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.

 

[BorisTheBlade82]

from what i have read there is not much difference if any between a 5800U & a 5800H when it comes to gaming

You mean without a dGPU? Yes, that is correct (ofc excluding edge cases).

 

[ahimsa42]

You mean without a dGPU? Yes, that is correct (ofc excluding edge cases).

correct-igpu only

 

[Rannar]

from what i have read there is not much difference if any between a 5800U & a 5800H when it comes to gaming

Renoir/ Cezanne vega is mostly mem bandwith and ROP limited, not power limited. it's iGPU stops to powerthrottle between TDP ranges of 25w to 35w. Rembrandt RDNA2 will have double the ROPs and much more bandwith comparatevly. can make more use of higher TDP ranges. i have 4800H laptop without discrete GPU. While its very nice and manages to almost run everything it feel just one two punch too little. cant wait for 6800H.

 

[ahimsa42]

Renoir/ Cezanne vega is mostly mem bandwith and ROP limited, not power limited. it's iGPU stops to powerthrottle between TDP ranges of 25w to 35w. Rembrandt RDNA2 will have double the ROPs and much more bandwith comparatevly. can make more use of higher TDP ranges. i have 4800H laptop without discrete GPU. While its very nice and manages to almost run everything it feel just one two punch too little. cant wait for 6800H.

that makes sense but i guess we'll have to wait & see. it would be great if that ends up being the case as i would certainly go for the H instead of U if there is an appreciable increase in igpu performance-say around 15-20%.

 

[Spicy]

makes sense since Rembrandt is a mobile CPU so will be for laptops, all in one desktops & mini PC's.

Like 2000G, 3000G and 5000G. But there WERE consummer deskop versions of them.
If it is so, it's only for rationing/shortage reason, I think, and OEM priority, like 4000G.

(or Raphaël will have a BIG iGP)

 

[Joe NYC]

So Rembrandt will have AM5 socket for desktop.

Does that mean the mobos will need a new chipset or will they use the existing chipsets for AM4?

 

[majord]

Let's not forget: The desktop 3400g had 11 CUs and could routinely clock north of 1600mhz. 12 CUs of RDNA 2, while being more efficient and likely clocked above 2Ghz in some parts, isn't a massive uplift. Assuming that the DDR5 can keep it fed, I'd still be shocked if it consistently reached RX560/1050 levels of in game performance.

This doesn't make much sense.. Desktop Cezzane is already not far off RX560/1050 levels when paired with fast [~4000] DDR4. - 80-90% of it @ 1080p

12CU RDNA2 + newer gen DCC + DDR5, should absolutely walk all over Cezzane , so go figure the rest I guess.

 

[ahimsa42]

This doesn't make much sense.. Desktop Cezzane is already not far off RX560/1050 levels when paired with fast [~4000] DDR4. - 80-90% of it @ 1080p

12CU RDNA2 + newer gen DCC + DDR5, should absolutely walk all over Cezzane , so go figure the rest I guess.

agreed. i think that Rembrandt may end up having far better igpu improvements than most people expect. it will hopefully end up being the 60fps 1080p apu (on most games anyway) that everyone has been waiting for.

 

[jpiniero]

agreed. i think that Rembrandt may end up having far better igpu improvements than most people expect. it will hopefully end up being the 60fps 1080p apu (on most games anyway) that everyone has been waiting for.

And then it will be OOS because people will mine it.

 

[Panino Manino]

I don't understand the skepticism.
Did people forgot when Navi came out? Even this last week "bad" 6600 GPU beats the Vega 64 with easy at lower resolutions. I expect performance to at least double what we have now.

 

[LightningZ71]

The "bad" 6600 still has at least some Infinity Cache to help hide it's VRAM throughput deficit. Rembrandt has none and dramatically lower VRAM bandwidth. It's not going to be amazing.
(Edited to fix "throughput")

 

[Panino Manino]

The "bad" 6600 still has at least some Infinity Cache to help hide it's VRAM throughout deficit. Rembrandt has none and dramatically lower VRAM bandwidth. It's not going to be amazing.

I bet it will.
There's less bandwidth but also less CUs to feed.

 

[jpiniero]

I bet it will.
There's less bandwidth but also less CUs to feed.

You are talking about 50% more CUs, higher core frequency and possibly double memory bandwidth (if you buy pricey DDR5). Rembrandt should be a lot faster.

It's still going to be a lot slower than a 6600. Maybe more like 1650 non-Super with 6200 DDR5.

 

[Shivansps]

The "bad" 6600 still has at least some Infinity Cache to help hide it's VRAM throughput deficit. Rembrandt has none and dramatically lower VRAM bandwidth. It's not going to be amazing.
(Edited to fix "throughput")

Remember the rumor of the Navi 24 with 16CU, 112GB/s and 16MB IC. If these specs are real, 12CU may be perfectly fine with DDR5 on RMB.

The problem is that the 12CU one will be used for the 8C SKU only.

 

[ahimsa42]

a mini pc model HM90 was just announced with a 4900H from Minisforum and has16GB ram & SSD for under $700. something like this with a 6800H & DDR5 would be amazing and do everything most people need in a very compact SFF.

 

[Panino Manino]

You are talking about 50% more CUs, higher core frequency and possibly double memory bandwidth (if you buy pricey DDR5). Rembrandt should be a lot faster.

It's still going to be a lot slower than a 6600. Maybe more like 1650 non-Super with 6200 DDR5.

That's what you understood? I didn't meant to say that it would be as fast as a midrange discrete GPU!

