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.

 

[ryanjagtap]

Increasing the L2 size and moving to a 32MB SLC with good partitioning rules (always reserve x% for iGPU etc.) Would do a lot for overall APU performance. That should be workable on N5.

I think they already doubled the L2 cache with respect to vega, they had it on one of the presentation slides. The graphics part of the APU is connected with the core complex via infinity fabric. They will have to redesign so that the latency from infinity fabric to L3 cache does not become a bottleneck. They have that modularity to since renoir so that they can easily change blocks of the APU as and when required like they did with cezzane (Just changed the CCX block of the APU)

 

[DisEnchantment]

The graphics part of the APU is connected with the core complex via infinity fabric. They will have to redesign so that the latency from infinity fabric to L3 cache does not become a bottleneck. They have that modularity to since renoir so that they can easily change blocks of the APU as and when required like they did with cezzane (Just changed the CCX block of the APU)

What you are thinking is only what marketing slides have led you to believe.
The Graphics part of the APU is connected to the Northbridge block which is then connected to the Infinity Fabric/DF via IOH (not directly to IF)
Similarly none of the USB/NVME/SATA etc. are connected to the Infinity fabric directly as shown in slides, they are connected to the Northbridge block connected to IOH, which in the case of Desktop is on the IOD and in the case of Mobile is on die.

Infinity Fabric is connected to L3 in AMD CCX design, the L3 is a cache coherent master and is responsible for snooping and maintaining coherency across CCXs.
So Graphics cores cannot access the L3 of the CPUs.

SLC is like the Infinity Cache in current RDNA2 GPUS( whatever is ejected/victimized from L2 goes to Infinity cache which is connected to UMC)
A much better question would be how could AMD design the Infinity cache to act like SLC. Will be interesting to see how AMD address this issue in the future, because currently the GPU goes to NB/IOH then to UMC, and the UMC handles requests from both CCXs and DF fabric masters.

Coherency between CPU cores and GPU CUs is possible using IF however with Trento + Aldebaran but they follow a different topology and address different use case.

This is how it looks like


On Linux this is the enumeration for the LilacKD-RMB SoC

-[0000:00]-+-00.0 Advanced Micro Devices, Inc. [AMD] Device 14b5
+-00.2 Advanced Micro Devices, Inc. [AMD] Device 14b6
+-04.0 Advanced Micro Devices, Inc. [AMD] Device 14b7
+-04.1-[33-61]--
+-08.0 Advanced Micro Devices, Inc. [AMD] Device 14b7
+-08.1-[62]--+-00.0 Advanced Micro Devices, Inc. [AMD/ATI] Device 1681 --> GPU
| +-00.1 Advanced Micro Devices, Inc. [AMD/ATI] Device 1640
| +-00.2 Advanced Micro Devices, Inc. [AMD] Device 1649
| +-00.3 Advanced Micro Devices, Inc. [AMD] Device 161d
| +-00.4 Advanced Micro Devices, Inc. [AMD] Device 161e
| +-00.5 Advanced Micro Devices, Inc. [AMD] Raven/Raven2/FireFlight/Renoir Audio Processor
| +-00.6 Advanced Micro Devices, Inc. [AMD] Family 17h (Models 10h-1fh) HD Audio Controller
| \-00.7 Advanced Micro Devices, Inc. [AMD] Raven/Raven2/Renoir Sensor Fusion Hub

 

[ahimsa42]

i wonder if the 680M tests were on a 6800U or H? i think price & availability are going to be the only wildcards for RMB laptops. if they are of reasonable cost & widly available i don't see how they they don't sell a ton of them.

 

[maddie]

Bandwidth and/or power ?

Power being the issue would be great for the desktop version.

 

[jpiniero]

IMO they don't want to make the IGP too good so they can sell the low end dGPUs.

 

[uzzi38]

There is only a 52% difference between the i7-11370H LPDDR4x-4266 versus 6800H 54W LPDDR5-6400. As I said AMD needs another big improvement, otherwise no chance.

They only "need a big improvement" because you're expecting Meteor Lake to literally double GPU performance, and relying on results from a single review as proof.

I could just as easily use the PurePC review as proof that actually Rembrandt is ~120% after taking out the Far Cry 6 benchmarks where Xe just fails to even finish the benchmark run.

That goes for everyone here mind you. Stop getting ahead of yourselves with just a single review. That can be the ComputerBase test, the one on Zhihu or PurePC. Wait for more to come in, then we can make assumptions on performance.

 

[IntelUser2000]

I think you're underselling 30%. If we saw that kind of jump in a CPU no one would call it anything less than huge.

