Discussion AMD Cezanne/Zen 3 APU Speculation and Discussion

[dr1337]

https://twitter.com/x/status/1292844896374718470

Fresh leak out today, not much is known but at least 8cu's is confirmed. Probably an engineering sample, core count is unknown and clocks may not be final.

This is very interesting to me because cezanne is seemingly 8cu only, and it seems unlikely to me that AMD could squeeze any more performance out of vega. A cpu only upgrade of renoir may be lackluster compared to tigerlake's quite large GPU.

What do you guys think? Will zen 3 be a large enough improvement in APU form? Will it have full cache? Are there more than 8cus? Has AMD truly evolved vega yet again or is it more like rdna?

 

[Shivansps]

Measurably faster and noticeably faster are two different things unfortunately. 1CU +100 mhz, or more with OC is 12.5% more compute resources plus 5-15% more frequency, yet, for configurations that have similar memory configurations, the net performance gain is usually no more than 2-3 fps, or, usually under 10%. Whereas something as simple as increasing rank interleaving show outsized gains to performance over raw bandwidth increases.

VEGA is, in a word, not that great for memory bandwidth efficiency (as compared to more modern or competing architectures). It does well with gargantuan memory bandwidth (maybe it was designed for HBM?) But seems to suffer with regular dram. It IS an improvement over Polaris in many ways, but, not much in the memory department.

Its a pitty i couldnt do a better comparison with the 5700G, i had already finished the 5600G test by the time i was told i could test a 5700G, and i didnt had the same rams anymore at that point.
The best comparison is both the 5600G and 5700G at 2.3Ghz with 2xDDR4 3600, and the 5700G is always better except for one case. And i know 2xDDR4-3600 is not enoght, 4xDDR4-3600 may had given better results.

 

[Asterox]

A little more detailed, or hm longer Vega 7 iGPU test.

PC Specs:

Ryzen 5 5600G 3.90 ghz (Boost 4.40 ghz)
CPU Cooler: AMD Ryzen Cooler Stock
SSD: 120GB SATA WD Green
GPU: Radeon Vega 7 Graphics
16GB DDR4 3200Mhz (2x8 dual channel / CL 16 / Model: G.SKILL AEGIS)

 

[jpiniero]

AMD Ryzen 5 5600G "Cezanne" APU die has been pictured up close - VideoCardz.com

Fritzchens Fritz delivers new high-resolution imaging of AMD Cezanne silicon AMD Ryzen 5000 APU processor pictured (very) up close. The latest infrared images from Fritzchens Fritz show a detailed structure and internals of AMD Cezanne APU. These are the latest mobile (and also recently became...

videocardz.com

Did this get posted here? It really looks like AMD literally took Renoir, replaced the Zen 2 cores with Zen 3, and shifted down some other parts leaving a decent amount of white space. Talk about lazy.

 

[LightningZ71]

Lazy, or, using the resources available at the time. AMD was certainly doing a bit better by that point, but they weren't exactly swimming in cash yet either...

 

[moinmoin]

We essentially knew that already:
"The Ryzen 5000 Mobile processor series is an upgrade over last year’s Ryzen 4000 Mobile processor series. AMD co-designed both of these processor families to re-use key parts of the chip design, enabling a fast time-to-market and quicker upgrade cycle"

AMD Ryzen 9 5980HS Cezanne Review: Ryzen 5000 Mobile Tested

www.anandtech.com

The turn around time from Zen 3 launch to Zen 3 APU was significantly faster this time.

Edit: The white space below the iGPU should actually fit exactly 4 CUs, matching the total 11 CUs available in Raven Ridge and Picasso. AMD likely did tests whether that's worth the effort and opted against it due to the known and much discussed lack of memory bandwidth.

 

[Shivansps]

Edit: The white space below the iGPU should actually fit exactly 4 CUs, matching the total 11 CUs available in Raven Ridge and Picasso. AMD likely did tests whether that's worth the effort and opted against it due to the known and much discussed lack of memory bandwidth.