 

[Rannar]

It's still going to be a lot slower than a 6600. Maybe more like 1650 non-Super with 6200 DDR5.

Rumors tell max 5200 DDR5/ 5500 LPDDR5 support - at least as of now. it's their first IMC for DDR5 so will probably not support highest data rate memory from the get go (similar to how DDR4 IMC progressed over the generations from 2133 to 3200 support). I would expect official 6400 and up data rates supported with generation after Rembrandt. of course on desktops you can always try to overclock...

 

[Spicy]

I believe that APUs have good memory overclocking potential, thanks to their monolithic architectures. Right?

 

[LightningZ71]

No, that's better latency numbers due to not having a ser/des link between the I/O die and the CCD.

 

[LightningZ71]

You're looking at about 80GB/sec in bandwidth with worse latency numbers than the Vega 8 APUs and no infinity cache, and I'm seeing people expecting better performance than dGPUs with over 100GB/sec of bandwidth.

That is overly optimistic.

 

[jpiniero]

You're looking at about 80GB/sec in bandwidth with worse latency numbers than the Vega 8 APUs and no infinity cache, and I'm seeing people expecting better performance than dGPUs with over 100GB/sec of bandwidth.

That is overly optimistic.

At least 6400 should be available by the time Rembrandt ships, and presumably should be able to handle even though it's overclocked. Which would be 100.

 

[LightningZ71]

Ddr5-6400 is 90GB/sec, and this is an APU, so that's shared with the CPU cores. In addition, this will be in laptops for most of its volume, and laptops that allow DRAM overclocking are VERY rare, and almost never just using an iGPU for their 3d work.

On a desktop, overclocked to within an inch of it's existence, you might have something...

 

[majord]

I maintain around 1050ti levels for a desktop 12CU part, 1050 levels for mobile.. nominal. Depends on clocks and RAM speeds. The other unknown is ROPS. Will they move to 32 or just remain at 16 . In fact I'd go so far as to say this would be THE primary factor , not RAM bandwidth.

Citing a lack infinity cache I don't really buy.. it's not like it makes RDNA2 merely 'competitive' for a given bandwidth - it's incredible the level of performance a 6600/XT gets from just 224-256GB/s bandwidth. with just 32MB of 836GB/s ICache. ~100GB/s of theoretical B/W from DDR5 in Dual channel seems plentiiful for that level of performance to me. Shared or otherwise

 

[LightningZ71]

What's incredible about it? With a cache hit rate of 50%, the effective bandwidth of that arrangement is over 500GB/sec. With 500GB/sec of bandwidth, you had better have impressive performance! And, at less than 1/5th of that bandwidth, you're not getting too far too fast.

 

[TESKATLIPOKA]

You're looking at about 80GB/sec in bandwidth with worse latency numbers than the Vega 8 APUs and no infinity cache, and I'm seeing people expecting better performance than dGPUs with over 100GB/sec of bandwidth.

That is overly optimistic.

Really? Let's see.

Link

Ryzen 7 5700G specs: 8C16T 3.8/4.6GHz 65W TDP ; IGP: 2000Mhz 8CU(512SP:32TMU:8ROPS?) (2048 GFLOPs)
Memory: 2x 8GB DDR4-3200 (dual channel) memory @ 16-16-16-36 -> 51.2 GB/s for both CPU+IGP.

RX 560 specs: 16CU(1024SP:64TMU:16ROPs); Frequency: 1168(1244) MHz -> 2611GFLOPs; BW: 112GB/s; TDP: 80W

RX 560 has 119% higher BW just for Itself, TDP is also higher but is for the whole card, GFLOPs is 27.5% higher, Pixel Fillrate(ROPs) is most likely ~2.5x higher, and you gain only 33-35% better performance. It doesn't look like the limited bandwidth is negatively affecting the performance to me, maybe by just a few %.

dGPUs vs 5700g IGP:
RX 560 is 35% faster and 33% faster at 99th percentile.
GTX 1050 is 61% faster and 48% faster at 99th percentile.
GTX 1050 Ti is 79% faster and 68% faster at 99th percentile.
GTX 1650 is 128% faster and 115% faster at 99th percentile.

Now what about Rembrandt?
Example: 8C16T Zen3+ 65W; 12CU(768SP, 48TMU, 16ROPs), I will keep It at 2GHz so that would mean 3072 GFLOPs.
Memory: DDR5 4800Mhz -> 76.8 GB/s.
Just this alone would mean 50% higher GPU performance.

Now let's calculate IPC between Polaris and RDNA2.
RX 6600XT review
RX 580(36CU;2304SP:144TMU:32ROPs; 256bit 8GHz GDDR5; 185W TBP) vs RX 6600XT(32CU; 2048Sp:128TMU:64ROPs; 128bit 16GHz GDDR6 + 32MB IF; 160W)
Performance: 48 vs 99 at Full HD (+106%)
GFLOPs: 6175(1340MHz) vs 10,605(2589MHz) ( +72%)
Bandwidth: 256 GB/s vs 256 GB/s + 32MB IF
I got 20% better IPC.

Let's calculate performance over 5700g IGP: 100*1.5*1.2 = 180%.
Now It would look like this:
Rembrandt IGP is 33-35% faster than RX 560.
Rembrandt IGP is 12-22% faster than GTX1050.
Rembrandt IGP is 1-7% faster than GTX1050 Ti.
GTX 1650 is 19-27% faster than Rembrandt IGP.
I got this with only 4800MHz DDR5 and 2GHz IGP. What would happen If we use 6400Mhz(+33% BW) and 2600MHz clocks(+30%)?