Intel ARC GPUs, Xe Architecture for dGPUs [2022-]

[MfA]

Why would Intel try to reverse engineer any of Nvidia's interfaces when it means playing by their rules ?

Because they can't offer developers as good a deal, they can compensate with more money but it might be cheaper to be able to use the same interface for a drop in replacement. No need to reverse engineer once NVIDIA stops being coy and releases the new DLSS without NDA, can just look at UE/Unity. Of course if it's patent encumbered that won't help.

I didn't say Intel was using the NPU, I said there was a good opportunity for them to do it. As for whether it's possible, Microsoft suggests it in the DirectSR documentation.

 

[Erinyes]

24 % more performance at 45 % lower consumption at the same node is underwhelming?

I'm not sure what Phoronix is testing, they haven't listed much detail. From the Windows gaming reviews I had seen (such as AT), the performance was 15-20% higher than 780M in most cases. Have you come across a good Windows gaming review which shows better? I'd like to explore further.

Strix Point does seem memory bandwidth limited as they didn't increase the memory speed. While it certainly might be appreciably better at lower power levels as they could clock lower, the absolute performance increase to me is a bit underwhelming tbh.

 

[Kaotik]

I'm not sure what Phoronix is testing, they haven't listed much detail. From the Windows gaming reviews I had seen (such as AT), the performance was 15-20% higher than 780M in most cases. Have you come across a good Windows gaming review which shows better? I'd like to explore further.

Strix Point does seem memory bandwidth limited as they didn't increase the memory speed. While it certainly might be appreciably better at lower power levels as they could clock lower, the absolute performance increase to me is a bit underwhelming tbh.

Are we looking at the same AnandTech review?

The AMD Ryzen AI 9 HX 370 Review: Unleashing Zen 5 and RDNA 3.5 Into Notebooks

www.anandtech.com

Ryzen AI HX 370 IGP is beating 7940HS IGP by
CoH: 23 %
Cyberpunk: 69 %
F1 23: 25,6 %
Returnal: 57,6 %
TW:WH3: 43,4 %
Average: 43,8 %
Despite having 7W lower TDP. Heck, it's beating even the desktop parts with 65W silly

 

[Erinyes]

Are we looking at the same AnandTech review?

The AMD Ryzen AI 9 HX 370 Review: Unleashing Zen 5 and RDNA 3.5 Into Notebooks

www.anandtech.com

Ryzen AI HX 370 IGP is beating 7940HS IGP by
CoH: 23 %
Cyberpunk: 69 %
F1 23: 25,6 %
Returnal: 57,6 %
TW:WH3: 43,4 %
Average: 43,8 %
Despite having 7W lower TDP. Heck, it's beating even the desktop parts with 65W silly

You're right, looks like my memory is off. The performance is a fair bit better and a lot more power efficient as you note. Well Lunar Lake definitely has it's work cut out then.

 

[snarfbot]

Lunar lake is generally better from what I've seen. By a small but not insignificant margin.

 

[no-X]

According to NotebookCheck, Lunar Lake has a bit slower iGPU than Strix Point, which is ~5 % faster. A few percents don't matter, the main problem is the support. Lunar Lake doesn't offer frame-gen, features for reducing latencies nor driver-based upscaling. Also price for this performance isn't really interesting. You can buy cheaper laptop with dedicated GPU for a fraction of the price, and you'll get several times higher gaming performance.

 

[DegustatoR]

Intel Arc "Battlemage" GPU Surfaces with 20 Xe2 Cores, 2.85 GHz Clock Speed, and 12 GB VRAM

Intel's upcoming Arc "Battlemage" G21 GPU has made an appearance in Geekbench benchmarks, offering a glimpse into the future of the company's discrete graphics offerings. This next-generation GPU, part of Intel's Xe2 graphics architecture, shows promising performance that puts it almost on par...

www.techpowerup.com

 

[tuna]

According to NotebookCheck, Lunar Lake has a bit slower iGPU than Strix Point, which is ~5 % faster. A few percents don't matter, the main problem is the support. Lunar Lake doesn't offer frame-gen, features for reducing latencies nor driver-based upscaling. Also price for this performance isn't really interesting. You can buy cheaper laptop with dedicated GPU for a fraction of the price, and you'll get several times higher gaming performance.

What kind of features can reduce latencies?

 

[Kaotik]

What kind of features can reduce latencies?

I'm thinking input latency aka Anti-lag on AMD, Reflex on NVIDIA

 

[Subtlesnake]

According to NotebookCheck, Lunar Lake has a bit slower iGPU than Strix Point, which is ~5 % faster. A few percents don't matter, the main problem is the support. Lunar Lake doesn't offer frame-gen, features for reducing latencies nor driver-based upscaling. Also price for this performance isn't really interesting. You can buy cheaper laptop with dedicated GPU for a fraction of the price, and you'll get several times higher gaming performance.

