Speculation: Ryzen 4000 series/Zen 3

Page 120 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

YAYgee

Junior Member
May 4, 2020
20
11
41
Are HDMI 2.1 and AV1 (decode) expected for desktop Renoir (DCN 2.1 <=> 2.0 in Navi and VCN 2.5 <=> 2.0 in Navi)? It sounds like Tiger Lake will support AV1 decoding.

B550 boards with HDMI 2.1 will soon be available. E.g. Aorus Pro (supporting a maximum resolution of 4096x2160@60 Hz; support for HDMI 2.1 version, HDCP 2.3, and HDR). Why only support 4K@60Hz?
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
Are HDMI 2.1 and AV1 (decode) expected for desktop Renoir (DCN 2.1 <=> 2.0 in Navi and VCN 2.5 <=> 2.0 in Navi)? It sounds like Tiger Lake will support AV1 decoding.

B550 boards with HDMI 2.1 will soon be available. E.g. Aorus Pro (supporting a maximum resolution of 4096x2160@60 Hz; support for HDMI 2.1 version, HDCP 2.3, and HDR). Why only support 4K@60Hz?
That spec has me puzzled as well, though that is DCI 4K and not UHD, so I guess it might be "DCI 4K60 (or UHD120)"? Sure hoping so at least. Don't think Renoir has a hardware AV1 decoder (think that would have been mentioned in some laptop review if that was the case).
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Base clock typically doesn't apply when the iGPU is under load in a TDP-limited scenario, so that argument doesn't stand up to scrutiny. We have no idea what CPU clocks Renoir on the desktop can maintain when the iGPU is being stressed.

It probably can run both the CPU and iGPU at full unlike the mobile variants because desktop chips have plenty of room for both.

But running on the iGPU means the frame rates are going to be quite low, so the CPU doesn't have to be running at full. The CPU core itself will use much more power if paired with RTX 2080 for example.
 
  • Like
Reactions: lightmanek

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
It probably can run both the CPU and iGPU at full unlike the mobile variants because desktop chips have plenty of room for both.

But running on the iGPU means the frame rates are going to be quite low, so the CPU doesn't have to be running at full. The CPU core itself will use much more power if paired with RTX 2080 for example.
That depends on the game and quality settings. It's absolutely possible to run high FPS (and thus load the CPU with lots of draw calls) even on an iGPU, though it's obviously easier with a powerful dGPU. Still, the fact is that base clock only applies when the iGPU is idle. We have no reason to suspect a massive CPU efficiency increase with Renoir, so it stands to reason that power draw under a similar load at similar clocks will be similar. Which means that the iGPU will at beat eat into the boost headroom of the CPU, at worst (and more likely) cause it to go below base clock unless power draw restrictions are lifted. This is expected and perfectly okay, but one should have realistic expectations.
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
Can’t wait to pick up a Ryzen 7 4700 G or GE to replace my i3-530. Hoping the they will be in the $200-250 range.
The Ryzen 7 3700X is currently $295 on Newegg. This has the same number of cores and threads, the same TDP, similar clocks, but adds an iGPU. And you expect it to get a significant price cut? Yeah, I don't believe for a second that these APUs will be cheap.
 
  • Like
Reactions: Makaveli and teejee

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
The Ryzen 7 3700X is currently $295 on Newegg. This has the same number of cores and threads, the same TDP, similar clocks, but adds an iGPU. And you expect it to get a significant price cut? Yeah, I don't believe for a second that these APUs will be cheap.

