• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Intel "Nova Lake-AX" Specifications Surface: 28 CPU Cores, 48 Xe3 GPU Cores, and LPDDR5X

AleksandarK

News Editor
Staff member
Joined
Aug 19, 2017
Messages
3,255 (1.13/day)
Yesterday, we covered the leak of Intel's plan to take on AMD in the APU space with a "Nova Lake-AX" SoC that features a relatively powerful CPU core configuration paired with an enhanced Xe3 GPU core cluster. Today, one of the most reliable leakers, Raichu, confirms what the new Nova Lake-AX configuration could look like. When it comes to the CPU, the AX SKUs will feature a downsized CPU cluster, utilizing only a single tile with eight "Coyote Cove" P-cores and 16 "Arctic Wolf" E-cores, alongside the four-core LPE island, totaling 28 cores. Where the Nova Lake-S is expected to have up to two of those CPU tiles, the AX variant cuts out the second tile to make space for one of the biggest iGPUs we've seen Intel put in its SoCs.

Coming in at 384 Execution Units (EUs), this roughly translates into 48 Xe3 cores, assuming the standard eight EUs per Xe core configuration. However, there could be some internal changes to the way Xe3 cores handle render slices, so the total number of Xe3 cores remains unknown, except for the 384 EU count. Intel is also pairing this SoC with LPDDR5X memory, which operates at 9,600 or 10,667 MT/s, providing sufficient bandwidth to the 28 CPU cores and 384 EUs over a 256-bit bus. Additionally, Raichu claims that the launch is uncertain, which means that Intel is likely evaluating the platform for profitability if it decides to produce it in high volumes. Similarly to AMD's APU, there could be significant interest, so we will have to wait and see if Intel greenlights the project for the masses. The picture below is our own modification of Arrow Lake-H breakdown, and not what the actual Nova Lake-AX SoC would look like.



View at TechPowerUp Main Site | Source
 
The B580 has 20 Xe² cores, with 48, this purported APU would put two of those in the ground. That'd make Celestial an insane gain over Battlemage. I want to believe, but... with the B770 nowhere in sight, and Intel's marketing quiet as a mouse... I have to be skeptical
 
Cant wait for this nova to explode.. whats the supposed TDP of this thing?!
 
Huuuh. There were rumors Intel made a pitch for next gen consoles. Kinda wonder if this was what they had in mind.
 
The B580 has 20 Xe² cores, with 48, this purported APU would put two of those in the ground. That'd make Celestial an insane gain over Battlemage. I want to believe, but... with the B770 nowhere in sight, and Intel's marketing quiet as a mouse... I have to be skeptical
As you said, B580 has 5 Render slices (20 Xe² cores, 160 XVEs(EUs)). Link
Yeah, Raichu mentioned 384EUs, which would mean 6x more than Lunar Lake(Arrow Lake) or 2.4x than B580, but let's be realistic.

There is not enough BW for that, B580 has 456 GB/s while this one only 341GB/s with 10,667 MT/s LPDDR5X.

Then there is still the question of size, B580 with only 160XVEs is already 272 mm² using 5N from TSMC.
Yeah, this would use N3 or N2, so there will be significant reduction in size not to mention some things would be on a different tile, but with 384XVEs(EUs), It should be still >150-200mm2 in my opinion and that's still huge when this is just the GPU tile.

I think those 384EUs are equivalent to 192EUs, but more capable, something like AMD did with RDNA3 and dual-issue.
Nothing else really makes sense to me to be truthful.
 
Last edited:
If those EUs are to be believed that should be some really impressive GPU performance for an SoC. I'd love to see it.

Of course-- they still need to compete at the software level if they want to take on AMD.
 
While I would like to see such a large APU, it seems unfeasible to supply enough bandwidth with a limited 256bit bus and current LPDDR5X technology. Not to mention the heat, if we're talking about laptops, that would be problematic.
 
Cant wait for this nova to explode.. whats the supposed TDP of this thing?!
I'm willing to bet that it's probably going to be less than you imagine. SOC design seems to have an advantage when it comes to efficiency compared to a dGPU setup with similar performance. Halo strix uses less power compared to an equivalent laptop. And RDNA 3 isn't as efficient as ADA. The real question is, will the AX CPU be tuned like an HX chip, or an H? There's at least a 120W difference between the top core 9 HX and the top core 9 H, when you average the measured power in laptop reviews. HX is basically a 200w CPU, when the H is more like a 77w CPU.
1752748526160.png
1752748644177.png
 
I would be inclined to believe that Intel changed something with the cores like Turing => Ampere and now report the number as twice as high.
 
