Reference to Inference-optimized GPU.
Also to next gen Shores(Jaguar Shores) product line.
Jaguar Shores(if it's not cancelled) is planned to be using Xe4 IP along with UALink and Ultra Ethernet
Last edited:
Reference to Inference-optimized GPU.
Also to next gen Shores(Jaguar Shores) product line.
Jaguar Shores(if it's not cancelled) is planned to be using Xe4 IP along with UALink and Ultra Ethernet
Xe3 on Pantherlake seems to indicate 25-30% perf/watt improvements over Xe2. That's actually ok. It won't be multi-generation behind anymore. Maybe just 1 year. If we take TPU efficiency results and extrapolate from B580, it would be RDNA4 class.
Pantherlake improves greatly on the Depth Writes micro bench test, which has to do with hidden surface culling. It'll utilize bandwidth and memory subsystem better.
Mesh Rendering test improvement is for high polygon count workloads.
Also variable register allocation and 25% increase thread counts per core improves utilization, which is a known problem even for Battlemage.
Something interesting:
Pantherlake improves greatly on the Depth Writes micro bench test, which has to do with hidden surface culling. It'll utilize bandwidth and memory subsystem better.
Mesh Rendering test improvement is for high polygon count workloads.
Also variable register allocation and 25% increase thread counts per core improves utilization, which is a known problem even for Battlemage.
Something interesting:
Last edited:
It was known that Celestial isn't Xe3, it is on Xe3p which is a significant architectural advancement according to Tom Petersen.
I would say 25-30% better perf/watt ISO-process is a significant architectural advancement. It would be nice if in a year they get that again.
@DrMrLordX It looks like Xe3 is just for Pantherlake and Xe3P is for Celestial dGPU. I guess they could use derivatives in future value products.
For now these are known but they may be used in some more future products as well
(Vanilla) Xe3 :
whole PTL line(both 12Xe3 IGPU tile and 4Xe3 IGPU tile),
Wildcat Lake IGPU(2Xe3 units)
4Xe3 tile used in PTL is to be reused in: NVL Desktop and -HX SKUs, SKUs of NVL(4P+0E+4LPE) CPU config and may be also in NVL-H (4P+8E+4LPE) CPU config in a similar fashion to PTL
Xe3P :
12Xe3P iGPU tile in NVL-H(4P+8E+4LPE) CPU config
NVL-AX(Halo): 24Xe3P
Celestial dGPUs
They are just going to re-use the 4Xe3 IGP tile(Intel 3 node) on some places that probably doesn't warrant a new die for may be cost/economics and other reasons(doesn't need that much iGPU perf or paried with a dGPU).
For premium segment or where they think it will make a difference they will go Xe3P, eg. top GPU config(12Xe3P) of NVL-H.
Xe3 in PTL has(up to) 6units in a render slice.
Irrespective of iGPU tile in NVL, whole NVL line will get the media engine upgrade over PTL(since it would be in a different tile).
Edit: removed some Xe3P reference
Last edited:
Wildcat Lake doesn't have Ray Tracing support. So obviously there can be more than just the base die configs. It's their own chip, they could do whatever they want with it.
They are trying to take advantage of the theoretical advantage of the Tile configs, which is being able to pair what you want with it. Before, they would have had to recycle the whole generation, including the CPU, now they can recycle what they choose. Have a latest core, but older GPU. Yea, for most desktops it's more than enough. Especially in desktops where the TDP constraint is much lower. 4 Xe3 is quite decent. You are getting Lunarlake graphics.
I do wish they would try something other than the accountants want, which is a low end CPU paired with the best GPU. Why not 2+4 core with 12 Xe3P? Oh I know, because forcing people to pay works, otherwise it's just too logical. Dream would be the NVL-AX coming with 24 Xe3 but Core 3 series or even Pentium with the 24 Xe3 version being $100 higher price than regular iGPU versions.
Same with AMD. Why not Strix Halo with full fat GPU but with Ryzen 3 CPU instead and save $200?
Last edited:
DrMrLordX
Lifer
- Apr 27, 2000
- 22,947
- 13,032
- 136
In the case of Strix Halo, they're charging $1500+ for the entire kit on the 395+. They'd probably charge $1300 ($200 less as you suggested) if they followed your advice and had a weaker CPU portion of the SoC on a cheaper model. At that point it wouldn't seem worth it to sacrifice the cores + clocks just to save $200.
In the case of a lower-end Nova Lake, it might make more sense.
It sounds to me like Xe3p will be a much bigger upgrade over Xe3 than Xe3 over Xe2.
Isn't it 16 and 32 Xe3p on NVL-AX?
Top GPU config for NVL-AX is 24 Xe3P units AFAIK.
Edit: not 100% sure, will have to check.
Last edited:
The fact that the systems are that pricey suggests in reality those CPUs, or at least systems using have to be quite expensive.
A socketed version of Strix Halo at $300 with Ryzen 3 would be ideal. You aren't talking RTX 5090 performance so that kind of CPU is enough plus you save money. But again, being consumer-friendly is too much logic for accountants.
This is the same reason why chipset iGPUs were actually free while on-die GPUs always cost more than they claim it is. Cause they always bundle best GPUs with best CPUs, meaning you pay the CPU tax always.
I'm not really buying that until they show it. Xe3 is already substantial at 25-30% ISO. "Bigger upgrade" means the gains will have to be greater than that. Fantastic, but skeptical. Xe3 already has really nice changes like variable register allocation.
Last edited:
DrMrLordX
Lifer
- Apr 27, 2000
- 22,947
- 13,032
- 136
Ideal for whom? AMD and their partners get to sell 64 or 128GB full fat configs with every 8060s they sell right now. You don't really think they're going to risk that revenue just so people can have fewer CPU cores?
From the AMD who refuses to give us dual CCD V-cache CPUs yet wastes that precious V-cache die on 6-core CPUs?
Yeah, keep dreaming
Ideal for who? You are really asking this question? Mind you they got to this perilous position BECAUSE they were run by greedy, disgusting people. If they actually brought something and cared about market and customers, they wouldn't be in this position.
Ideal for who?
Consumers, customers, us!
Last edited:
Strix Halo would have been cancelled (bc OEMs would not have been interested) if not for AI Hype. Hence all the high memory configs.
I have a feeling that Nova-AX will be cancelled and we'll see if any nVidia products end up being released.
I have a feeling that Nova-AX will be cancelled and we'll see if any nVidia products end up being released.
I think not. Intel is anxious to show the world that they can do stuff better than AMD. Why else would they bet their company on tile strategy despite not having enough experience with it to put out a good enough product with tiles?
The sheer amount of slides they have created for Panther Lake speaks volumes about their desperation. It's like they were told, THIS IS IT. This thing's gotta sell or we are finished!
Xe3 looks good. If it wasn't for the Nvidia deal I would be much more positive about the GPU division. The Nvidia deal is antithetical to their GPU plans.
Few months ago I heard ARC team was severely slashed and big things canned, adding up the news of Nvidia joins, I guess what you say is likely to happen.
The team failed the PPA of every ARC too hard, if they insist to participate in NVL-AX's graphic, I can't imagine how hard they will fail. Not to mentioned the purpose of such product is doubtful.
They laid off the graphic team and Nvidia joined months later. Not antithetical but very logical.
Last edited:
If the B770 is released, we know that ARC is fine and has a future. If not, that will be a bad sign.
poke01
Diamond Member
- Mar 8, 2022
- 4,338
- 5,655
- 106
There are very few people in the GPU division. It’s all crickets. The NV deal is smart.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 23K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter