Question AMD Rembrandt/Zen 3+ APU Speculation and Discussion

[izaic3]

Alright, so we've had some leaks so far. I don't know if any of it's been confirmed yet, as it's pretty early, but here is what I've surmised so far (massive grain of salt of course):

• Zen 3 "enhanced"
• DDR5 and LPDDR5 support
• PCIE4
• 6nm
• CVML
• Navi 2/RDNA/RDNA 2
• 12 CU
• Socket FP7/AM5
• Possible USB4
• 2021(unlikely) -2022 (more likely) release date

If if turns out to have RDNA 2 and 12 CU, I could see iGPU performance potentially almost doubling over Cezanne.

If I've made any mistakes or gotten anything wrong, please let me know. I'd also love to hear more knowledgeable people weigh in on their expectations.

 

[andermans]

The other question is how likely it will be that Van Gogh will support the full 6.4 GT/s LPDDR5. Like AFAIU Genoa will support DDR5-5200, wouldn't something similar be likely for VanGogh?

 

[uzzi38]

I don't think the standard 9W Van Gogh will have that much lower average clockspeed for IGP than 15W Cezanne.
I consider 128bit 3200Mhz as good enough for 15W Cezanne.
6.4GHz lDDR5(LPPDR5) with 64bit memory controller will be at best comparable to 3.2GHz DDR4 with 128bit memory controller, but most likely worse thanks to higher latencies.
If you say 128bit LPPDR5 is an overkill for 8CU RDNA2 I agree, but for example 64bit paired with anything other than 6.4GHz LPDDR5 doesn't look so great to me, but maybe I am too optimistic about Van Gogh's clocks.

Cezanne requires about 15W (not including uncore) to sustain max clocks on it's iGPU. Add on uncore and CPU core power and you're looking at more like 25W.

Van Gogh has the benefit of being a low power part sure, but even within that 9W power budget, you can count on 4-5W being allocated to everything but the GPU on a GPU load. There really isn't much headroom. 4-5W for the GPU as well across 8 CUs is absolutely nothing. We're legitimately talking about <1GHz per CU.

 

[uzzi38]

The other question is how likely it will be that Van Gogh will support the full 6.4 GT/s LPDDR5. Like AFAIU Genoa will support DDR5-5200, wouldn't something similar be likely for VanGogh?

I doubt it. Expecting LPDDR5-5500 at best.

 

[LightningZ71]

It sounds like, to me, that Van Gogh is taking the strategy of using a wider iGPU running at lower clocks to give good performance at low power. Kind of like how Picasso was used in the MS Surfaces, with the extra CUs enabled, but running at lower clocks.

Do we think that there will be a TDP-UP version of Van Gogh that can handle 30 watts, or is Van Gogh really just aimed at Ultra low power tablets, 2 in 1 convertibles, and hand held gaming consoles? At those low power levels, a single LPDDR5 64bit channel seems like it would meet a lot of design objectives.

 

[scineram]

Not sure we can answer much without knowing who the original customer was.

 

[Shivansps]

8CUs sounds like a lot for a <10W ULV chip with 64bit lpddr5.

 

[TESKATLIPOKA]

8CUs sounds like a lot for a <10W ULV chip with 64bit lpddr5.

If the clock is low then 8CU is not a problem. Lakefield also has a relatively big 64EU IGP, but the clocks are very very low. This has the highest effectivity.

 

[TESKATLIPOKA]

Cezanne requires about 15W (not including uncore) to sustain max clocks on it's iGPU. Add on uncore and CPU core power and you're looking at more like 25W.

Van Gogh has the benefit of being a low power part sure, but even within that 9W power budget, you can count on 4-5W being allocated to everything but the GPU on a GPU load. There really isn't much headroom. 4-5W for the GPU as well across 8 CUs is absolutely nothing. We're legitimately talking about <1GHz per CU.

After looking at your and izaic3's charts I have to agree, not sure about IGP boost, but average clocks for IGP should be <1GHz. RDNA2 is a lot more efficient than Vega, but If within 9W TDP the IGP has only 3-4W allocated, then Vega IGP with 8-9W will have much more headroom for higher clocks.

 

[izaic3]

Has anyone heard any rumors or have any speculation as to whether or not Rembrandt will finally include Thunderbolt? USB 4 was rumored, though it's been a while since I've heard anything new on that.

 

[scineram]

So when will Rembrandt come to desktop, if ever? Must be AM5, since DDR5 only. But what branding? Will they shove it into Ryzen 6000 for low end Zen 3 stuff, like Lucienne?

 

[uzzi38]

Btw I've been posting the charts here too: https://forums.anandtech.com/thread...na-architectures-thread.2579999/post-40487064

Also, if you want I can DM the spreadsheet. Just didn't want to post it in case somebody goes and messes with the data.

EDIT: Oh wait, I figured out how to do a read-only link. I think. Here: https://docs.google.com/spreadsheets/d/1R57Xfp0pES5AFwbT-zAf8FXfE_iIRmkY7iKN0-Azbdc/edit?usp=sharing

I'm in the process of updating the spreadsheet, but I ran some tests with some estimations for RMB performance using my 6700XT, trying to match compute throughput as best I could. It's in the second sheet.

Should provide a good explanation as to why RMB is so interesting on the iGPU. Only things I'm really missing is Renoir and I was hoping to get specific TGL+LPDDR4X numbers, but haven't been able to find a review that detailed TDP at 28W, used LPDDR4X AND also noted just the graphics score alone, so for now I've used extremely high performing scores for 96EUs.

But yeah, I think full desktop 1650 performance in a 35W package is totally feasible.

