Discussion AMD Cezanne/Zen 3 APU Speculation and Discussion

[dr1337]

https://twitter.com/x/status/1292844896374718470

Fresh leak out today, not much is known but at least 8cu's is confirmed. Probably an engineering sample, core count is unknown and clocks may not be final.

This is very interesting to me because cezanne is seemingly 8cu only, and it seems unlikely to me that AMD could squeeze any more performance out of vega. A cpu only upgrade of renoir may be lackluster compared to tigerlake's quite large GPU.

What do you guys think? Will zen 3 be a large enough improvement in APU form? Will it have full cache? Are there more than 8cus? Has AMD truly evolved vega yet again or is it more like rdna?

 

[maddie]

He's gotta throw Intel a bone once in a while. I doubt he'd want to pay for inferior products just for reviews.

Isn't he still working for Intel or did he leave?

 

[Hitman928]

Isn't he still working for Intel or did he leave?

He left PCPer for Intel and is still working for Intel.

 

[Asterox]

People are assuming vega has poor performance. It does not. Mobile Vega is extremely fast and power efficient. That is why AMD continues to use it.

Yes Vega is ok, or more that ok vs mighty Intel iGPU-s.

But keep in mind, that 8 CU RDNA2 in iGPU gaming will eat 8 CU Vega GPU for dinner.

We all now, how good Radeon VII/60 CU+HBM memory is vs Navi 40 CU GPU.

AMD Radeon RX 5700 XT and Radeon RX 5700 Review: New Prices Keep Navi In The Game

AMD's Radeon RX 5700 XT and Radeon RX 5700 deliver better performance than Nvidia's GeForce RTX 2060 and 2060 Super, but without real-time ray tracing support.

www.tomshardware.com

 

[Hitman928]

seems to even run better on 2500u than in Ryan's video

and this is the multiplayer not the campaign

That video has low (and 1 medium setting) and 50% resolution scale at 1080p. The video Ryan showed was high settings at 1080p with 100% resolution scaling. It really is a strong showing for Tigerlake but not knowing more details about power settings as well as it being single player only in Ryan's video makes it hard to judge how impressive it is. As always, we'll have to wait for independent reviews to see how it stacks up.

 

[TESKATLIPOKA]

RDNA1 has 40% better IPC than Polaris and Vega has pretty much the same IPC If not worse as Polaris.

 

[TESKATLIPOKA]

People are assuming vega has poor performance. It does not. Mobile Vega is extremely fast and power efficient. That is why AMD continues to use it.

RDNA1 is still better in absolute performance and performance/W, but performance/size ratio is probably on the side of Vega.

 

[moinmoin]

Has anybody done size measurements of each CU of all the designs on the different nodes? Didn't find anything searching around.

 

[Shivansps]

I'm starting to think that AMD is sticking with VEGA on the 7nm APUs for footprint reasons. From what I'm reading, the equivalent number of NAVI/RDNA CUs would consume a non-trivial amount more die area than the existing VEGA cores do. Given how tight AMD is on wafer capacity at the moment, I don't see them wanting to put any more die area towards the iGPU than they absolutely have to.

If Cezanne can bring the increased core capabilities of Zen3 and the increased effective cache performance of the unified 8 core CCX, and provide another 10% clock speed uplift on the Vega CUs, it will still be a worthy generational upgrade to Renoir. Now, will it be better than Tigerlake 4C in mobile? I don't know. Obviously, it'll have a core count advantage. It'll have a tweaked process. Tigerlake should rectify a lot of what was wrong with Ice Lake, meaning more Mhz and lower power draw. It'll also have a further improved iGPU. It'll bring with it Intel's noted advantages when it comes to extracting performance and low latency from their memory controllers as well. By that point, I would hope that AMD is specifying DDR4-3600 for their mobile platform, and maybe faster LPDDR4X for those types of systems, to help deal with the greater memory demands of a faster iGPU.

Either way, I still see the biggest loser here being Nvidia as they continue to loose volume on their MX lines of dGPUs. I suspect that, as Renoir has made everything from the MX330 on down completely irrelevant, and bumps up against the MX350 (which is essentially a low power 1050 mobile), we'll see Cezanne do the same for the MX350 as well, and start to push slightly into low power, cut down mobile 1650 territory. Look at the current laptop market. Whereas a few years ago, Mobile 1050 laptops were commanding prices at $900+, you can get 4GB 1650s in the under $700 category now.