While I would like to see such a large APU, it seems unfeasible to supply enough bandwidth with a limited 256bit bus and current LPDDR5X technology. Not to mention the heat, if we're talking about laptops, that would be problematic.

2.5x higher peak bandwidth than PTL-H with 2x more Xe cores.
 
Last edited:
While I would like to see such a large APU, it seems unfeasible to supply enough bandwidth with a limited 256bit bus and current LPDDR5X technology. Not to mention the heat, if we're talking about laptops, that would be problematic.
Afaik this is what current M4 pro chips have, they do step up for a 512 bit bus for the higher models though... so you may be right. Also m4 max struggles heat a bit - but if anything intel is the master of dealing with heat lol.

Will be interesting to see.
 
The 'internal' slide looks fake. The configuration of tiles is exactly the same as on the Arrow Lake S, and iGPU tile has a single row of four Xe cores. Where is "mega-iGPU" on this slide? Besides, 'Nova Lake AX" is allegedly supposed to be a mobility SKU with LPDDR memory support. The rendering shows a desktop CPU. The blue colour shades of fonts are not the same either. Someone added "Intel Nova Lake-AX" with a different shade of blue. What is this nonsense about?
 
Last edited:
The AMD gang with those "explosion" (konosuba tone) comments..

just 4 more actual cores from previous gen, hmmm..
 
Huuuh. There were rumors Intel made a pitch for next gen consoles. Kinda wonder if this was what they had in mind.
Even if they did make a pitch, the slide published in this article looks fake, because there is no mega-iGPU tile anywhere to be seen. The CPU presented looks exactly the same as Arrow Lake S, with tiny iGPU tile containing four Xe cores. I don't think they would ever show such slide to Sony or Microsoft. It'd be embarrassing.
 
Cant wait for this nova to explode.. whats the supposed TDP of this thing?!
Supernova level of heat.

I'm willing to bet that it's probably going to be less than you imagine. SOC design seems to have an advantage when it comes to efficiency compared to a dGPU setup with similar performance. Halo strix uses less power compared to an equivalent laptop. And RDNA 3 isn't as efficient as ADA. The real question is, will the AX CPU be tuned like an HX chip, or an H? There's at least a 120W difference between the top core 9 HX and the top core 9 H, when you average the measured power in laptop reviews. HX is basically a 200w CPU, when the H is more like a 77w CPU.
View attachment 408121View attachment 408122
That version of Halo strix is under powered. Halo strix can go much higher, 140w, then settles down to 120w. I think that version in the asus z13 is rated at 90w.
 
Things to add:

Backside Power Delivery may allow voltage reduction, and therefore lower heat.

Higher shader execution units than the bandwidth or TDP allows still provides performance in modern shader-heavy games.

2x performance of a B580, in a cost effective APU is ideal for a next-gen Xbox.
 
As you said, B580 has 5 Render slices (20 Xe² cores, 160 XVEs(EUs)). Link
Yeah, Raichu mentioned 384EUs, which would mean 6x more than Lunar Lake(Arrow Lake) or 2.4x than B580, but let's be realistic.

There is not enough BW for that, B580 has 456 GB/s while this one only 341GB/s with 10,667 MT/s LPDDR5X.

Then there is still the question of size, B580 with only 160XVEs is already 272 mm² using 5N from TSMC.
Yeah, this would use N3 or N2, so there will be significant reduction in size not to mention some things would be on a different tile, but with 384XVEs(EUs), It should be still >150-200mm2 in my opinion and that's still huge when this is just the GPU tile.

I think those 384EUs are equivalent to 192EUs, but more capable, something like AMD did with RDNA3 and dual-issue.
Nothing else really makes sense to me to be truthful.

I agree, but it will also honestly depend on how efficient is the use of the available bandwidth, though. If Intel manages to improve their memory compression algorithm, it could well be competitive.
 
We've just acquired PCs and laptops at work that all use Core Ultra processors with some flavour of Xe graphics and the driver headaches just for doing office work have been never ending. Screens not waking from sleep, display corruption on videos and video calls, outright BSODs, driver resets, absolutely terrible. I can only imagine how flaky they are if you try to do things more demanding that MS Office, writing PBI reports, and chatting on Teams.

They're going to need to fix this and fix it properly to be taken seriously, these are well into "send them back" territory now. Odd because I can't remember serious intel GPU driver issue for desktop use in the 20+ years they've been shipping integrated graphics, chipset and CPU based.
 
Back
Top