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.

 

[leoneazzurro]

He don't need to prove, it's logical that 4266 is faster than 3200 on a 5800U because of the bandwidth bottleneck. It should definitely improve the fps by at least 10%

Dude, DDR4-4266 is not LPDDR4-4266X. There are fundamental differences in the way these two memories work, like LPDDR5-6400 is not DDR5-6400 so the results are by no means comparable.

 

[ahimsa42]

all will be resolved in a few weeks when hopefully in game benchmarks begin to surface. from what AMD said at CES, there will be a lot of choices of RMB laptops from many vendors so some may have high speed LPDDR & others may not. as long as there are decent options for igpu only systems i think most people will be able to find something which works for them-the amount they will have to pay & the availability though are two issues which are entirely unknown at this point.

 

[moinmoin]

For both Renoir and Cezanne availability had been such a slow build up over the months after launch that people had to be convinced that they actually exist. I expect the same to happen again with Rembrandt and will as usual refer to my Geizhals links for proof of (initially scarce) availability at least in Germany.

 

[uzzi38]

Dude, DDR4-4266 is not LPDDR4-4266X. There are fundamental differences in the way these two memories work, like LPDDR5-6400 is not DDR5-6400 so the results are by no means comparable.

Memory bandwidth should be similar between the two, although lower latency DDR4 will have a minute difference to mem b/w vs the LPDDR4X. That latency increase gets compounded by the fact that RNR/CZN-U runs the FCLK at 1:2 rate in mobile. So where TE only shows a ~10% improvement with DDR4-4266, realistically in mobiel you're going to be looking at 7-8% at best.

 

[eek2121]

I have a feeling that RMB iGPU performance claims are going to be similar to Polaris perf/W claims.

Wait, what? I'm pretty sure TGL-U launched before M1.

TGL-U was 12-28W and there were lots of "15W" chips. I am saying to expect 15W chips to gradually go away and the new standard to be somewhere around 28W.

 

[leoneazzurro]

Memory bandwidth should be similar between the two, although lower latency DDR4 will have a minute difference to mem b/w vs the LPDDR4X. That latency increase gets compounded by the fact that RNR/CZN-U runs the FCLK at 1:2 rate in mobile. So where TE only shows a ~10% improvement with DDR4-4266, realistically in mobiel you're going to be looking at 7-8% at best.

In reality we don't know, because the way the memory controller works with LPDDR is quite different, and so the efficiency and latency with LPDDR may be very different respect to a test with a desktop DDR4 system (which has also different latencies compared to laptop memories). This is why in a comparison called "apples to apples" not only he has not the apples on one side because he has no Rembrandt, but he is also putting some oranges by using a different memory type.

 

[mikk]

Dude, DDR4-4266 is not LPDDR4-4266X. There are fundamental differences in the way these two memories work, like LPDDR5-6400 is not DDR5-6400 so the results are by no means comparable.

It doesn't matter. LPDDR-4266 has quite a big bandwidth advantage over DDR4-3200, you can be sure that a bandwidth limited 5800U will run faster by at least 10% in many typical gaming workloads. We don't need a proof for that, it's logical.

 

[leoneazzurro]

It doesn't matter. LPDDR-4266 has quite a big bandwidth advantage over DDR4-3200, you can be sure that a bandwidth limited 5800U will run faster by at least 10% in many typical gaming workloads. We don't need a proof for that, it's logical.

This would be true if the two memory types were equivalent but they are not as they work differently and one is optimized for power and the other for performance. And yes, you need to prove that with practical tests before drawing a conclusion, just because reality is quite often different from assumptions.

 

[ahimsa42]

i find the controversy over the 6500XT to be quite interesting. to me, all of the issues with getting a pc graphics card points strongly to getting a laptop instead as the igpu alone in RMB will be fine for most people & hardcore gamers can just spend the extra for a dgc. i also think the obsession with frame rates & high resolution is hilarious as with out a counter how many people would notice the difference between 60fps @ 1080 & 100fps @ 1440? if a game is fun to play, are more FPS & slightly better visuals going to make that much of a difference that someone is willing to pay large sums of $ for them? i am very glad that i am quite content with 60fps @ 1080 myself.

 

[Thibsie]

Well, if a youtuber or hardware review site has nothing to bash, how can he/it justify its existence ?
I don't say this is an axcellent product, I just say, it is completely logical to try and serve the market with what's already somewhere in a drawer, making sure it doesn't disappear in a crypto farm somewhere.

 

[ahimsa42]

Well, if a youtuber or hardware review site has nothing to bash, how can he/it justify its existence ?
I don't say this is an axcellent product, I just say, it is completely logical to try and serve the market with what's already somewhere in a drawer, making sure it doesn't disappear in a crypto farm somewhere.

people often forget that AMD's only goal is to generate profits so this is always going to be their primary decision maker-the same with review sites.

 

[deasd]

6900HX GB

https://twitter.com/x/status/1485860473862254592

 

[eek2121]

Came here to post that. Honestly? Landed about where I personally expected it. Doesn’t beat Alder Lake, but Alder Lake has no issues using more than 100W of power.

 

[coercitiv]

Came here to post that. Honestly? Landed about where I personally expected it. Doesn’t beat Alder Lake, but Alder Lake has no issues using more than 100W of power.