I would argue that the 4800/4900H have roughly the same performance as the 2700X in most benchmarks that aren't crazy AVX2 heavy or just solidly small memory footprint, single core in nature. So, going by that logic, I don't think that suggesting that it would be about $75 more expensive than the most recent retail price listings for the 2700X is that far out of line. It appears that, in most cases, a 3600X with a $50 used video card would perform on par, or faster than it. That's the environment that it's selling into. Remember, it may have 8 cores, but it's also got half the L3 and a smaller chip surface area to dissipate it's heat output over. It also appears to always run the memory bus async, so there's some added latency there that being monolithic probably won't hide.
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
I would argue that the 4800/4900H have roughly the same performance as the 2700X in most benchmarks that aren't crazy AVX2 heavy or just solidly small memory footprint, single core in nature. So, going by that logic, I don't think that suggesting that it would be about $75 more expensive than the most recent retail price listings for the 2700X is that far out of line. It appears that, in most cases, a 3600X with a $50 used video card would perform on par, or faster than it. That's the environment that it's selling into. Remember, it may have 8 cores, but it's also got half the L3 and a smaller chip surface area to dissipate it's heat output over. It also appears to always run the memory bus async, so there's some added latency there that being monolithic probably won't hide.
You have a very idealistic view of how companies decide on pricing for their products. Do you honestly believe that AMD will price the upcoming APUs based on a combination of used and new parts optimized for best value? Because if you do, you are going to be disappointed. Prices are decided based on the perceived worth of the product compared to competing new products - used part prices do not factor into this whatsoever (at least outside of the extreme low end, which is why the used market usually represents great value). The same goes for comparisons to previous-gen parts: those are heavily discounted, have long since abandoned any MSRP, and AMD has likely long sinced ceased caring what those sell for beyond getting the margins they want on clearing out their inventory.

The only relevant comparison is current generation alternatives - no matter how comparable other alternatives might be in terms of performance. For current generation alternatives, as I said, the 3700X is currently on sale at Newegg for $295, with the 3600X at $269. Do you honestly think they would price an 8c16t APU below the 3600X? That is quite frankly ridiculous, even if the lower cache means the 3600X might be faster in most consumer workloads. You're also placing too much emphasis on the performance disadvantages of Renoir. While memory speed is indeed decoupled from IF clock, there's nothing saying you couldn't manually set them to matching speeds should that perform better for you. Also, a smaller chip surface to dissipate heat? What? The Renoir die is 150mm2. A Zen 2 CCD is less than half of that, and located toward the side of the IHS rather than centrally as Renoir dice are no doubt going to be. And given matching TDPs, heat density will be at worst equal, but typically lower for Renoir - though a lot of that area is of course taken up by the iGPU too. CPU performance if limited to TDP will of course be lower under iGPU load, but ... it's an APU. That's how they work, the tradeoff is expected. I would, at best, expect these APUs to be ever so slightly cheaper than similar CPUs, though I wouldn't be surprised at price parity either. I would be really happy if you turned out being right, but I'd much rather have realistic expectations than be disappointed after forming a belief on a poorly thought out basis.
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
AMD will launch the APUs at the same MSRP as the CPUs with the same amount of cores. I still remember before AMD launched desktop Raven Ridge I even argued that they'd make the APUs more expensive than the comparable CPUs. Making them cheaper would be financially insane if the cost per mm² 7nm silicon is the biggest part of the manufacturing cost since Renoir chips are already larger than Matisse chips.
 
Mar 11, 2004
23,031
5,495
146
SMT4, AVX-2048, a 300% IPC increase and quantum teleportation wifi. That's what I'm putting my money on.

And for the twist, Intel will have GF fab it on their freshly announced 5nm FD-SOI! Which you have it all wrong, the quantum teleportation will be with the new heatspreader. But it'll require a lot of bandwidth which required a truly next gen interconnect (its what Jim Keller's been working on), "Cold Width Solution"
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
I have no idea if this is legit,

Cezanne APU is Zen 3 + Vega (again)



image-20200526140253545.png


Google Translation

AMD side: Zen 3 and various APUs
AMD's new products are still relatively clear. The product they are about to release soon should be Matisse Refresh, which is an upgraded version of the frequency of several currently selling Zen 2 processors. It is believed that they will have the XT model suffix. Then there is the desktop version of the Renoir processor, which may debut with Matisse Refresh, or it may be launched at some point later.

The above two series do not exist in AMD's official information. The only available information is the Zen 3 architecture and the schedule of the new generation of processors. It can be seen that AMD will launch a new generation of processors based on Zen 3 this year. Code-named Vermeer, what can be learned from various channels is that Zen 3's single-threaded performance will be improved by double-digit levels, which means that it will definitely not be as simple as changing the CCX structure, and the kernel micro-architecture will also have Larger upgrade.

