News Intel reportedly prepping supercharged Nova Lake-AX mobile chips for gaming — Team Blue’s high-performance APU to rival AMD’s Strix Halo

[Admin]

A popular Intel leaker says that Team Blue is preparing a competitor to AMD's Strix Halo chips, slated to arrive in 2027.

Intel reportedly prepping supercharged Nova Lake-AX mobile chips for gaming — Team Blue’s high-performance APU to rival AMD’s Strix Halo : Read more

 

[TheSecondPower]

The Strix Halo packaging is more similar to Meteor Lake and Arrow Lake packaging than other AMD lines. In Strix Halo AMD uses power-optimized bins of their desktop chiplets, and assuming Nova Lake is similar to Arrow Lake then Nova Lake-AX could do the same thing. It would need a SoC tile with a wider memory interface and a huge GPU tile and probably also a large GPU cache like Strix Halo has.

I kind of hope Intel goes another direction. What I wanted from Strix Halo was a mobile CPU that can game without a discrete GPU. (So 8 cores/16 threads and 40 CUs.) What we got was a 16-core CPU with 40 CUs. It's neat but too expensive for a simple portable gaming laptop.

 

[thestryker]

I kind of hope Intel goes another direction. What I wanted from Strix Halo was a mobile CPU that can game without a discrete GPU. (So 8 cores/16 threads and 40 CUs.) What we got was a 16-core CPU with 40 CUs. It's neat but too expensive for a simple portable gaming laptop.

Unfortunately the way I see this is an easy way to tick the "AI" checkbox for shareholders. This is something Intel can do to address "AI" without developing new technology.

In theory PTL should be pretty good for gaming since it should have at least one SKU with 12 Xe3 cores. Though this is still going to be a situation relying on upscaling for higher graphics settings. Personally I just hope that it scales well at lower wattage for handheld as LNL is the best 15W performer now and it uses 8 Xe2 cores.

 

[Notton]

Strix Halo has an 8C/16T, 32CU config, the Max 385.
They also have a really weird 6C/12T, 16CU config with the Max 380. Presumably, it's still better than the 370HX.

The next iteration Medusa Halo? is rumored to be 48CU and 384-bit mem bus.

 

[usertests]

Should be easy enough for them to scale things up with chiplets and make something like Strix Halo or Apple M4 Max/etc. I also want to see them take on AMD's 3D V-Cache, both in desktop and in APUs.

I kind of hope Intel goes another direction. What I wanted from Strix Halo was a mobile CPU that can game without a discrete GPU. (So 8 cores/16 threads and 40 CUs.) What we got was a 16-core CPU with 40 CUs. It's neat but too expensive for a simple portable gaming laptop.

You can point to pretty much any AMD APU and complain about it having more than enough CPU performance, but not enough GPU performance. Renoir, Cezanne, Rembrandt, Phoenix, Strix Point were all designed with an expectation that it can be paired with dGPUs, but even the Halo hits 4060 mobile performance while giving you 16 CPU cores. Though the 8-core, 32 CU model should be getting most of the GPU performance (>80%?) while slashing the cores in half.

Maybe the only product you can point to that bucks the trend is the custom APU in the Steam Deck, quad-core Zen 2 with 8 CUs RDNA2. So it will be very interesting to see what chip makes it into the Steam Deck 2.

In theory PTL should be pretty good for gaming since it should have at least one SKU with 12 Xe3 cores. Though this is still going to be a situation relying on upscaling for higher graphics settings. Personally I just hope that it scales well at lower wattage for handheld as LNL is the best 15W performer now and it uses 8 Xe2 cores.

Hopefully the new 18A node is enough to keep it on par with Lunar Lake power efficiency, even as it drops on-package memory and uses larger dies with more cores.

They also have a really weird 6C/12T, 16CU config with the Max 380. Presumably, it's still better than the 370HX.

It has to be a nothingburger SKU that some OEM asked for. There's no good reason to disable 40 CUs down to 16 CUs, which is what must be happening with that. But if someone pays for it, AMD would make it, and it takes care of the most garbage yields.

The next iteration Medusa Halo? is rumored to be 48CU and 384-bit mem bus.

