Zen 2 APUs/"Renoir" discussion thread

Page 17 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146
It is worse, but the 4700G with 4133 vs 3400G 3200 was like what? 10%? Remember that we know, from the 3200G that Vega 8 at 1600mhz more or less matches the 3400G stock with the same rams, at 2100mhz with those rams should be way faster than that if not GPU bottlenecked. Its expected that gpu-light games, and specially shader-light games to perform better on Renoir due to the massive GPU clock diferences.

BTW those results are in line with the Korean review as well. Keep in mind that it is DDR4-4133 vs 3200, but it could be worse.
Shader-light?

This isn't a CPU. All modern games created in the last few years are inherently designed to take advantage of well over 704 shaders. What you're getting mistaken by is the bandwidth requirements which vary from game to game. Some games require greater memory bandwidth than others.

Picasso will perform better with games that need less bandwidth. This number of games is going to rapidly shrink with time.
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
Shader-light?

This isn't a CPU. All modern games created in the last few years are inherently designed to take advantage of well over 704 shaders. What you're getting mistaken by is the bandwidth requirements which vary from game to game. Some games require greater memory bandwidth than others.

Picasso will perform better with games that need less bandwidth. This number of games is going to rapidly shrink with time.

Thats a good point, but remember that the shader workload is not the same in CSGO, LOL, DOTA, Rocket League, than in Witcher 3, AC, RE, RDR2 or Death Stranding, etc...
I used a Vega 8 for gaming for a year, depending on the game, you there are cases were you just cant match the 3400G by OCing unless you go very very far (1700mhz), and bandwidth has nothing to do with that.

But what your are saying is correct.

BTW, i expect a massive uplift in GPU requirements due to next gen, all Vega apus will have a hard time with new games.
 

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
Since RAM bandwidth is the ultimate limiting factor for APUs I wonder if AMD is just keeping RDNA2 for DDR5 since that would likely be even more bottlenecked by the limits of DDR4.
 
  • Like
Reactions: Tlh97 and Thibsie

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,075
136
Since RAM bandwidth is the ultimate limiting factor for APUs I wonder if AMD is just keeping RDNA2 for DDR5 since that would likely be even more bottlenecked by the limits of DDR4.
then the console specs make no sense relative to PS4/XB1X , so im going to say no.
 

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
then the console specs make no sense relative to PS4/XB1X , so im going to say no.
??? The consoles' (current and next gen) nearly have an order of magnitude higher bandwidth, that's no comparison to what PC APUs get. We are currently around the bandwidth level an Xbox 360 had.
 

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,075
136
??? The consoles' (current and next gen) nearly have an order of magnitude higher bandwidth, that's no comparison to what PC APUs get. We are currently around the bandwidth level an Xbox 360 had.
i talking about shader flop to memory bandwidth ratio. if RDNA2 had worse or the same bandwdith per flop then Vega then the console designs would look very different.
 

Thunder 57

Platinum Member
Aug 19, 2007
2,647
3,706
136
I think the mistake was to use 7nm for Renoir, AMD should have realised that 7nm demand was going to be way too high due to console chips.

Could AMD make a 15W 8 core on 12nm? I don't think so. At least certainly not at the clocks of Renoir.

I trought RDNA2 had better bandwidth efficiency,

Do we know AMD is skipping Navi and going to RDNA2 for the next APU? I strongly suspect that, but wasn't sure if it has been confirmed. RDNA always felt like it would be short lived and RDNA2 was the "real" next gen GCN. I think they wanted to get rod of the GCN name because of its reputation. RDNA is certainly way better a least efficiency wise, but I think RDNA2 will be an even bigger leap.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146
Since RAM bandwidth is the ultimate limiting factor for APUs I wonder if AMD is just keeping RDNA2 for DDR5 since that would likely be even more bottlenecked by the limits of DDR4.
Either that or they're planning on the next major uptick to everything but CPU with Rembrandt (and evem then 5nm goes brrrrr), and Cezanne is just Renoir w/Zen 3 to try and speed up launch to make it compete with TGL-U/H by launching in Q1 21.