 

[uzzi38]

Btw @mikk you may also want to see the above and perhaps realise why I was saying that you shouldn't expect LPDDR5 for Xe to catch up in performance. I've provided both 1.7GHz and 2GHz figures because they're essentially an upper and lower bound for what should be possible within a 28W TDP with everything considered.

 

[izaic3]

I'm in the process of updating the spreadsheet, but I ran some tests with some estimations for RMB performance using my 6700XT, trying to match compute throughput as best I could. It's in the second sheet.

Should provide a good explanation as to why RMB is so interesting on the iGPU. Only things I'm really missing is Renoir and I was hoping to get specific TGL+LPDDR4X numbers, but haven't been able to find a review that detailed TDP at 28W, used LPDDR4X AND also noted just the graphics score alone, so for now I've used extremely high performing scores for 96EUs.

But yeah, I think full desktop 1650 performance in a 35W package is totally feasible.

That second page is very helpful, thank you!

 

[tomatosummit]

Some info from executablefix
Summary is rembrant is 6nm, 12cu rdna2, am5, ddr5 and zen3+
If it's following the usual yearly cadence I'd say this could put the am5 launch in 22q1 for oems at least and I'm suspecting more and more that zen3+ is just whatever is fabbed on n6 including whatever warhol was supposed to be at some point in time.

https://twitter.com/x/status/1388249804787048456

https://twitter.com/x/status/1390939182210732032

https://twitter.com/ExecuFix/status/1390963672277819395?ref_src=twsrc^tfw

 

[Asterox]

Some info from executablefix
Summary is rembrant is 6nm, 12cu rdna2, am5, ddr5 and zen3+
If it's following the usual yearly cadence I'd say this could put the am5 launch in 22q1 for oems at least and I'm suspecting more and more that zen3+ is just whatever is fabbed on n6 including whatever warhol was supposed to be at some point in time.

https://twitter.com/x/status/1388249804787048456

https://twitter.com/x/status/1390939182210732032

https://twitter.com/ExecuFix/status/1390963672277819395?ref_src=twsrc^tfw

Very good, i was expecting same Renoir or Cezanne APU scenario or only up 8 CU.

Future Desktop AM5 APU version, hm it is logical again up to 12CU RDNA2.

 

[uzzi38]

One more:

https://twitter.com/x/status/1390996826506080257

LPDDR5-5500 support as well.

 

[movdx]

Isn't 550 a little? I would like something more like LPDDR5-6400.

 

[IntelUser2000]

But yeah, I think full desktop 1650 performance in a 35W package is totally feasible.

So you think Remembrandt will be 3-3.5x faster than desktop Cezanne for iGPU?

 

[uzzi38]

So you think Remembrandt will be 3-3.5x faster than desktop Cezanne for iGPU?

Ah no, that's a mistake of mine, mixed up my numbers. 2.5x desktop Cezanne on the iGPU.

Wait, I'm going off the top of my head here, but desktop Cezanne is like 1.2k pts in TS, right? If so, them yeah, 2.5x.

 

[tomatosummit]

Ah no, that's a mistake of mine, mixed up my numbers. 2.5x desktop Cezanne on the iGPU.

Wait, I'm going off the top of my head here, but desktop Cezanne is like 1.2k pts in TS, right? If so, them yeah, 2.5x.

Are you comparing desktop to desktop or laptops?
At the desktop power draws I'm also expecting around 2x performance for the most part,
Mobile parts are always a mess with power limits, so I think a mobile rembrant could be more than 2-2.4x performance of a mobile cezanne with the additional process and rdna efficiency imrovements and lpddr5 if nothing is hamstrung by useless laptop design.

 

[Shivansps]

a 2400mhz Vega 8 with DDR4-4400 is petty much right below a base GTX1050, a Vega 11 would outperform it already at that speed. If RMB with DDR5 AND 12CU of RDNA2 dont outperform the GTX1650 non Super it would be very dissapointing.

 

[tomatosummit]

a 2400mhz Vega 8 with DDR4-4400 is petty much right below a base GTX1050, a Vega 11 would outperform it already at that speed. If RMB with DDR5 AND 12CU of RDNA2 dont outperform the GTX1650 non Super it would be very dissapointing.

IF any of these apus are available ever, then I think it'll also depend on what ddr5 is available. It'll take a while to go appreciably above 6000mt by my guess for a couple of years.

 

[eek2121]

With rumors of Warhol being cancelled, I wonder if AMD may have decided to use Rembrandt up and down the stack and skip chiplets for Ryzen 6000.

 

[TESKATLIPOKA]

a 2400mhz Vega 8 with DDR4-4400 is petty much right below a base GTX1050, a Vega 11 would outperform it already at that speed. If RMB with DDR5 AND 12CU of RDNA2 dont outperform the GTX1650 non Super it would be very dissapointing.

We are still talking about an IGP here, and you would be disappointed If It wasn't faster than a GTX 1650, which has GDDR6 and no CPU to share memory and TDP with? Why should It outperform GTX 1650 and by how much in your opinion?
It's enough If It's performing similarly or close enough, and It will depend mainly on available bandwidth and clockspeed.
In desktop I can imagine that It could perform very well against GTX 1650, thanks to high clocks(~2.3-2.5GHz) and TDP(95 or 125W), while GTX 1650 has fewer Cuda than the mobile version(896 vs 1024).
In mobile I expect It will be slower.

 

[TESKATLIPOKA]

With rumors of Warhol being cancelled, I wonder if AMD may have decided to use Rembrandt up and down the stack and skip chiplets for Ryzen 6000.

If they want to release more cores than 8, then they can't skip chiplets, unless they will keep the current 12-16 core lineup and just rename It until Zen4.
In mobile I think we won't see more than 8 cores.