News Intel Nova Lake CPUs reportedly get a GPU overhaul — Xe3 Celestial and Xe4 Druid IPs used for graphics, media, and display

Desktop iGPU improvements are arguably more important for Intel, because they go into so many OEM office PCs, and end up flooding the used market later.

Fully enabled Arrow Lake-S iGPUs are faster than the 5700G (Vega 8). Also faster than the 8500G (Radeon 740M - RDNA3 @ 4 CUs) at 1080p but slower at 720p according to TechPowerUp's review.

Only Core Ultra 5 235/225 models have compromised iGPUs (3 and 2 Xe Alchemist cores respectively).

If Intel moved to 8 Xe3 Celestial cores for Nova Lake-S, maybe you could see tripled performance, around an Arc A380 or RX 6400. Just a guess. That would have legitimate low-end 1080p performance.

On the media engine side, Lunar Lake is the latest. It adds H.266 (VVC) decode only, and I think AV1/H.265/VP9 support is the same, so only up to 10-bit encode for AV1/VP9, and 12-bit encode with 4:2:0 chroma subsampling for H.265. So there's room for improvement. There hasn't been much news about the AV2 codec so it probably won't be ready to go into a 2026 product like Nova Lake.
 
Last edited:
Meteor Lake splits the media and display capabilities from core graphics. The media and display units were placed on a separate System-on-Chip (SoC) chiplet, manufactured using TSMC's N6 process, while the graphics engine resided on a separate tile produced with TSMC's N5 technology.
One upshot of this could be that Quicksync support becomes available on '-F/-KF' suffix processors, which is not the case currently (with the encode block being part of the GPU and disabled along with it). Though with the GPU moving off-die, it's entirely possible that it wont be subject to the yield behaviour that created the need for the '-F' line in the first place.
 
One upshot of this could be that Quicksync support becomes available on '-F/-KF' suffix processors, which is not the case currently (with the encode block being part of the GPU and disabled along with it). Though with the GPU moving off-die, it's entirely possible that it wont be subject to the yield behaviour that created the need for the '-F' line in the first place.
ARL still has F SKUs (below 285) and disabled Quick Sync on them despite the display/media engines being in the SoC tile and there being a separate Graphics tile.
 
If Intel moved to 8 Xe3 Celestial cores for Nova Lake-S, maybe you could see tripled performance, around an Arc A380 or RX 6400. Just a guess. That would have legitimate low-end 1080p performance.
I would be surprised if they did this even though utilizing tiles makes it very much possible. If they shifted to using EMIB for connecting the Graphics tile it would be trivial to change the tile configuration. I'm not sure this would be a fiscally responsible move compared to using Foveros for the whole die as they do now though.

If it was you could see F SKUs that simply didn't have a graphics tile, regular SKUs that had binned 4 core tile and then maybe G SKUs with binned 12 core tile (assuming PTL leaks are correct about top bin being 12 Xe3 cores). If the performance was there I could easily see them commanding a decent premium since there would be no performance disadvantages as there have been with AMD's APUs.
 
Intel's next-gen Nova Lake CPUs for desktops are expected to use Xe3 (Celestial) for the integrated graphics, while media and display move to Xe4 (Druid).

Intel Nova Lake CPUs reportedly get a GPU overhaul — Xe3 Celestial and Xe4 Druid IPs used for graphics, media, and display : Read more
And who will produce it? TSMC, seized together with Taiwan by China until the end of the year, or maybe IF, which is losing contracts because it is unable to supply enough wafers because China has cut it off from raw materials? The latest attempts at sanctions on EDA show how panicked the US is
 
Desktop iGPU improvements are arguably more important for Intel, because they go into so many OEM office PCs, and end up flooding the used market later.

Fully enabled Arrow Lake-S iGPUs are faster than the 5700G (Vega 8). Also faster than the 8500G (Radeon 740M - RDNA3 @ 4 CUs) at 1080p but slower at 720p according to TechPowerUp's review.

Only Core Ultra 5 235/225 models have compromised iGPUs (3 and 2 Xe Alchemist cores respectively).

If Intel moved to 8 Xe3 Celestial cores for Nova Lake-S, maybe you could see tripled performance, around an Arc A380 or RX 6400. Just a guess. That would have legitimate low-end 1080p performance.

On the media engine side, Lunar Lake is the latest. It adds H.266 (VVC) decode only, and I think AV1/H.265/VP9 support is the same, so only up to 10-bit encode for AV1/VP9, and 12-bit encode with 4:2:0 chroma subsampling for H.265. So there's room for improvement. There hasn't been much news about the AV2 codec so it probably won't be ready to go into a 2026 product like Nova Lake.
I'd say that improvements aren't what really matters, what matters is that it's there and is able to do all the normal 2D computer things. The fact that it can run games, in the OEMs' mind, is nothing more than an afterthought.
 
I'd say that improvements aren't what really matters, what matters is that it's there and is able to do all the normal 2D computer things. The fact that it can run games, in the OEMs' mind, is nothing more than an afterthought.
Yet Intel still basically doubled iGPU performance going from Alder/Raptor Lake to Arrow Lake, leapfrogging the Radeon 610M iGPU in AMD's desktop non-APUs.

If they do nothing at all but swap Xe1 with Xe3, that should yield a substantial improvement. But they could have plenty of die area to add additional cores, if the node is much better. The Arrow Lake-S iGPU tile is made on TSMC N5P. If Intel eats their own dog food for Nova Lake... 18A?
 
But will AMD still persist with their trash tier iGPU in Medusa Ridge and insult us with 2CU's yet again.
I'd just consider it a wise expense of transistors where they matter most.

Spending transistors on an iGPU where either a dGPU is present or the use case is CPU-only, is mostly a waste. I was surprised they even put in an iGPU back in with Zen 4.

And once I actually got off my high rocker and started testing it, I found it was rather capable and useful, even with a dGPU plugged in. Because it a) is a rather decent 2D display engine for 4k at HDR and 144 Hz, the best my screens are capable of and b) because it allows dGPU pass-through processing without much of a performance penalty. It even allowed refresh rates and pixel depths the dGPU didn't support natively on its physical outputs. The biggest surprise was how easily it got along with pretty near every dGPU I threw at it, from ancient Maxwell Nvidias to an Intel B580.

And then it allows my to have a display output for the hypervisor when I use the dGPU via pass-through in a VM, which can be handy.

Imagine having a Strix Halo iGPU and then adding the dGPU for real gaming or AI power: all those extra transistors wasted!

Intel in the past has made CPUs which used more than half of the die area spent on an iGPU, that they basically gave away for free, e.g. GT3 variants of Broadwell and Skylake which even included 64/128MB of eDRAM, only to keep the chipset and dGPU competition from gaining a foothold.

AMD crashed that gate by converting that iGPU chip area into twice the number of cores instead and the rest is history.

Sometimes it's just not that obvious how clever they are.
 
  • Like
Reactions: thestryker