Intel will win the H competition by pushing TDP way past 45W but that's not an option for U parts. I think their 2+8 die will face dire competition, and ironically this is where we were convinced hybrid approach would shine the most.

What I'm most interested right now is dGPU gaming benchmarks: with finite power budgets the CPU with better gaming efficiency will leave more power headroom to the "ISO" GPU. Should be fun to watch and interpret.

 

[ahimsa42]

hopefully we will start seeing some real world benchmarks in early Feb.

 

[gdansk]

Intel will win the H competition by pushing TDP way past 45W but that's not an option for U parts.

You say it's not an option but the 2+8 have up to 55W PL2. For some reason I suspect the PL2 will last long enough for some workloads (*cough* Geekbench).

 

[tamz_msc]

You say it's not an option but the 2+8 have up to 55W PL2. For some reason I suspect the PL2 will last long enough for some workloads (*cough* Geekbench).

Yeah but AMD has the advantage of having 8 "P" cores at 15 W compared to 2 P + 8 E cores at 15 W on Intel. I believe AMD will comfortably win in the -U segment.

 

[clemsyn]

6900HX GB

https://twitter.com/x/status/1485860473862254592

Looks like 12800H beats the upcoming best AMD mobile processor in Geekbench. Its still just one benchmark though. Can't wait for reviews on this new processor.

 

[Abwx]

Looks like 12800H beats the upcoming best AMD mobile processor in Geekbench. Its still just one benchmark though. Can't wait for reviews on this new processor.

12800H has a 115W PL2.

Geekbench MT scores of a Lenovo device using a 12900HK tested at different power settings range from 7850, wich is surely at 28-30W, to 14400 or so that are likely performed at around 115W.

Geekbench Search - Geekbench

browser.geekbench.com

 

[tamz_msc]

12800H has a 115W PL2.

Geekbench MT scores of a Lenovo device using a 12900HK tested at different power settings range from 7850, wich is surely at 28-30W, to 14400 or so that are likely performed at around 115W.

Geekbench Search - Geekbench

browser.geekbench.com

Lower than expected Geekbench scores are almost always due to some other background process interfering during the pause between each sub-test, not necessarily due to lower power limits. So I won't give much importance to those low scores.

 

[Abwx]

Lower than expected Geekbench scores are almost always due to some other background process interfering during the pause between each sub-test, not necessarily due to lower power limits. So I won't give much importance to those low scores.

Background process that use 45% of the total throughput.?
Why doing a GB test in those conditions.?

That s a SAME device with same ST score for each run, only the MT results are different, wich tell that they were done at different power levelS.

LENOVO 82RF - Geekbench

Benchmark results for a LENOVO 82RF with an Intel Core i9-12900H processor.

browser.geekbench.com

 

[IntelUser2000]

Yeah but AMD has the advantage of having 8 "P" cores at 15 W compared to 2 P + 8 E cores at 15 W on Intel. I believe AMD will comfortably win in the -U segment.

2+8 is roughly equal to 6, so, yes Rembrandt will have an advantage.

We'll see how they do with 4+8 and 6+8. They'll likely need 6+8 to win against 28W Rembrandt. 4+8 is going to be very interesting.

I think their 2+8 die will face dire competition, and ironically this is where we were convinced hybrid approach would shine the most.

It does sound like Alderlake is Intel's Ryzen sometimes.

2+8 might be better suited to 9W than 15W, just like 8+8 is too little for the -S space.

In laptops I find interesting how NOBODY cares about how Tigerlake-U loses massively against Cezanne and Renoir. So whatever Alderlake gains is in this very area where currently no one cares about.

But Intel did need to address the massive MT performance lead at some point.

Background process that use 45% of the total throughput.?
Why doing a GB test in those conditions.?

It doesn't matter. It happens, and that's why relying on user submitted benchmarks for comparisons are asinine. The massive score discrepancy happened way before people even knew how to adjust TDP settings, or were able to. Or back in Core 2 days where there was no such thing. Who the hell knows the exact reason for all that? No one does. It's a waste of time analyzing.

That's why you rely on review day with Apples-to-Apples comparisons with software, hardware, OS, drivers, firmware all equal as much as possible even if it doesn't look like it might affect it, it might, so you do it for the sake of doing it right.

 

[tamz_msc]

Background process that use 45% of the total throughput.?
Why doing a GB test in those conditions.?

I'm saying this from my own testing of Geekbench. MT scores can be severely affected by interfering background tasks.

Dell Inc. Inspiron 14 5410 vs Dell Inc. Inspiron 14 5410 - Geekbench

browser.geekbench.com

35% difference in my own case.

 

[Abwx]

I'm saying this from my own testing of Geekbench. MT scores can be severely affected by interfering background tasks.

Dell Inc. Inspiron 14 5410 vs Dell Inc. Inspiron 14 5410 - Geekbench

browser.geekbench.com

35% difference in my own case.

45% difference with 14C is not the same thing, your 35% would be 10-15% at most with such a CPU.

Besides that s SEVERAL tests from Lenovo, i dont think that they would mess the tests this way, and moreover that s clean systems with about no software other than Windows and eventually an antivirus + firewall.

 

[deasd]

6600H

https://twitter.com/x/status/1486303520789872643