That's in standard mode, which runs at 17W. The Asus VivoBook S 14 it's competing against has a 54W max TDP according to its NotebookCheck review.

Edit: The Asus Zenbook S16 apparently runs at up to 33W.

 

[no-X]

What kind of features can reduce latencies?

AntiLag / AntiLag 2

That's in standard mode, which runs at 17W. The Asus VivoBook S 14 it's competing against has a 54W max TDP according to its NotebookCheck review.

Edit: The Asus Zenbook S16 apparently runs at up to 33W.

17W for PL1, actual PL2 is 37 watts.

 

[Erinyes]

You have to also keep in mind that Lunar Lake iGPU is on 3nm vs 4nm for Strix Point, and it has faster 8533 mhz memory speed vs 7500 mhz, both of which will give it a decent advantage.

That said, it is probably the first time Intel is very competitive on iGPU perf, with good power use as well. AMD has it's work cut out for Medusa Point as Panther Lake is supposedly 12 next gen Xe3 cores, but similar memory bandwidth, unless they go for 9600mhz or higher which is not yet officially certified by JEDEC.

 

[tuna]

I'm thinking input latency aka Anti-lag on AMD, Reflex on NVIDIA

Those are purely software, and really a bandaid around bad CPU-GPU synchronization.

 

[Kaotik]

Those are purely software, and really a bandaid around bad CPU-GPU synchronization.

Pure software or not, Intel doesn't currently offer anything similar

 

[Kaotik]

That's in standard mode, which runs at 17W. The Asus VivoBook S 14 it's competing against has a 54W max TDP according to its NotebookCheck review.

Edit: The Asus Zenbook S16 apparently runs at up to 33W.

Tom's Hardware has Strix at 28W beating Lunar at 30W

We benchmarked Intel's Lunar Lake GPU with Core Ultra 9 — overall performance improved, though driver concerns still remain [Updated]

Intel's integrated Battlemage GPU put to the test.

www.tomshardware.com

 

[arandomguy]

I know the interest in directly comparing these but there some additional complexities involved that make it difficult to do so.

The power/frequency curves for both chips aren't directly overlapping each other. You're not going to find an actual single "fair" TDP setting for both. I believe Lunar Lake's min-max TDP configurable range is is half that of Strix Point's for example. So it's unlikely any single point on the curve is going to fair to either side.

Secondly with laptops the actual configured TDP limit really needs to be validated during the test runs even if we are going to compare those. What the processors actually get set at can be different from what's actually being drawn.

 

[Subtlesnake]

Tom's Hardware has Strix at 28W beating Lunar at 30W

We benchmarked Intel's Lunar Lake GPU with Core Ultra 9 — overall performance improved, though driver concerns still remain [Updated]

Intel's integrated Battlemage GPU put to the test.

www.tomshardware.com

Only 3 games tested though. Obviously it's going to go back and forth depending on the title.

 

[Erinyes]

Only 3 games tested though. Obviously it's going to go back and forth depending on the title.

Here's a good review from Jarrod's tech comparing more games. Strix Point is ~11% faster than Lunar Lake but there are some games where LL is slower than ML, if not LL would be closer. Could be driver related? Or perhaps CPU limited where the higher core count of ML made a difference.

 

[trinibwoy]

Secondly with laptops the actual configured TDP limit really needs to be validated during the test runs even if we are going to compare those. What the processors actually get set at can be different from what's actually being drawn.

This seems to be a real problem. Hardware Canucks has LL soundly beating Strix Point at matched TDP. Their results are very different to all the other reviews I’ve seen.

 

[T2098]

I know the interest in directly comparing these but there some additional complexities involved that make it difficult to do so.

The power/frequency curves for both chips aren't directly overlapping each other. You're not going to find an actual single "fair" TDP setting for both. I believe Lunar Lake's min-max TDP configurable range is is half that of Strix Point's for example. So it's unlikely any single point on the curve is going to fair to either side.

Secondly with laptops the actual configured TDP limit really needs to be validated during the test runs even if we are going to compare those. What the processors actually get set at can be different from what's actually being drawn.

Yeah, these are really tough to review. If it was me, I'd be testing these with the laptop lid closed, and a single standard external display, same external display for each device under test. The internal screen is a very significant portion of total laptop power draw in most cases, and that variable needs to be controlled.

And then I'd be using a standard USB-C PD power adapter, same for each device, measuring power draw at the wall, and then try to plot a performance vs cTDP curve for each device (if configurable.)

You'd only need to use a handful of representative benchmarks to get a good idea of how they scale, I'd try to do a nearly pure GPU test for one of them, along with a game that was heavy on both the CPU and the GPU side of things to see how well the system balances the power demands of each half of the system with the limited TDP budget available.