Question AMD's next-gen Ryzen 6000 APU: Zen 3+ CPU, RDNA 2 GPU expected in 2022 - TweakTown

[JPB]

AMD's next-gen Ryzen 6000 APU: Zen 3+ CPU, RDNA 2 GPU expected in 2022

AMD's next-gen 'Rembrandt' APU packs DDR5, USB 4.0, PCIe 4.0, AM5 socket, Zen 3+ CPU cores, RDNA 2 GPU core

AMD has a gigantic pipeline of products coming out over the coming years, but there is a metric boat load of new information that has just leaked through Expreview.

The new Ryzen 6000 series APUs are known as "Rembrandt" and is a massive upgrade over current APUs, where it will pack the new Zen 3+ CPU architecture, and upcoming RDNA 2 GPU architecture. We're expecting AMD to make the new Rembrandt APU on TSMC's new 6nm node (which is an optimized version of TSMC's N7).

AMD's next-gen Zen 3+ architecture will have a bunch of improvements over the current Zen 2 chips that are here, but it is the GPU side that is interesting. AMD will replace Vega GPU cores with RDNA 2 GPU cores, which will surely provide a gigantic leap in performance -- especially considering this is an APU.

This means we should expect the Rembrandt APU to have ray tracing technology, meaning we could have near next-gen console levels of performance which is pretty damn amazing. By then (2022) the Zen 3+ CPU architecture and RDNA 2 GPU architecture could leap frog what is built into the consoles.

Not just that, but AMD is expected to have Rembrandt APUs support the AM5 socket (this is new) which will bring DDR5 support, LPDDR5 support, USB 4.0 support, and of course PCIe 5.0.

AMD will launch its new Rembrandt APUs in 2022, and I'm sure we'll have more details between now and then.

 

[AtenRa]

Vega may be smaller than in RDNA1, but RDNA1 has a much better IPC(35-40%), so It doesn't necessarily mean Vega has better performance/size ratio.

Vega 64 is 4% faster at FullHD and 9% faster at 4K than 5600xt.
Vega 64 -> 64CU, ~1530mhz, 484GB/s
5600XT -> 36CU, ~1700mhz, 288GB/s
If you scaled them both down to 8CU to be comparable to IGP in APU while keeping their average clockspeed It would look like this:
Vega 64 -> 8CU, ~1530mhz, 61GB/s
5600XT -> 8CU, ~1700mhz, 64GB/s
Vega would end up with lower bandwidth requirement, but RDNA1 would be significantly faster thanks to much better IPC and higher clocks so It would have much better performance/bandwidth ratio than Vega.

Naturally RDNA1 is not the same and RDNA2 will most likely end up bigger on same process, but performance or bandwidth requirement should be better rathen than worse compared to RDNA1.
So my conclusion is that It doesn't make any sense using Vega architecture for the next gen of APUs.

RDNA1 CU is double that of VEGA, you cannot fit the same amount of VEGA and RDNA1 CUs in the same size.
Also, VEGA 20 is alot faster than VEGA 10, so you should be comparing VEGA 7 to RX5600XT or VEGA 20 to RDNA1.

 

[TESKATLIPOKA]

RDNA1 CU is double that of VEGA, you cannot fit the same amount of VEGA and RDNA1 CUs in the same size.

Do you have a link to actual Vega CU size?
GPU(IGP) doesn't consist only from CUs and I never said you can fit the same amount of RDNA1 CU and Vega CU in the same space.

Also, VEGA 20 is alot faster than VEGA 10, so you should be comparing VEGA 7 to RX5600XT or VEGA 20 to RDNA1.

This just doesn't make sense to me.
Why should I compare 7CU IGP(Vega 7) to RX5600XT?
Why should I compare Vega 20 to RDNA1?
RX5600XT is also RDNA1 and Vega 20 has 1TB/s which is a total overkill so If I compared It to anything based on RDNA1, It would end up much worse for Vega. I was comparing Vega 64 vs RX5600XT because they have similar performance.

 

[Shivansps]

The first RDNA2 APU will likely end up having about 5CU, 6CU max.

 

[TESKATLIPOKA]

The first RDNA2 APU will likely end up having about 5CU, 6CU max.

Why do you think so? With that amount It won't be any faster than Renoir IGP, unless clockspeed would be higher.