Let's talk about AMD's layout in the mobile market. A few days ago, there is a form of integration of AMD ’s future processor roadmap. Let ’s take a look. The data comes from @KOMACHI_ENSAKA:

What is more useful to us in this form is the layout on MSDT and APU next year. You can see that AMD has laid out an APU called Cezanne next year, and then there is a new APU codenamed Rembrandt, but the table does not mark them. What kind of architecture will be used, but we got the latest news from anonymous sources.

Cezanne will use a combination of Zen 3 + Vega 7 instead of the rumored RDNA 2, and it will continue to use TSMC's N7 process. Like Renoir, it will have both low-voltage and standard-pressure versions. Rembrandt is a huge update. The CPU architecture will use Zen 3+, and the GPU will finally get rid of the Vega architecture and switch to RDNA 2. In terms of process, it will be updated to TSMC's 6nm, which is the optimized version of the N7 process. And will support DDR5, LPDDR5, in addition, USB4 and PCIe 4.0 two new buses will also be available on this generation of architecture.

AMD is also planning a processor called Van Gogh (Van Gogh) at ultra-low voltage. It is like a simplified version of the PS5 and Xbox Series X custom processors. It also uses the Zen 2 + RDNA 2 architecture, but TDP restrictions At 9W, for thin and light mobile devices.

If you use a road map to simply represent the layout of AMD on the desktop CPU and APU, this can be the case, but there is no guarantee that the time point will be correct.

Summary: CPU market changes
In the CPU market in the next two years, if nothing unexpected happens, Intel will experience a period of relatively weak, they are seriously dragged down by the process technology, they decided too late to put the new core on the desktop processor, resulting in their own desktop The processor has begun to dominate at the same frequency. How much Rocket Lake-S can recover is still unknown. In the mobile market, it is still a mystery whether Tiger Lake and the optimized 10nm + process can make the energy consumption better. The implementation of the roadmap by AMD is still relatively good. At the end of this year, we should see the Zen 3 processor. With the success of Renoir, more manufacturers will choose to use it in the mobile market next year. APU.
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
And for the twist, Intel will have GF fab it on their freshly announced 5nm FD-SOI! Which you have it all wrong, the quantum teleportation will be with the new heatspreader. But it'll require a lot of bandwidth which required a truly next gen interconnect (its what Jim Keller's been working on), "Cold Width Solution"
So the IHS has quantum entangled copper atoms on the top and bottom surfaces, transferring heat through quantum mechanics? Brilliant! That must have an effective infinite watts per meter Kelvin, with infinite instantaneous heat transfer. Though I guess the better solution then would be for them to separate the "receiving" end of the quantum heat teleporter IHS separately so that you could, say, stick it in your freezer. Maybe that's for the next generation? I guess Intel could just ship all the "receiving"-side IHS parts to, say, Antarctica, and give all their users infinite free sub-zero cooling? It would come with a significant risk of condensation though. Lots of shipping damage too. "Yeah, hi, I'd like to RMA my CPU, it arrived encased in a solid block of ice so cold my fingers stuck to it and now I have gangrene in my frostbitten fingers and there's ice chips all over my motherboard."

Now we just need them to implement native 10GbE (preferably bugged so that it delivers 10-100Mbps like their 2.5GbE solution) and Xe graphics and we'd be golden.


On a more serious note:
I have no idea if this is legit,

Cezanne APU is Zen 3 + Vega (again)



image-20200526140253545.png


Google Translation
This is a bit disappointing, though I guess not surprising. A bit of a back-and-forth in terms of pros and cons.
Pros: Zen 3 upgrade path on AM4 up to at least 8 cores, which I frankly had thought wouldn't happen at all. Likely improved process and efficiency for even higher iGPU clock than Renoir, and the possibility of more Vega CUs (though it might also stay the same of course).
Cons: No Navi APUs on AM4, so no real upgrade incentive for anyone buying Renoir unless they need more CPU perf. Necessitates a platform upgrade including expensive new RAM for anyone wanting to move to these APUs.