LPDDR6 was just announced. I think it boosts bits per channel in a way that makes that 384-bit bus reachable with the same number of chips.

 

[TheSecondPower]

Unfortunately the way I see this is an easy way to tick the "AI" checkbox for shareholders. This is something Intel can do to address "AI" without developing new technology.

In theory PTL should be pretty good for gaming since it should have at least one SKU with 12 Xe3 cores. Though this is still going to be a situation relying on upscaling for higher graphics settings. Personally I just hope that it scales well at lower wattage for handheld as LNL is the best 15W performer now and it uses 8 Xe2 cores.

Strix Halo does seem to be a hit for running LLMs and is selling for a great deal of money so Intel may be smart to copy that formula.

Phoenix, Meteor Lake, Strix Point, Lunar Lake, and Arrow Lake have mostly different iGPUs yet all come pretty close in performance. I'm not pleased with where they land, because for nearly any semi-modern game the settings have to be turned way down. They're all held back by the shared 128-bit CPU DDR/LPDDR RAM they use. Even the most modest dGPUs have dedicated 128-bit GDDR memory. An APU with 192-bit memory or Infinity Cache would be a lot more usable in games.

 

[Mr Majestyk]

If reports of AMD sticking to RDNA3.5 for Medusa Halo, while Nova Lake will get Xe3 with some aspects of Xe4, then Intel could deliver a knockout blow. Already even Lunar Lake's Xe2 can compete with AMD 890M. A full fat Xe3-Xe4 iGPU using even DDR5X 10000MT/s in a system with 128GB on 256 bit bus and allowing 120W+ I think would cause AMD a lot of grief. Let alone what Niviudia has in store.

 

[thestryker]

They're all held back by the shared 128-bit CPU DDR/LPDDR RAM they use.

I imagine LPDDR6 should be a thing by the time NVL launches so that ought to be around 50% more bandwidth which will help. Xe2 seems to do more with its bandwidth than RDNA 3/3.5 so that could mean better things for Xe3.

An APU with 192-bit memory or Infinity Cache would be a lot more usable in games.

I think for the big APUs this will continue to be the case. We'll see wider bus and dedicated cache for the GPU side. I don't really think we'll see either of these things outside of those.

In theory they could target the lower end gaming laptop segment, but that would require a different overall design because those rely on cheap CPUs with the cheap GPUs. Historically AMD and Intel have largely put the best IGP with higher end CPU so the price premium is quite high.

From a technical standpoint there is nothing of course stopping them from putting a wider bus and better GPU on a midrange part and selling accordingly.

 

[usertests]

If reports of AMD sticking to RDNA3.5 for Medusa Halo, while Nova Lake will get Xe3 with some aspects of Xe4, then Intel could deliver a knockout blow.

We already have a leaker sharing specs but doubting this thing will exist: https://videocardz.com/newz/intel-n...-cores-xe3p-gpu-with-384-eus-launch-uncertain

I think for the big APUs this will continue to be the case. We'll see wider bus and dedicated cache for the GPU side. I don't really think we'll see either of these things outside of those.

I think 16-32 MiB of Infinity Cache has to come to the 128-bit APUs eventually. 16 MiB would match the 6500 XT and be optimized for 1080p, if that. Strix Halo gets 32 MiB.

MLID has been teasing the possibility of Infinity Cache in Medusa Point, but not committing to it.

 

[TheSecondPower]

I imagine LPDDR6 should be a thing by the time NVL launches so that ought to be around 50% more bandwidth which will help. Xe2 seems to do more with its bandwidth than RDNA 3/3.5 so that could mean better things for Xe3.

I think for the big APUs this will continue to be the case. We'll see wider bus and dedicated cache for the GPU side. I don't really think we'll see either of these things outside of those.

In theory they could target the lower end gaming laptop segment, but that would require a different overall design because those rely on cheap CPUs with the cheap GPUs. Historically AMD and Intel have largely put the best IGP with higher end CPU so the price premium is quite high.

From a technical standpoint there is nothing of course stopping them from putting a wider bus and better GPU on a midrange part and selling accordingly.

