buffer massive texture repositories locally. Heck, with PCIe 5, they could use single lanes.

 

[igor_kavinski]

Great idea. Pat killed Optane but maybe they can now revive it since Mr. Memory hater is gone.

 

[Win2012R2]

Many years ago Radeon Pro SSG had SSD on board that was addressable directly from the GPU, this gave some interesting optimisation opportunities as data could go directly into GPU memory bypassing CPU/PCIE switches/main RAM - this is what we need en masse in consumer GPUs to get proper support, PLUS dedicated ASIC for decompression of assets to avoid using GPU for it, pretty certain RDNA4 won't have it, BUT I am hopeful Blackwell will and that will force RDNA5/UDNA get it too, after that finally some parity with consoles!

Great idea. Pat killed Optane but maybe they can now revive it since Mr. Memory hater is gone.

Without Micron it's dead

 

[gdansk]

Pat killed Optane

It was dead in all ways except formally before Pat was CEO.

 

[igor_kavinski]

If we think about it just before going to sleep or first thing upon waking up in the morning, we can affect reality itself and bring Optane back!

 

[igor_kavinski]

It was dead in all ways except formally before Pat was CEO.

It was still seeing use by individuals with deep pockets and also server customers with needs that couldn't be met by anything other than Optane.

 

[Win2012R2]

It was dead in all ways except formally before Pat was CEO.

Intel always had bad attitude towards memories...

Optane was doing great - main problem was Intel's decision (which Pat did not reverse) to force Optane RAM DIMMs to work only with expensive Xeons - and DIMMs were ain't exactly cheap either, so that meant low volume. Endurance and low latency were perfect to use them on board of GPUs for direct addressing, or some other novel uses, it was very pathetic what Pat did.

 

[igor_kavinski]

it was very pathetic what Pat did.

Yeah and I'm surprised that people keep defending him by saying he was an engineer. What kind of engineer kills a sweet lovely baby like Optane??? Any engineering feelings he had, he lost them at EMC. Probably got fed up with anything storage related.

 

[Win2012R2]

I don't think Micron was particularly happy with having cheap large capacity Optane DIMMs to compete with their own ridiculously priced 128 GB DIMMs (which still are).

I thought CXL will be the saving grace that would work amazingly well for Optane - sticking them into DIMM was just too expensive/challenging, requiring CPU support, but with CXL it would have worked great, what a loss.

 

[gdansk]

It was still seeing use by individuals with deep pockets and also server customers with needs that couldn't be met by anything other than Optane.

You're looking it as an interesting niche technology. But as a product it was a dead end due to lack of demand at the price it could be made. And CXL killed it years before Pat.

 

[Win2012R2]

CXL killed it years before Pat

CXL was amazing opportunity for Optane - opening markets beyond Xeons it was limited to (in DIMM format), which is probably why Pat killed it. This would have worked wonders with GPUs being able to pull data with low latency, could have solved lots of issues with streaming assets etc.

 

[gdansk]

CXL was amazing opportunity for Optane - opening markets beyond Xeons it was limited to (in DIMM format), which is probably why Pat killed it.

Nope. It was dead end when CXL became an actuality. NAND flash does storage better/cheaper and DRAM does memory better. There wasn't enough demand and that's why both Micron and later Intel killed it.

Optane enthusiasts have this weird "well it was good" blindness. That never ever matters. Did it sell well? In a niche. At a scale where they couldn't decrease production costs. But I digress this is a GPU thread.

 

[adroc_thurston]

CXL was amazing opportunity for Optane - opening markets beyond Xeons it was limited to (in DIMM format), which is probably why Pat killed it.

a) more tiers is more tears
b) 3DXP would never be cost per bit competitive with any established memory

 

[Win2012R2]

Did it sell well?

Well I did my bit

I disagree about CXL: DRAM is still very very expensive, after so many years in server space we still have only 64 GB DIMMs that are reasonably priced, now 96 GB is popping in but 128 GB remains ridiculously priced. NAND on CXL was nowhere near good enough - from memory Optane was able to address data at 64 bytes granularity, and modern NAND is what, 16k pages now?

I doubt Micron wanted competition with its high capacity DRAM business, maybe that was the reason why Intel also thrown in the towel, still - they could have easily carried it for at least a few more years to get CXL working, considering that memory is the huge limit for AI such low latency addressable storage COULD have sold now as GPUs could have pulled data directly from it.

 

[511]

Intel always had bad attitude towards memories...

Not just Intel AMD as well maybe PTSD of 1980s when Grove changed buisness model due to japanese oversupply

 

[igor_kavinski]

Grove - A special breed of dog known for hunting memory chips and destroying them.