Let me rewrite that for you.

30% for graphics is not huge.

Of course 30% for CPU is huge. That's absolutely ginormous if you are talking about single thread, and big even in multi-thread where you can tack on more cores.

But for graphics it's not. Context, context. On a mobile platform you'll see that kind of difference by different thermal and system settings. The Iris Xe G7 is 20-30% faster than Vega 8, but it needs the dual channel memory. And only when both are at 28W.

You can really see the impact that not having any infinity cache is having on the RDNA2 architecture. If they can manage a way to get 16MB of it on the next gen APU, it would be good for another nice bump in performance.

It seems to me the Chinese test is showing it doesn't really need fast RAM so the Infinity Cache benefits will be lower as well.

@mikk Don't forget that Intel is not changing graphics architecture in Alderlake nor Raptorlake. Intel NEEDs the gains in Meteorlake since they are using the same GPU for 3 generations.

It's nice they'll be doing another big jump with Arrowlake. Even if AMD does nothing with Phoenix Point, Remembrandt is enough to take on Raptorlake graphics, and more!

 

[LightningZ71]

If there wasn't a need for more memory bandwidth for 12CU/6WGP designs, I promise you AMD wouldn't have bothered with any at all in NAVI24, which has a better raio of CUs to memory bandwidth than the 680m. NAVI24 is an exercise in being the absolute least that they could get away with.

The 6CU version is already well fed with power and memory bandwidth, so neither seem to do much for it.

 

[IntelUser2000]

The dGPUs behave quite differently so I can't say much about that, but the test clearly shows only 8% gain from 33% increase in memory bandwidth. That's 24% for doubling, which is below even the typical 30% gain.

The dGPUs are expected to run faster and on higher settings as well so Infinity Cache is better but the gain might be quite muted on 680M. If they scale the CUs up yes I can see it.

 

[Panino Manino]

Would be nice if AMD had 2 dies.
4-6 cores with full 12 CU and 8-12 cores with 6CU.
Those higher core counts APUs are meant to be paired with discrete GPUs so it's a waste to give so many CUs to these APUs.

But Intel also does this, right?
It goes against logic, people who want to pay less for weaker CPUs are the ones that most need that extra iGPU.
Will this ever happen one day?

 

[jpiniero]

Would be nice if AMD had 2 dies.
4-6 cores with full 12 CU and 8-12 cores with 6CU.
Those higher core counts APUs are meant to be paired with discrete GPUs so it's a waste to give so many CUs to these APUs.

But Intel also does this, right?
It goes against logic, people who want to pay less for weaker CPUs are the ones that most need that extra iGPU.
Will this ever happen one day?

The IGP that's in Raphael's IO die is the closest thing you will get. That might be less than 6 CUs.

 

[ahimsa42]

lol-so much for RMB affordability & availability i guess

Next-Gen AMD, NVIDIA & Intel Powered Laptops Threatened With Further Delay As Supply Chains & Prices Under Pressure, SSDs Become More Expensive Too

It looks like latest laptops utilizing AMD, Intel & NVIDIA chips will be hard to purchase as supply & pricing threatens the whole market.

wccftech.com

 

[leoneazzurro]

Well' that is a big no-no for every laptop availability, not only RMB based ones.

 

[Shivansps]

I think the problem here is power, simple as that. This is really good news for the desktop version, but i guess well see.
That alone already tell you why there is no infinity cache.

 

[moinmoin]

Well' that is a big no-no for every laptop availability, not only RMB based ones.

For anything that goes through China potentially, whether it's manufactured or just assembled there. Considering China still tries to stick to their zero covid policy that could spell huge troubles considering the high virality of omicron.

 

[uzzi38]

Another Rembrandt review, this time they have actually useful numbers for how much idle power consumption has improved:

ASUS ROG Zephyrus G14 review (GA402RK - Ryzen 9, RX 6800S)

This is our detailed review of the 2022 Asus ROG Zephyrus G14, in the top-tier GA402RK configuration built on an AMD Ryzen 9 6900HS processor and an AMD

www.ultrabookreview.com

(They also have the 6900HS beating their 12900H sample in MT performance at 50W, whilst losing at 70W).

 

[andermans]

They only "need a big improvement" because you're expecting Meteor Lake to literally double GPU performance, and relying on results from a single review as proof.

I could just as easily use the PurePC review as proof that actually Rembrandt is ~120% after taking out the Far Cry 6 benchmarks where Xe just fails to even finish the benchmark run.

That goes for everyone here mind you. Stop getting ahead of yourselves with just a single review. That can be the ComputerBase test, the one on Zhihu or PurePC. Wait for more to come in, then we can make assumptions on performance.