 

[Shivansps]

Why do you think so? With that amount It won't be any faster than Renoir IGP, unless clockspeed would be higher.

Im thinking about the final size, best case scenario is having a IGP as big as Vega 11 in Picasso. And thats around 6CU, i belive.

 

[TESKATLIPOKA]

I don't understand why the best case scenario for RDNA2 IGP would be only the size of Vega 11 IGP. Why couldn't It be bigger If there is enough bandwidth to feed It?
RDNA1 CU is only ~2mm2 on 7nm process and If RDNA2 CU is 25% bigger then that's 2.5mm2. That would mean you need only 10mm2 for 4 more CUs. I think that's worth It If you have enough bandwidth to keep It fed.

 

[LightningZ71]

I would like to think that the first RDNA APU would still have about 8 CUs. If my memory serves, for RDNA, they have to be in groups (pairs, quads?). For segmentation purposes, that's about the floor to what they could use. Giving a spread of 8/6/4/2 CUs for the different tiers. With RDNA's greater throughput per CU per clock, that shouldn't be too much of a problem. If we also expect N5P to scale clocks better, its also not unreasonable to have the CUs hitting 2.4-2.6Ghz clocks. Still more than enough throughput to be heavily memory bandiwdth constrained by the dual channel DDR5 its expected to have.

 

[amrnuke]

RDNA1 CU is double that of VEGA, you cannot fit the same amount of VEGA and RDNA1 CUs in the same size.

Is that because an RDNA CU is less dense (to promote speed) or because it has more transistors?

 

[TESKATLIPOKA]

RDNA CU has more transistors than Vega CU, the performance has to come from somewhere.

 

[TESKATLIPOKA]

8CU 1750mhz IGP+IGP in Renoir has only 3200Mhz DDR4 most of the time and that's just 51.2GB/s. From my above calculation the 8CU Vega IGP should need 61GB/s and that's at 1530mhz, so at 1750mhz It should need 69.8GB/s. On the other hand 8CU RDNA1 at 1750mhz should need 65.9GB/s.
My conclusion is that you can have 8CU 1750mhz RDNA1 IGP even with 3200mhz DDR4 and It wouldn't be any more bandwidth limited than Renoir IGP, yet the performance provided would be better. Another question is how much power does RDNA consume compared to Vega, If the clockspeed and number of CUs and ROPs is the same.

 

[Shivansps]

8CU 1750mhz IGP+IGP in Renoir has only 3200Mhz DDR4 most of the time and that's just 51.2GB/s. From my above calculation the 8CU Vega IGP should need 61GB/s and that's at 1530mhz, so at 1750mhz It should need 69.8GB/s. On the other hand 8CU RDNA1 at 1750mhz should need 65.9GB/s.
My conclusion is that you can have 8CU 1750mhz RDNA1 IGP even with 3200mhz DDR4 and It wouldn't be any more bandwidth limited than Renoir IGP, yet the performance provided would be better. Another question is how much power does RDNA consume compared to Vega, If the clockspeed and number of CUs and ROPs is the same.

Wait. Vega 11 at 1600mhz performs faster than Vega 8 at 1600mhz in Picasso with DDR4-3200.

 

[TESKATLIPOKA]

How much faster? Is It at least 25% faster? Throughput is 37.5% better.
At worst It would mean my calculation is not accurate, but then It should also apply to RDNA and not just Vega.

 

[Shivansps]

How much faster? Is It at least 25% faster? Throughput is 37.5% better.
At worst It would mean my calculation is not accurate, but then It should also apply to RDNA and not just Vega.

About 10% you are memory limited on DDR4-3200 whiout a doubt. It may also be due to be 4/8.
With DDR4-3466 the diference is larger, but by that point i was unable to test because my friend was able to hit 1700mhz and DDR4-3466 on his 3400G and i could not test beyond 1650mhz and DDR4-3333 on my 3200G.

The good thing about Renoir is not only able to hit beyond 2Ghz, but DDR4-4266 is also possible at maybe 1:1 fabric, im expecting Renoir, as a full 7mm monolithic CPU to hit higher fabric speeds than Matisse.

 

[LightningZ71]

According to what I've read around the internet, Renoir memory bus is uncoupled from the inrnal fabric and seems to run at varying ratios depending on the memory speed. Granted, this was in laptops, so we don't know what's available via desktop UEFI configs.