As AMD and Intel improve their mobile iGPU products, and memory technology also moves along, there's less and less of a demand for the volume/Low end mobile dGPU products. Nvidia is going to have to do some more innovation in that area to keep their volumes up in that market.

Why would you need an equivalent number of RDNA CUs? it is clear that, on DDR4 at least, that 8CU RDNA will be way too much.
22CU RDNA = 36CU Polaris 20 (more or less)

6CU RDNA should be faster than 8CU Vega at the same clock.

With LPDDR5 and RDNA 2 things starts to change A LOT, but AMD will be killing his own mobile entry level dgpu market there.

 

[maddie]

Has anybody done size measurements of each CU of all the designs on the different nodes? Didn't find anything searching around.

I got around 3.2mm^2/CU for Renoir, all included (ROPS, ACEs, caches, etc). Was going to write a post on a comparison to VEGA 7 CUs, but stopped. Was doing this because I think that VEGA (Renoir) is modified and not the old Vega. VEGA 7 is 8x and Renoir 4800U has 8CU, 1.75 GHz and 15W. Roughly extrapolating out to 120W max for the 64CU portion of the GPU @ 1.75GHz

I believe major changes in Vega have taken place.

 

[JustMe21]

Why would you need an equivalent number of RDNA CUs? it is clear that, on DDR4 at least, that 8CU RDNA will be way too much.
22CU RDNA = 36CU Polaris 20 (more or less)

6CU RDNA should be faster than 8CU Vega at the same clock.

With LPDDR5 and RDNA 2 things starts to change A LOT, but AMD will be killing his own mobile entry level dgpu market there.

But, AMD doesn't have a lot of wins for their entry level dGPUs, so it make sense to beef up their iGPU. Since RDNA2 should be more power efficient, have variable rate shading, and better compression than Vega, why not just leave the 8 CUs if it won't take up extra die space. The current Vega 8 just isn't enough for entry level 1080p gaming, but an RDNA2 8 CU, even with DDR4 would probably do well enough.

 

[DrMrLordX]

But, AMD doesn't have a lot of wins for their entry level dGPUs, so it make sense to beef up their iGPU.

That doesn't make any real sense. AMD used to have the beefiest iGPU around (Kaveri, Carrizo) and it did them very little good. OEMs have repeatedly demonstrated that they will pair top-end iGPUs with dGPUs making them mostly irrelevant in actual products. iGPU-driven OEM machines make up too small of OEM's total sales to be terribly relevant to AMD's total revenue stream.

 

[LightningZ71]

Even if they have to stick to DDR4-3200, the RDNA CUs are more memory bandwidth efficient than VEGA CUs. Cutting down to 6 RDNA CUs can limit the iGPU in other ways. As we have already seen, there are already fewer ROPs in VEGA8 in Renoir. That seems to be a hinderance to total performance for VEGA 8 in some situations, even though it has high enough clocks to have more available FLOPS to throw at graphics processing.

It seems likely to me that, even though the "enhanced 7nm" process that Cezanne is going to be produced on has an improvement in density, the Zen3 cores, having more transistors than Zen2 cores, will take up a greater percentage of the die space, meaning that there will still be considerable pressure on the VEGA cores to live in a tiny space. Better thermals and higher clocks should enable another gradual evolutionary performance improvement for Cezanne, and I'm sure that the thin and light, iGPU only market will greatly appreciate that. It may not matter much for desktop, at least, not as much as the improved CCX design and greater single thread cache availability will help the CPU cores, but, it will help.

What follows, being produced on 5nm, and including RDNA cores with DDR5, should be something special. I suspect that iGPU performance will be notably past the RX560 and may even approach RX570 levels in CERTAIN, VERY SPECIFIC, situations. More generally, when used with image scaling and sharpening, 1080p gaming should be very comfortable for most titles with excellent frame rates. I wouldn't suggest it as a daily driver, but, its not unreasonable to use the same approach and drive a 1440 panel with one for those road warriors out there.

Again, the progress of Intel and AMD on iGPUs has been significant in the last few years. It's going to hurt the volume, bottom end products from both Nvidia and AMD as anything below the 1650ti and 5500M should be of very little actual benefit over the iGPUs being offered in the next generation mobile chips from both vendors.

As for beefy iGPUs being irrelevant, I dare say that the 2 in 1 market is showing that there is a use for those things. 4500u/4700u 2 in 1 tablets are doing quite well in gaming with reasonable expectations right now, with almost all of them not having a dGPU. I believe that, if they could actually find their way into the market, 4800u/LPDDR4X-4333 quad channel (4 x 32bit) based convertibles should be quite compelling for most everyone that's not playing AAA games competitively.