So far for Rembrandt we have it potentially being:

- 5nm?
- RDNA2
- Zen3(+?)
- Supports DDR4 (will have an AM4 version), LPDDR5 and DDR5(?). Maybe LPDDR4X too. Nearly 100% sure on the first, mostly sure on the middle two, last is a complete guess.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146
Thats a good point, but remember that the shader workload is not the same in CSGO, LOL, DOTA, Rocket League, than in Witcher 3, AC, RE, RDR2 or Death Stranding, etc...
I used a Vega 8 for gaming for a year, depending on the game, you there are cases were you just cant match the 3400G by OCing unless you go very very far (1700mhz), and bandwidth has nothing to do with that.

But what your are saying is correct.

BTW, i expect a massive uplift in GPU requirements due to next gen, all Vega apus will have a hard time with new games.

On point 1, yes and that's caused by memory bandwidth requirements. The lower the bandwidth you actually need per CU, the more CUs can spend time not starved for memory and actually working.

Also, as for having to OC really really far, compare TFLOPs between your two examples of 3400G stock and 3200G at 1700mhz.

3400G at 1400mhz stock operation is still greater compute power than a 3200G at 1700mhz. If you're playing a game with low memory bandwidth requirements, the 3400G's lead will be more sustained as those shaders can do their job just fine regarfless. One with jigh requirments, nd suddenly compute is no longer the constraint.

As for the last bit - it depends based on the game in question. Especially at the beginning, many games are going to aim for 4K60, so at 1080p or 720p the APUs will hold up just fine.

Or you can wait for RMB for some juicy RDNA2 with all the good stuff like VRS packed in.
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
Could AMD make a 15W 8 core on 12nm? I don't think so. At least certainly not at the clocks of Renoir.

It is hard to come up with something similar to Renoir on mobile because it is very very good.

I think it is posible, to make, just not with the same perf or power efficiency.
 

amd6502

Senior member
Apr 21, 2017
971
360
136
It is hard to come up with something similar to Renoir on mobile because it is very very good.

I think it is posible, to make, just not with the same perf or power efficiency.

I doubt porting Zen2 to 12nm would be worth it.

If there were a real demand for deskstop focused APUs then the way to do it would be to leverage the 7nm efficiency of the chiplets with a 12nm IO hub with integrated 12 to 14CU Vega igpu.

im starting to belive it is best to keep Renoir mobile-only.

Because of the Intel development and the 7nm production scarcity that is around for quite some time, the majority of it will be mobile only imho.

I think rare salvage parts like 43**g's will be limited availability and OEM only for a year. DIY will see 4600g and 4700g shortly though as those are much more common die salvage derived from dies that did not make the mobile voltage cutoff, and they also have a role to fill that cannot be filled by the 3400g.

Im under the impression that Renoir Vega has a lower "IPC" than Picasso Vega, the 4700G with DDR4-4133 is about 10-15% faster than the stock 3400G with just DDR4-3200. But the 3200G at 1600mhz with DDR4-3466 already matches a 3400G stock with DDR4-3466, im having a hard time beliving that the 4700G with a a massive clock and memory boost can only yield a 15% over that(actually less, due to the 3466vs3200). They not only decreassed the CU size, but they also cut the ROP size to half, that may have something to do with that. Ive asked TechEpiphany to do a Vega 8 IPC comparison...

Can you link the halved ROP info? As far as IPC, I'm skeptical. IPC will always go down as you increase frequency. This is especially so for memory intensive code where latency is an issue. Run your program at 4ghz to measure IPC, and then run it again at 1ghz to remeasure. At 1 ghz the program will see a memory latency of 1/4 of what it was in terms of processor cycles. So your CAS 12 memory will look like it's CAS 3 when you're downclocking your processor to 1ghz. (As far as IPC, the effect of dramatically downclocking is similar to having a much improved cache.) So when comparing IPC, always run the comparison at a equal or at least similar frequency.
 