Thats a bold statement when we know that more CUs have more performance even on heavy memory limited scenarios, and it is being tested a lot now, that dual rank has a huge effect on IGP perf. Look at these results with dual rank:

The hard reality here is that 8CU was good enoght for AMD, because Picasso was always terrible at mobile and Intel had nothing. The only problem AMD have right now is that Xe changed things, but it with RMB thats a non-issue, and with Cezanne the only problem they have is on desktop, and only with the 5300G, that would be a hard sell at 6C/12T i5 prices with I3 CPU performance and 3200G IGP perf.

 

[moinmoin]

How is that a bold statement when we already know the decision AMD took there? AMD is willingly wasting significant wafer area for these white spaces. For that to make financial sense to AMD the opposite, using that space for actual functionality, must have cost even more in their internal calculation than the decision they ended up with.

Not that it really matters considering (unlike with the desktop chips) AMD appears to want to uphold the yearly cadence for mobile chips, so we can expect it to be replaced by Rembrandt soon.

 

[LightningZ71]

I don't think that it would have any real difference to performance in the end. The very high clocked 8CU solution is roughly as fast as the 11 CU solution that was in raven ridge. It's not in the right area to use for bus interfaces. It's not in a good spot for more cache. It's too small for an effective NPU. You might be able to fit a fixed function encoder/decoder there, but would it really help ASPs? And, more importantly, would the added cost of the additional IP and development resources be recoverable? I don't think that it would.

 

[IntelUser2000]

Yes, design considerations aren't only about die size and silicon! Faster time to market is also very important. Being able to make different variations quickly is also important. Intel chips have "white space" too. They aren't copying and pasting in paint, nor do they simply press a button and say "Fabricate CPU" and you see a nicely cooked batch of Zen 5 boxes 5 mins later!

It shouldn't matter to us, unless any of you are working for AMD. Good products at a decent price is all that matters.

 

[Shivansps]

5600Gv5700Gv1050TI with 4xDDR4-3600 (one of the best configurations you can have)

 

[Gideon]

5600Gv5700Gv1050TI with 4xDDR4-3600 (one of the best configurations you can have)

Wow, it's much closer to 1050 Ti than i thought. I can see Rembrandt with decent DDR5 being up to 1060 or 1650 Super performance for 1080p.

Nothing earth-shattering of course, but it would finally be something I would personally consider an acceptable stop-gap GPU replacement @ 1080p. Current APUs work for some titles but still require 720p for some others, which I wouldn't consider acceptable at this day and age.

 

[Shivansps]

Not sure if dropping to 720P is actually needed with these APUs, except maybe some corner case like CP2077, most things should run at 900P if it doesnt run ok at 1080p.

But yeah, RMB should provide good enoght 1080P performance.

 

[blckgrffn]

Wow, it's much closer to 1050 Ti than i thought. I can see Rembrandt with decent DDR5 being up to 1060 or 1650 Super performance for 1080p.

Nothing earth-shattering of course, but it would finally be something I would personally consider an acceptable stop-gap GPU replacement @ 1080p. Current APUs work for some titles but still require 720p for some others, which I wouldn't consider acceptable at this day and age.

While that seems pretty optimistic based on some of what I saw there (didn’t watch much because I don’t typically watch 20 minute videos, would love the graph set) the 5700G was within striking distance on some games and 30% slower in others - I will grant you this - RDNA 2 and DDR5 in optimal configs should consistently beat the 1050ti. I think that, actually, is pretty darn good. Lots of people would love to have a 1050ti right now and would pay $100+ for the privilege.

Nearly all F2P games would work really well at that level, an Minecraft and Starcraft and all those evergreen games would run well too.

 

[Shivansps]

I dont know how to make the performance forecasting based on raw numbers, but i would expect that:

-IPC: RDNA2 CU > Vega CU
-Bandwidth Efficiency: RDNA2 > Vega
-CU Number: RMB > Cezanne
-Bandwidth: DDR5 > DDR4 (is memory rank expected to still make a difference on DDR5? considering DDR5 changes how memory channels work?)

RMB looks to be a massive upgrade in everything.

 

[uzzi38]

I dont know how to make the performance forecasting based on raw numbers, but i would expect that:

-IPC: RDNA2 CU > Vega CU
-Bandwidth Efficiency: RDNA2 > Vega
-CU Number: RMB > Cezanne
-Bandwidth: DDR5 > DDR4 (is memory rank expected to still make a difference on DDR5? considering DDR5 changes how memory channels work?)