 

[TESKATLIPOKA]

From the looks of this the major improvements of celestial will be PPA LNL iGPU was ~39mm2 N3B and 12 core Xe3(+50% EU) is 54.68 on N3E
View attachment 112725

I am looking and looking, but I don't see that major PPA improvement.
8 XE2 iGPU tile in LNL is ~39mm2
12XE3 iGPU tile in PTL is ~54.7mm2
If you linearly scaled XE2 to 12XE2 It would be ~58.5mm2, but in reality It should be less.
So basically comparable tile sizes and any PPA improvement will entirely depend on achievable higher clocks or architectural improvements, which is still unknown.

As for my opinion about Battlemage.
The price looks good for that performance, I also like 12GB Vram, RT, OC.
But It's seriously too big for that performance and TDP is also not very good.
Yeah, I as a gamer care only about price or Perf/$, but If they have no profit or even a loss on these GPUs, then It's highly likely Intel will pull out.

 

[511]

I am looking and looking, but I don't see that major PPA improvement.
8 XE2 iGPU tile in LNL is ~39mm2
12XE3 iGPU tile in PTL is ~54.7mm2
If you linearly scaled XE2 to 12XE2 It would be ~58.5mm2, but in reality It should be less.
So basically comparable tile sizes and any PPA improvement will entirely depend on achievable higher clocks or architectural improvements, which is still unknown.

N3B is dense than N3E by approx 5%(N3P gains the density back from N3B except for Sram) so that would mean around 61mm2 and it is 54.7 mm2 around 11% shrink on a samish node while improving PPA is good and as you said architectural improvements are unknown

As for my opinion about Battlemage.
The price looks good for that performance, I also like 12GB Vram, RT, OC.
But It's seriously too big for that performance and TDP is also not very good.
Yeah, I as a gamer care only about price or Perf/$, but If they have no profit or even a loss on these GPUs, then It's highly likely Intel will pull out.

Yeah their margin is the only issue cause as a product it is Good

 

[TESKATLIPOKA]

N3B is dense than N3E by approx 5%(N3P gains the density back from N3B except for Sram) so that would mean around 61mm2 and it is 54.7 mm2 around 11% shrink on a samish node while improving PPA is good and as you said architectural improvements are unknown

My mistake, I thought N3E has higher density than N3B, but let's not forget that 50% more XE doesn't meant 50% bigger tile size.
We will be able to compare It when PTL is out, It will have both 8XE and 12XE variants.

 

[511]

My mistake, I thought N3E has higher density than N3B, but let's not forget that 50% more XE doesn't meant 50% bigger tile size.
We will be able to compare It when PTL is out, It will have both 8XE and 12XE variants.

For this one yes cause outside of the Xe cores everything is on Main tile with IMC so the only thing that remains is Xe Cores you cannot find a better comparison.
There is no 8Xe3 core only 12Xe3(N3E) and 4Xe3(Intel 3) exist funnily we can compare Intel 3 vs N3E

 

[TESKATLIPOKA]

For this one yes cause outside of the Xe cores everything is on Main tile with IMC so the only thing that remains is Xe Cores you cannot find a better comparison.
There is no 8Xe3 core only 12Xe3(N3E) and 4Xe3(Intel 3) exist funnily we can compare Intel 3 vs N3E

Yeah, It looks like almost everything is part of Render Slices.

 

[tajoh111]

N3B is dense than N3E by approx 5%(N3P gains the density back from N3B except for Sram) so that would mean around 61mm2 and it is 54.7 mm2 around 11% shrink on a samish node while improving PPA is good and as you said architectural improvements are unknown

Yeah their margin is the only issue cause as a product it is Good

When you in Intel's financial position, margin is the most important factor.

Intel is bleeding money like crazy and for arc to be viable in the long run, they need to make money.

Without good finances, this product cannot generate the financials to keep the driver team going and support going.

People thinking Intel is making money on this thing are delusional.

Polaris was a lower margin product and was priced at 239 dollars. But it could make money because the costs were much lower at the time.

Battlemage die is larger than polaris and using 5nm technology. Polaris was manufactured with 14nm samsung technology which was probably 4000 dollars per wafer. TSMC was 5000 dollars per wafer for 16nm and samsung was worse and cheaper. 5nm TSMC is 16,000 dollars per wafer(I don't see any reports of 5nm decreasing in price really and mostly the opposite).

Add in inflation, cost to ship things have gotten more expensive and no way is Intel making a profit.

The reason Intel has to price this thing so low is because next gen parts are not too far off which will drive this into even lower prices. Intel needs to sell as many of these things as possible as the RTX 5060 and RX 8600 should be 20 to 30 percent faster than this chip for around 300 dollars.

Intel was left holding the bag with R and D and TSMC wafer obligations that they can't produce nothing. Gamers buying this should not complain when ARC pulls out and they are left with videocard paper weights in two years.