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

Page 259 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

DrMrLordX

Lifer
Apr 27, 2000
22,934
13,021
136
Ideal for who? You are really asking this question?

Yes.

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.

Are you still talking about AMD? Or Intel? Because I'm not sure which "they" is in play here. Intel is in their current position because they sucked at doing their jobs when in a position of advantage. AMD is in their current position because they executed when they needed to and targeted high margin markets successfully (after scraping by on semi-custom revenue for a bit). Notice that AMD didn't get where they are today on low-end mobile offerings.

Ideal for who?
Consumers, customers, us!

You shouldn't expect a desperate Intel and a gun-shy AMD to risk lowering their margins just so you can have a bigger iGPU on a budget mobile SoC. OEMs already don't want to sell those types of products, or else you'd see an i3 with a mobile 4070 or 5060 (or 5070). If OEMs don't want to produce such a product, odds are good Intel and AMD don't want to either, for similar reasons.

So it's not as consumer-friendly as you might think, if the company producing said product is losing money doing so, since the product won't stay on the market for long. Someone's got to make money or else they're not going to do it.

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 theory, and it's that Intel, NV, and AMD would like to kill off the mdGPU (for various reasons). Well maybe not NV but still. OEMs do stupid things that are suitable for OEMs but not necessarily for their suppliers (AMD, NV, Intel). AMD is probably not going to release any more mdGPUs and will just focus on halo and premium chips for people that want AMD graphics in a package suitable for mobile applications. Intel meanwhile appears to be integrating (somehow) NV products into their mobile lineup, taking configuration options away from the OEMs and putting that power into Intel's hands.

I could be wrong here, but within a few generations, it's entirely possible that most/all mobile platforms won't expose enough PCIe lanes for OEMs to integrate mdGPUs into products AND neither AMD nor NV will offer to sell mdGPUs to OEMs.
 

mikk

Diamond Member
May 15, 2012
4,306
2,393
136
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.


It is a beefed up Battlemage Xe2. There are some good optimizations and enhancements which might close some bottlenecks nicely, but I wouldn't call the changes significant architectural advancement. Xe3p (according to rumors) is "halfway Xe4 with some features backported from Xe4". There must be a reason why Petersen calls Xe3p significant architectural advancement. PTL Xe3 looks really good nevertheless.
 

DavidC1

Golden Member
Dec 29, 2023
1,894
3,037
96
It is a beefed up Battlemage Xe2. There are some good optimizations and enhancements which might close some bottlenecks nicely, but I wouldn't call the changes significant architectural advancement. Xe3p (according to rumors) is "halfway Xe4 with some features backported from Xe4". There must be a reason why Petersen calls Xe3p significant architectural advancement. PTL Xe3 looks really good nevertheless.
I don't care about feature changes. These products are a black box. 25-30% gains ISO a generation is significant. Can Xe3P do that over Xe3? If the checklists are significant but the gains are 10-15% I would say the big gains are Xe3 not Xe3P. If it offers 25-30% over Xe3 then the magnitude of changes are just as significant as Xe3, and no more. That's what matters to buyers anyway. Who cares about 18A, RibbonFET, WTFBBQ if it doesn't result in tangible gains?
 

regen1

Member
Aug 28, 2025
147
209
71
.....160GB LPDDR5X offering for H2'2026 or later that will be AI inference optimized around power efficiency and cost. It sounds interesting but technical details beyond those basics were light and it's going to be a long while before we see Crescent Island.


Key features include:

  • Xe3P microarchitecture with optimized performance-per-watt
  • 160GB of LPDDR5X memory
  • Support for a broad range of data types, ideal for “tokens-as-a-service” providers and inference use cases
Intel’s open and unified software stack for heterogeneous AI systems is currently being developed and tested on Arc Pro B-Series GPUs to enable early optimizations and iterations.
 
Last edited:

ToTTenTranz

Senior member
Feb 4, 2021
713
1,184
136
Looks like Xe3P might have native FP4, FP8 support

Panther Lake's Xe3 shader units seem to support FP8.

1760536511254.jpeg

And so does Panther Lake's new NPU:

1760536540019.jpeg
1760536598856.jpeg




Though FP8 is missing from the XMX engines, so FP8 matrix calculations' throughput on the GPU might actually be lower than on the NPU:

1760536815627.jpeg
 

regen1

Member
Aug 28, 2025
147
209
71
Panther Lake's Xe3 shader units seem to support FP8.

View attachment 132076

And so does Panther Lake's new NPU:

View attachment 132077
View attachment 132078




Though FP8 is missing from the XMX engines, so FP8 matrix calculations' throughput on the GPU might actually be lower than on the NPU:

View attachment 132079
Yeah but that's likely not extensive, we will get the details soon enough.
XMX engine remains unchanged from previous gen, just that more units have got packed since the units have increased per render slice.
Xe3P on other hand(from the slides of OCP 2025) indicates support for FP4/MXP4, FP8, likely with newer XMX units
 

regen1

Member
Aug 28, 2025
147
209
71
NVL-AX is 16 and 32 Xe3p afaik, it makes sense they can reuse this 32 Xe3p version for other GPUs as well.


edit


Is this a 2x6 XVE per core setup? Not sure how accurate this render is though.
It looks like 12XVEs per Xe3P unit ? In that case NVL-AX could be 32Xe3 even putting aside the confusion of old and new EU count convention.
(Config could be 32*12 = 384EUs for top).
But still hard to say anything with confirmation(if the render itself is accurate).
 

poke01

Diamond Member
Mar 8, 2022
4,288
5,622
106
It looks like 12XVEs per Xe3P unit ? In that case NVL-AX could be 32Xe3 even putting aside the confusion of old and new EU count convention.
(Config could be 32*12 = 384EUs for top).
But still hard to say anything with confirmation(if the render itself is accurate).
Is this on 512b? Could be since it’s MoP
 

mikk

Diamond Member
May 15, 2012
4,306
2,393
136
It looks like 12XVEs per Xe3P unit ? In that case NVL-AX could be 32Xe3 even putting aside the confusion of old and new EU count convention.
(Config could be 32*12 = 384EUs for top).
But still hard to say anything with confirmation(if the render itself is accurate).


That would give 50% more EUs over Xe2+Xe3 with the same Xe core count right? NVL-H 12Xe3p would get 50% more units with the same amount of Xe cores and we pretty much know it stays at 12 Xe cores. Of course we can't be sure the render image is accurate.
 

regen1

Member
Aug 28, 2025
147
209
71
That would give 50% more EUs over Xe2+Xe3 with the same Xe core count right? NVL-H 12Xe3p would get 50% more units with the same amount of Xe cores and we pretty much know it stays at 12 Xe cores. Of course we can't be sure the render image is accurate.
Yeah, seems so from the image but we will have to see for actual thing and implementation.
As per Intel though the focus should be on Xe units count rather than on number of EUs .
 

regen1

Member
Aug 28, 2025
147
209
71

vv.png
Lunar Lake seems extremely efficient in 10-30W range here(some of it also due to MoP, node, etc.).
Too bad they didn't do a game comparison test for Z2E vs 258V