I think the big question for how Meteor Lake stacks up against AMD is what they're going to do with the clock speeds. There is currently a big difference between clock speeds for Intels iGPUs and the rumored clock speeds for their dGPUs. if they can make that 50%-100% increase in clock speeds fit in their iGPU (e.g. with power budget etc.) then that does provide an extra jump.

Of course that likely is against Phoenix, and I'm curious what that does for GPU size. If it ends up with RDNA3 (which would make sense with 5nm) I wonder if it will have the same increase of 2x shaders per WGP that desktop RDNA3 is rumored to have.

 

[ahimsa42]

let the games begin! hopefully this will help force RMB to compete with these price & availability wise:

Intel 12th Gen Alder Lake-P & Alder Lake-U Laptop CPUs Launched: Over 300 New Enthusiast and Thin & Light Designs For Everyone

Intel has launched its 12th Gen Alder Lake-P & Alder Lake-U Laptop CPUs for all kinds of users ranging from enthusiasts to thin and light.

wccftech.com

 

[Saylick]

let the games begin! hopefully this will help force RMB to compete with these price & availability wise:

Intel 12th Gen Alder Lake-P & Alder Lake-U Laptop CPUs Launched: Over 300 New Enthusiast and Thin & Light Designs For Everyone

Intel has launched its 12th Gen Alder Lake-P & Alder Lake-U Laptop CPUs for all kinds of users ranging from enthusiasts to thin and light.

wccftech.com

Not sure what that post has to do with Rembrandt, seeing as it's entirely related to Alderlake mobile. Should probably post that over in the Alderlake thread instead... Just don't want to cause a derail here.

 

[Frenetic Pony]

Would be nice if AMD had 2 dies.
4-6 cores with full 12 CU and 8-12 cores with 6CU.
Those higher core counts APUs are meant to be paired with discrete GPUs so it's a waste to give so many CUs to these APUs.

But Intel also does this, right?
It goes against logic, people who want to pay less for weaker CPUs are the ones that most need that extra iGPU.
Will this ever happen one day?

Chiplets should help on layout efficiency for laptop SOCs, next year for Intel and possibly AMD as well.

 

[uzzi38]

Although I wish they tested more games: kind of comprehensive test from ComputerBased: AMD Radeon 680M: The Ryzen 6000 iGPU in the test: Benchmarks of the Radeon 680M in games - ComputerBase

Simple TL;DR for results is:

 

[leoneazzurro]

Well it seems it easily outperforms a RX560, even if it's not at GTX1650 level. But it's right where I'd expected it to be.

 

[DisEnchantment]

AMD will be selling a ton of these as "Pro" ( something like those AMD Ryzen™ 7 PRO 5750G ) variants for the enterprise. Lots of People may not like it, but Pluton+PSB is going to do well in enterprise.
OEMs will capitalize quickly on that efficiency as well and sell corporate boxes with the utmost minimalistic thermal solution possible.
Seems they had all the features discussed earlier below, except for the AIM-T which will be coming for Pro only.

Apparently this is not all
Rembrandt also has a new Xtensa LX6 Chip doing remote management duties in a framework called AIM-T (AMD Integrated Management technology), similar to Intel remote management engine. I guess this will go to all the Pro SKUs. I guess they just license in bulk a bunch of Cadence Tensilica IPs
New Z state deep sleep likely handled by the AOP (Valve also mentioned a feature like this on Steam Deck, instant wake up from days of suspend state)
Smartshift 2.0

Another thing to note also, lots of Rembrandt ACP firmware and support added to SOF framework.
Nice work on this, at least OEMs can work with Tensilica SDK directly to customize the Audio DSP FW with out of box support in SOF, QCM Hexagon SDK may still be superior but SOF is much more desired due to more industry support and being open source. (MTK is also using SOF along with Intel)

3 weeks to go, but zero leaks so far. AMD is tightening the screws looks like.

 

[Heartbreaker]

Well it seems it easily outperforms a RX560, even if it's not at GTX1650 level. But it's right where I'd expected it to be.

It would be enough to hold me over for a while until GPU cards return to sanity, even if it takes years because I won't pay scalper prices.

Buy you won't be able to buy one for DIY for a long time.

 

[Panino Manino]

Although I wish they tested more games: kind of comprehensive test from ComputerBased: AMD Radeon 680M: The Ryzen 6000 iGPU in the test: Benchmarks of the Radeon 680M in games - ComputerBase

Simple TL;DR for results is:
View attachment 57865

It's faster at RT than the 6500XT?