Overall it sounds okay to me; it lowers the chance I'll be annoyed by something much better arriving a few months after I build my HTPC, and it kind of makes sense that implementing RDNA might not be worth it until it can be matched with DDR5. Though it would have been really cool to be able to put that to the test.
 

dr1337

Senior member
May 25, 2020
311
516
106
Cezanne APU is Zen 3 + Vega (again)

I don't buy it, the article says its going to be vega 7??? But at the same time van gogh uses rdna2? It just doesnt line up. The article also says that rembrandt is going to be on 6nm... Given its 2022 time frame this seems incrediblely unlikely.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Cezanne APU is Zen 3 + Vega (again)
I don't buy it, the article says its going to be vega 7???
imho, the only way Cezanne is being Vega-ish is if it's going APU CDNA1 w/ Infinity Architecture. With Rembrandt following on with a CDNA2 design and 2nd gen Infinity Architecture. With a custom BGA for 4-way/8-way motherboards.

8-way being 64-core Zen3 cores, 64-cores of Wavefront64 ISA, plus whatever competes with Intel's GNA/IPU at AMD times 8.
 
Last edited:

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
imho, the only way Cezanne is being Vega-ish is if it's going APU CDNA1 w/ Infinity Architecture. With Rembrandt following on with a CDNA2 design and 2nd gen Infinity Architecture. With a custom BGA for 4-way/8-way motherboards.

8-way being 64-core Zen3 cores, 64-cores of Wavefront64 ISA, plus whatever competes with Intel's GNA/IPU at AMD times 8.
We'll never see CDNA in an APU. The total compute power of such a chip is so small that using a compute-focused architecture will be a drop in the bucket; utterly and completely meaningless and easily overshadowed by orders of magnitude by the 1-8 (or more) accelerators likely to be installed in a PC that needs that kind of compute. For anything else, RDNA does compute plenty good enough (including for mobile and low-end MSDT professional applications). All the while using CDNA in APUs would make that silicon entirely unsuited for consumer use (unless you want them to start developing yet another GPU driver series?).
 

amrnuke

Golden Member
Apr 24, 2019
1,181
1,772
136
Didn't someone at AMD actually explicitly refute Zen2+ at some point? Or am I mixing that up with something else?
 

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
Didn't someone at AMD actually explicitly refute Zen2+ at some point? Or am I mixing that up with something else?

Matisse refresh is not Zen2+, it is the same Zen2 die just binned better. Not on the same level of changes we had going from Zen1 to Zen1+
 

soresu

Platinum Member
Dec 19, 2014
2,617
1,812
136
We'll never see CDNA in an APU. The total compute power of such a chip is so small that using a compute-focused architecture will be a drop in the bucket; utterly and completely meaningless and easily overshadowed by orders of magnitude by the 1-8 (or more) accelerators likely to be installed in a PC that needs that kind of compute. For anything else, RDNA does compute plenty good enough (including for mobile and low-end MSDT professional applications). All the while using CDNA in APUs would make that silicon entirely unsuited for consumer use (unless you want them to start developing yet another GPU driver series?).
Depends, there is a lot of use for ML HW nowadays, including for VR use cases on the consumer side (foveated rendering, RT denoising).

I could see a future APU having 2-4 CDNA CU's exclusively for ML use in the same way that pretty much every new mobile SoC from any chip designer has an ML accelerator core/cores included now.

That being said I would not expect it in even Rembrandt, perhaps its successor though.
 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
Depends, there is a lot of use for ML HW nowadays, including for VR use cases on the consumer side (foveated rendering, RT denoising).

I could see a future APU having 2-4 CDNA CU's exclusively for ML use in the same way that pretty much every new mobile SoC from any chip designer has an ML accelerator core/cores included now.

That being said I would not expect it in even Rembrandt, perhaps its successor though.
RDNA can handle that nearly as well while being far more flexible and useful in a broader selection of use cases, all the while avoiding adding another layer of driver complications. As such, something like this would be far, far too expensive in terms of silicon area. There is an argument for ML hardware on PCs, but that would require universal APIs that don't exist yet, and an appreciable perf increase over just running the workload on (already incredibly flexible) GPU cores.
Might be on a different node though (N7P?)
Highly unlikely - that would be extremely costly for a limited-run product. Vermeer is coming in the next 6 months after all. It is far more likely that this is just a consequence of a maturing product and process and increased familiarity with both allowing for better bins.