Testing the performance of AMD Radeon 780M & 760M iGPUs with new drivers

[Redaktion]

Nearly three months after we initially reviewed the new integrated Radeon 780M, AMD has now finally released official drivers for the GPU. We are thus taking another look at the performance of both the Radeon 780M and the less-powerful Radeon 760M using synthetic as well as gaming benchmarks.

https://www.notebookcheck.net/Testing-the-performance-of-AMD-Radeon-780M-760M-iGPUs-with-new-drivers.740311.0.html

[Teskatlipoka]

780M is 21% faster in F1 2021 Full HD Ultra settings, that's the biggest difference in performance.
Why is the difference only this much and the rest is even less?Is bandwidth from DDR5 5600 not enough or It is power limited and can't clock high enough?
If you provided clockspeeds and power consumption in the game where the difference is the biggest, in this case F1 2021, It would be great.

[Neenyah]

Add CSGO (will be fully replaced with CS2 soon) to your tests please.

[Hotz]

780M is 21% faster in F1 2021 Full HD Ultra settings, that's the biggest difference in performance.
Why is the difference only this much and the rest is even less?

Maybe 8 CUs (as in the 760m) is the SweetSpot. As was the Vega 8 in previous generation.

[Lennitt]

These numbers make no sense, completely out of line with other reviews. These laptops must be defective

[Hotz]

These numbers make no sense, completely out of line with other reviews. These laptops must be defective

Which other reviews? I've watched some reviews on youtube, but most are cheating with FSR, and thus are not comparable with Notebookcheck reviews.

Regarding the performance of the 760m vs the 780m: when checking some youtube videos, it seems that the Vega 8 vs Vega 11 had similar results. The performance improvement was only 20%, but should have been 37% (according to the increase of CUs). So the performance doesn't seem to scale linearly anymore.

I guess this is an architectural feature (or limitation). In fact AMD itself said back at Vega 8, that 8 CUs is the SweetSpot and more CUs are not worth it. Maybe it's the same with Rembrandt and Phoenix. Afaik RDNA2 and RDNA3 are not completely new architectures, but still built on Vega. At least I've read something like this some months ago (not sure where exactly).

Please correct me if I'm wrong. I'm also just trying to make sense of it.

[Hotz]

Just checked a comparison video from Hubwood at Youtube - he gets similar values. The 780m only has about 27% improvement compared to the 760m. Not that much.

Also, his overlay shows that the 760m only gets 45W, while the 780m gets 55-65W.


Imagine giving it 65W, it may be even on par with the 780m. It's baffling... (but would again confirm the "Sweet Spot")

[Cruz]

While many exoected support in 23.7.2, the 700m series isnt listed as being supported in the driver official Release Notes, list stops at 600 series.  They often work anyway (like with 6700s & 6800s that are still not listed as having support a yr and a half after launch) but a few G14 2023 owners found they only partially worked w that model, some major problems made it impractixal to use. Details in the G14 subreddit.

My question is how did you determine they are intended to be officially supported?  Hopefully its not just by Amd website pointing u to the driver after keying in your hardware.  Id think Release Notes would be final word on whether supported.

[Mr Majestyk]

I will say the DDR5 5600 is totally inadequate to provide the needed bandwidth for the 780M. iGPU need bandwidth bandwidth and more bandwidth. Given next year AMD is releasing Sarlak with 16 cpu cores and a 40 CU RDNA3.5 iGPU, pray they support at least DDR5 7200, or even better the new LPDDR5X 9600 memory just announced by SK Hynix.

[Mr Majestyk]

I meant to say LPDDR5T.

[Hotz]

@Mr Majestyk:

LPDDR5 is generally soldered RAM, and therefore a bad candidate for comparison tests (it's also harder to find comparable devices).

[DS2]

These numbers make no sense, completely out of line with other reviews. These laptops must be defective

Which other reviews? I've watched some reviews on youtube, but most are cheating with FSR, and thus are not comparable with Notebookcheck reviews.

Regarding the performance of the 760m vs the 780m: when checking some youtube videos, it seems that the Vega 8 vs Vega 11 had similar results. The performance improvement was only 20%, but should have been 37% (according to the increase of CUs). So the performance doesn't seem to scale linearly anymore.

I guess this is an architectural feature (or limitation). In fact AMD itself said back at Vega 8, that 8 CUs is the SweetSpot and more CUs are not worth it. Maybe it's the same with Rembrandt and Phoenix. Afaik RDNA2 and RDNA3 are not completely new architectures, but still built on Vega. At least I've read something like this some months ago (not sure where exactly).

Please correct me if I'm wrong. I'm also just trying to make sense of it.

Search for Daniel Owen. There are very interesting tests there

No. RDNA is a new architecture, much more advanced than GCN (including vega). Techspot did a very detailed analysis showing the difference.

[Hotz]

Search for Daniel Owen. There are very interesting tests there

No. RDNA is a new architecture, much more advanced than GCN (including vega). Techspot did a very detailed analysis showing the difference.

I did not find any video title from him about architectural differences. Also I don't know what you mean by "interesting tests". Interesting tests are also direct comparisons from Notebookcheck and Hubwood.

And while I found some articles and slides about the technical side of RDNA in the web, they are quite shallow (addressing only a few improved parts), and do not explain if the RDNA architecture has fundamentally changed from the GCN architecture or not. It rather looks like some parts were increased in size, but the overall architecture remains very similar.

And if that is not the case,
how can it be that the 760m (RDNA3, RDNA2) behaves almost the same as the Vega 8 (GCN)?
how can it be that in both cases the 8 CU version is the sweetspot, and higher performance require significant higher power than before?
Doesn't sound like a completely new architecture to me...

[O'Connell]

I will say the DDR5 5600 is totally inadequate to provide the needed bandwidth for the 780M. iGPU need bandwidth bandwidth and more bandwidth. Given next year AMD is releasing Sarlak with 16 cpu cores and a 40 CU RDNA3.5 iGPU, pray they support at least DDR5 7200, or even better the new LPDDR5X 9600 memory just announced by SK Hynix.

AMD updated driver of Zen4, to support up to 10000 DDR5. Phoenix (that is really good APU) is only capped by DDR% bandwith of low frequency of SODIMM DDR5 that miniPC marks has put inside their product. Think which performances they can have with hiher frequency, for this see ETA Prime video of GPD Win max 2 with LPPDR5 7500 and he OC it up to 40W R7 7840U---> impressive : https:  //youtu.be/ocJbhKtNbBM.
Again, if they (i intend minisforum/Beelink etc) wants they can even put Innosilicon next year with 10000 LPPDR5X like here https:  //hothardware.com/news/innosilicon-lpddr5x-dram-10000-mbps-speed. just to show how can be good RDNA 3.5 at 65W. I really don't care how many CU's will have them, but how fast their random memory share informations is matter. myabe even next 2/3 years will put HBM2/3 to reach level of GTX1080Ti.

[NikoB]

LPDDR5 is generally soldered RAM, and therefore a bad candidate for comparison tests (it's also harder to find comparable devices).

This is not a problem when at least 32 or 64GB are soldered. And Zen4 Phoenix supports up to 128GB of soldered memory. The greed of laptop manufacturers who sell you cheap memory at 3-4 times the market price is the real problem. It is on such components as ram and ssd that they make the main margin on suckers. In addition, the guarantee for all this as part of a laptop is 1-2 years maximum, and separately 5-10 years for a much lower price.