I doubt LPDDR6 will change the relative performance of iGPUs to dGPUs; as dGPUs continue to advance iGPUs need faster memory just to stay the same distance behind. Although I'm remembering now that Arrow Lake or Meteor Lake added more cache to its Xe cores and you said Xe2 does more with the same memory, so that's promising. (Although Xe2 is also paired with the fastest memory from its time.)

Lunar Lake is neat; it has a modest CPU but didn't cheap out on the GPU. That's what handhelds need. With several handhelds on the market now maybe there's actually pressure to build a small APU with a big GPU like the Steam Deck uses. But hopefully one with a bigger GPU.

 

[thestryker]

I doubt LPDDR6 will change the relative performance of iGPUs to dGPUs; as dGPUs continue to advance iGPUs need faster memory just to stay the same distance behind.

Currently (I doubt it will go higher, but you never know) the best 128-bit LPDDR5 can do is 153.6GB/s and I don't think anything in the PC space supports that. From the initial tool support we know 128-bit LPDDR6 can support up to 230.4GB/s (initially, but I don't know what the absolute cap is) which puts it into RX 6600 territory. If they could translate that bandwidth into close to equivalent performance that would be really good for an IGP on something which wasn't focused on having big graphics. I think the certainly do need to be standardizing some cache just for graphics though.

Lunar Lake is neat; it has a modest CPU but didn't cheap out on the GPU. That's what handhelds need.

Yeah it has allowed LNL to be the best at ~15W even when compared to Z2E. PTL is also going to have a 4p/0e/4lpe configuration, but the LPE are undoubtedly on the SoC tile. It seems likely that the only place the 12 Xe3 configuration will appear is 4p/8e/4lpe which I think is a missed opportunity. There is supposed to be a 4p/4e/4lpe with 10 Xe3 cores though so maybe this will be the handheld choice.

 

[usertests]

I doubt LPDDR6 will change the relative performance of iGPUs to dGPUs; as dGPUs continue to advance iGPUs need faster memory just to stay the same distance behind.

Even if the latest dGPUs are constantly 3-10x faster than the latest iGPUs are or whatever the range is, I think the resolution/FPS targets are moving slower. So if all graphics hardware becomes 10x faster over a period of time, it doesn't become 10x harder to run games at 1080p60/ultra. This is an oversimplification that ignores VRAM, raytracing, and a lot more, but close enough hopefully.

If that's true then large generational increases in bandwidth, like moving from LPDDR5X to LPDDR6, should push iGPU capabilities up, while smaller incremental increases, like the next-gen APU memory controllers supporting 5-15% faster memory of the same type, should help maintain the status quo. (I bet there's a widening memory scaling gap in graphics but I haven't looked into it.)

The 230 GB/s @thestryker mentioned? If that's true it would bring 128-bit LPDDR6 APUs past the graphics portion of the Xbox Series S's GDDR6, which is at 224 GB/s.

If we get Infinity Cache in Medusa Point, that would help it punch further above its weight, increasing the effective bandwidth. But from what I remember the average hit rate is dependent on the size and resolution. So I guess a 16-24 MiB entry level in Medusa would help 1080p a lot but not 1440p as much. Solid 1080p performance with no compromises would be good enough.

 

[TheSecondPower]

Yeah it has allowed LNL to be the best at ~15W even when compared to Z2E. PTL is also going to have a 4p/0e/4lpe configuration, but the LPE are undoubtedly on the SoC tile. It seems likely that the only place the 12 Xe3 configuration will appear is 4p/8e/4lpe which I think is a missed opportunity. There is supposed to be a 4p/4e/4lpe with 10 Xe3 cores though so maybe this will be the handheld choice.

That's especially sad given the flexibility of Intel's advanced packaging. AMD mobile chips are monolithic so the little CPU and little GPU are doomed to the same die. But Intel's chips could combine a little CPU tile with a big GPU tile, Intel just chooses not to. The same could be said of Strix Halo; the 40 CU GPU is only offered with 2 CPU chiplets, but without even a packaging change it could be offered with just one chiplet for gaming laptops, AMD just chooses not to.