RMB looks to be a massive upgrade in everything.

Also perf/W should be drastically in RDNA2's favour.

 

[Gideon]

Also perf/W should be drastically in RDNA2's favour.

Exactly. RX 5700XT is about as fast as a Vega VII with 40CUs vs 60CUs and RX 6900XT is almost 100% faster with just 20 more CUs (granted with Infinity Cache and very high clocks).

1050 Ti is only about 50% faster in games that aren't limited heavily by Bandwidth (like Witcher 3). Rembrandt should have no trouble beating it as soundly:

• Rembrandt has 50% CUs than Cezanne
• It has DDR5 which even at the lowest 4800 MT/s should deliver more than 50% effective bandwidth (according to memory makers own claims). If the price of faster modules (say 6400 MT/s) ends up reasonable in 2022 it will eveen more BW still.
• RDNA2 usually clocks extremely well. While at 65WTDP (88W actual) we might be still somewhat limited I could easily see GPU clock speeds like 2.8 Ghz for 95W models (if there aree any) and overclocks

This makes me somewhat hopeful for it being in the rough GTX 1660 performance ballpark.

 

[leoneazzurro]

This makes me somewhat hopeful for it being in the rough GTX 1660 performance ballpark.

I think it's better to lower a bit the expectations as the bandwidth will be anyway shared with the CPU cores, and we don't even know if there will be Infinity Cache (probably not). Also, in the raw power there will be too much disadvantage with a 1660. I think it's more reasonable to expect something in the GTX1650/Rx470 range (which is anyway quite good for an APU).

 

[Gideon]

I think it's better to lower a bit the expectations as the bandwidth will be anyway shared with the CPU cores, and we don't even know if there will be Infinity Cache (probably not). Also, in the raw power there will be too much disadvantage with a 1660. I think it's more reasonable to expect something in the GTX1650/Rx470 range (which is anyway quite good for an APU).

Arrgh I made a typo at posting, I wanted to say 1060 not 1660, It will absolutely not reach the latter which is 20% faster than the 6GB 1060 and 30% faster than the 3GB model.

It will certainly depend on the game. In very bandwidth limited games (Witcher 3, etc) it will absolutely still be slower. Bbut on average I can see it winning, if it clocks anywhere near as well as RX 6xxx cards so-far.

Speculation

To illustrate I made this table with 3 hypothetical 6700G versions vs current stock of APUs and GPUs:

Model	Flops	Bandwidth
5700G	2.04 TFlops (8 CUs @ 2 GHz)	57.6 GB/s @ 3600 MHz DDR4
6700G (conservative)	3.37 TFlops (12 CUs @ 2.2 GHz)	76.8 GB/s @ 4800 MT/s DDR5 **
6700G (slightly optimistic but within reason, XMP memory)	3.68 TFlops (12 CUs @ 2.4 GHz*)	102 GB/s @ 6400 MT/s
6700G (theoretical limit at say 24/7 OC and very expensive memory)	4.3 TFlops (12 CUs @ 2.8 GHz*)	134 GB/s @ 8400 MT/s (though even to 12600 MT/s will be available I doubt it will run)
GTX 1050 Ti	2.14 TFlops (6 SMs @ 1.39 GHz)	112.1 GB/s
GTX 1060 3GB	3.93 TFlops (9 SMs @ 1.7 GHz)	192.2 GB/s
GTX 1060 6GB	4.375 TFlops (10 SMs @ 1.7 GHz)	192.2 GB/s

* Pulled out of thin air but based on mobile 6600XT game clock and 160W TDP from here.
Considering it's with GDDR6 and it's a game clock (not max clock) 2.4 Ghz should be quite achievable for a 12CU 65W (88W max sustained) desktop part

** According to Crucial it effective bandwidth (vs theoretical) is actually 30% better vs DDR4 at the same clock speed so it's much closer to 2x the actual bandwidth of 3600Mhz DDR4


The unknown variables (that decide the game)

In the end it depends on a lot of unknown factors that can totally skew the result:

• What MT/s will the Infinity Fabric actually supportl? Will 1:2 mode actually improve performance at extremely high memory clocks?
  • 4800 MT/s is a pretty safe bet but if it can't even go to, say 6400 MT/s (which cannot be ruled out on first gen), then it probably won't compete.
• What will be the final GPU clock-speeds and how much would a 95W TDP or 24/7 OC improve things
  • Vega 8 OCs to 2.4 Ghz at around 130W (according to the above video) if RDNA2 can do even 2.8 Ghz it will be in the 1060 ballpark

TL;DR:

Even with very conservative clocks (seeing what RDNA2 actually does vs Vega) Rembrant will have over 50% more compute power than 1050 Ti and close-enough memory bandwidth. It is almost guaranteed to be faster, even with cheapest DDR5.
1060 would still have a hefty bandwidth advantage but pretty similar compute power to the more optimistic projections. Expecting it to be faster is probably too much yeah, but it could easily be within 10%.

 

[RTX2080]

Efficiency comparison (sad no temperature but fan profile at the button instead)



edit: I forgot to post the source

https://twitter.com/x/status/1425103050864308236

 

[coercitiv]

Efficiency comparison (sad no temperature but fan profile at the button instead)

Performance regression coupled with a 20%+ system power increase?! The 11700 data is weird to say the least.

 

[Shivansps]

i would love to see it performing like a RX570+/GTX1650 Super, but i think RX5300 seems more realistic, in fact considering the RX5300 has 22CU RDNA1 with 168GB/s, i would expect the full die APU to perform a lower than that. Finding RX 5300 results is not easy, it is a very low volume card, and i dont trust the ones i saw.

 

[leoneazzurro]

i would love to see it performing like a RX570+/GTX1650 Super, but i think RX5300 seems more realistic, in fact considering the RX5300 has 22CU RDNA1 with 168GB/s, i would expect the full die APU to perform a lower than that. Finding RX 5300 results is not easy, it is a very low volume card, and i dont trust the ones i saw.

Well according to a test I saw the 5300XT in full HD performs around a RX470 (45%-75% better than a Rx560). I'd say that expecting performances around ( a bit lower) than that should be fine.

 

[Gideon]

i would love to see it performing like a RX570+/GTX1650 Super, but i think RX5300 seems more realistic, in fact considering the RX5300 has 22CU RDNA1 with 168GB/s, i would expect the full die APU to perform a lower than that. Finding RX 5300 results is not easy, it is a very low volume card, and i dont trust the ones i saw.

Good point. While the clockspeed is pretty low, it has tons of extra CUs and memory BW so you are probably correct.

Well according to a test I saw the 5300XT in full HD performs around a RX470 (45%-75% better than a Rx560). I'd say that expecting performances around ( a bit lower) than that should be fine.

Exacty. Regardless where it lands that's a very significant difference.

EDIT for clarity:
Would I be willing to build a new rig with an APU as a temporary GPU replacement right now? - No, not really
Would I do it with a RX 5300 equivalent Zen3+ APU? - If the GPU prices remain as they are - Yes, most certainly!

 

[blckgrffn]

Good point. While the clockspeed is pretty low, it has tons of extra CUs and memory BW so you are probably correct.


Exacty. Regardless where it lands that's a very significant difference.

EDIT for clarity:
Would I be willing to build a new rig with an APU as a temporary GPU replacement right now? - No, not really
Would I do it with a RX 5300 equivalent Zen3+ APU? - If the GPU prices remain as they are - Yes, most certainly!

Same. Right now I would likely still scrounge for a GTX 950 or similar. If you can get up to a GTX 970 or something the performance difference is so huge it makes sense for a "gamer" to try find a $150 CPU and a $200 GPU rather than go all in on a $350 ultimate APU.

 

[LightningZ71]

Remember, there is a non-trivial amount of interest in uSFF PCs that can be used in entertainment centers for gaming on living room TVs. Getting something in there that can competently play things at 1080p is a big deal. We're very close with the 5700g, with things like FSR upsampling 900p to 1080p with reasonable detail levels being playable in most situations. It's not unreasonable for the next gen of RDNA2 based iGPU equipped APUs to be able to hit decent 1080p/60 with decent quality settings. That makes for a lot of very functional laptops and STBs without needing a dGPU.