 

[xilli_fiberbit]

If i'am remember right cezanne will be produced in 7nm+, is-there a chance that vega will gains more CU ?

 

[Shivansps]

But, AMD doesn't have a lot of wins for their entry level dGPUs, so it make sense to beef up their iGPU. Since RDNA2 should be more power efficient, have variable rate shading, and better compression than Vega, why not just leave the 8 CUs if it won't take up extra die space. The current Vega 8 just isn't enough for entry level 1080p gaming, but an RDNA2 8 CU, even with DDR4 would probably do well enough.

AMD created an H version of Renoir and it is forcing OEMs to use them with their dGPUs...

And now that i think about it, what a strange way to segment things, the U, H and Gs are all diferent configurations except for the top ones. They have far too many diferent versions there.

 

[DrMrLordX]

AMD created an H version of Renoir and it is forcing OEMs to use them with their dGPUs...

Nothing's "forcing" OEMs to do so. OEMs choose to do so for their own reasons.

 

[Shivansps]

Nothing's "forcing" OEMs to do so. OEMs choose to do so for their own reasons.

I remember when mobile Renoir launched, that it was said than the H versions were only for use along a dGPU and that was the reason for the H versions to have lower CUs compared to the U versions of the same core count.

Ideally, you would want that the high mobile TDP skus to be the same configuration as the desktop ones, they are not.

 

[leoneazzurro]

There are a lot of laptops out there with Renoir H series and Nvidia GPU (practically the only ones with an AMD GPU came from MSI and Dell, all the other dGPU laptops are nVidia based). So this "forcing to use their dGPU" is incorrect. In the case, it is the vendors not using more powerful dGPU than the 2060 on AMD laptops...

 

[LightningZ71]

I understood it to be more of a "positioning to the realities of the market" thing. Most of the 45w designs are in larger laptops with better cooling systems, which also tends to be important for adding an effective dGPU. The H series, being the higher power variant, were always going to be used in those frames, so it makes sense that they be tailored to the market. If you look at the market, there are, what, two laptops for sale today that use an H class renoir APU WITHOUT a dGPU? I think both of those are actually manufactured by the same outfit. There's just perceived demand for laptops like that, as, if you're going to carry the bulk of a 45+ watt cooling system, and a battery big enough to make use of it, why wouldn't you throw in a few more dollars and get at least a serviceable dGPU?

Just look at the U market. How many 4800U devices are out there? From what we're hearing, Lenovo is no longer exclusive on it, yet, you can't hardly find any of them out there. That's the only U SKU with all 8 CUs enabled. Are there even any laptops out there at all that take maximum advantage of it by offering LPDDR4X 4x32 ram on the 4800u in TDP up mode with an adequate cooling system?

 

[leoneazzurro]

So far the only solution with a 4800H and no discrete GPU is the Schenker VIA Pro, which should be a Clevo o similar product.

 

[Zepp]

That video has low (and 1 medium setting) and 50% resolution scale at 1080p. The video Ryan showed was high settings at 1080p with 100% resolution scaling.

the 2500u is using DDR4 2400, what is tiger lake using?

 

[Hitman928]

the 2500u is using DDR4 2400, what is tiger lake using?

I don't know. Like I said, there's not enough info to really make a comparison against current market products and most likely there won't be until we get independent reviews.

 

[IntelUser2000]

Ideally, you would want that the high mobile TDP skus to be the same configuration as the desktop ones, they are not.

They do, other than the clock difference. The 4900HS has 8 CUs enabled.

So far the only solution with a 4800H and no discrete GPU is the Schenker VIA Pro, which should be a Clevo o similar product.

Clevo and Schenker sometimes uses configurations others barely use. They were the only one to use the first generation Iris Pro.

 

[coercitiv]

So far the only solution with a 4800H and no discrete GPU is the Schenker VIA Pro, which should be a Clevo o similar product.

Lenovo is bringing in the Lenovo Yoga Slim 7 Pro this month which is a 14" device sporting 4900H, 4800H or 4600H. No dGPU, it's ultrabook class.

 

[leoneazzurro]

Yup, just saw the press release. Prices seems also to be quite attractive for such a solution.

 

[Thibsie]

LPDDR4X or GDDR6, which one is it ?
That Lenovo spec sheet or wierd IMO.