Last edited:

YAYgee

Junior Member
May 4, 2020
20
11
41
Either that or they're planning on the next major uptick to everything but CPU with Rembrandt (and evem then 5nm goes brrrrr), and Cezanne is just Renoir w/Zen 3 to try and speed up launch to make it compete with TGL-U/H by launching in Q1 21.

So far for Rembrandt we have it potentially being:

- 5nm?
- RDNA2
- Zen3(+?)
- Supports DDR4 (will have an AM4 version), LPDDR5 and DDR5(?). Maybe LPDDR4X too. Nearly 100% sure on the first, mostly sure on the middle two, last is a complete guess.
Won't see APUs with RDNA2 before discrete GPU products with RDNA3; so yes Rembrandt for the next major uptick and Cezanne only gets a minor (graphics) bump/focus remains on CPU. Looks like desktop APUs are the unwanted stepchild that cannot be allowed to cannibalize AMD's entry-level GPUs. APUs primarily meant for OEM/SI channels while for retail not much budget will be made available (marketing, sales, support, stock, etc.; retail accounting for 20% of APU sales seems like the best case scenario for now). AMD rather services retail with CPU + discrete GPU products.

Any info on display and video blocks wrt Renoir, Cezanne and Rembrandt (mainly interested in HDMI 2.1 and AV1 decode)?
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146
Won't see APUs with RDNA2 before discrete GPU products with RDNA3; so yes Rembrandt for the next major uptick and Cezanne only gets a minor (graphics) bump/focus remains on CPU. Looks like desktop APUs are the unwanted stepchild that cannot be allowed to cannibalize AMD's entry-level GPUs. APUs primarily meant for OEM/SI channels while for retail not much budget will be made available (marketing, sales, support, stock, etc.; retail accounting for 20% of APU sales seems like the best case scenario for now). AMD rather services retail with CPU + discrete GPU products.

Any info on display and video blocks wrt Renoir, Cezanne and Rembrandt (mainly interested in HDMI 2.1 and AV1 decode)?
Pretty sure Renoir does HDMI2.1 - media and display engine were brought over from RDNA. AV1 decode - not with RDNA2 I don't think. Maybe with RDNA3. Whether or not Rembrandt will get the same treatment as Renoir when it comes to media/display engine I have no clue.
 
  • Like
Reactions: YAYgee

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
i talking about shader flop to memory bandwidth ratio. if RDNA2 had worse or the same bandwdith per flop then Vega then the console designs would look very different.
Oh you were interpreting my post as saying RDNA2 having worse or the same bandwdith, that's not what I was trying to get at. What I'm thinking is that to make the most of the new design you'd want plenty bandwidth as well, scale up like AMD did with CPU cores and appears to be doing with CUs in Big Navi as well. That's not possible with the APUs on AM4/DDR4 due to the inherent bandwidth cap. So they may prefer backporting several optimization parts like efficiency improvements and increase in frequency. Save the work on adapting the whole RDNA2 design for APUs for the time when the DDR5 bandwidth also allows some more room to significantly scale up the CU count for maximum impact.
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
Can you link the halved ROP info?

GPU-Z reports 8 ROPS for all Renoir Vegas, both mobile and desktop.
CPU-Z-e1595487516206.jpg


Even the 3000G has 16 rops
gpuz-overclocking.gif
 
  • Like
Reactions: Tlh97 and amd6502

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
I doubt porting Zen2 to 12nm would be worth it.

It would be better than Zen+ at 12nm (3200G/3400G)
It will also free a lot of 7nm volume, no need to have that much higher perf/watt in desktop. A 12nm GloFo Zen 2 APU with better iGPU than current Vega 8 would be way better and a lot cheaper for Desktop use vs 4700G.
This would also benefit OEMs, as AMD could keep 7nm APUs for OEM use only.

If there were a real demand for deskstop focused APUs then the way to do it would be to leverage the 7nm efficiency of the chiplets with a 12nm IO hub with integrated 12 to 14CU Vega igpu.

There is a huge market for Desktop APUs, the problem is that AMD doesnt have the resources to create more dies and invest money, time an engineers to OEMs to create all those products that those APUs could be used on (AIOs, Ultra SFF and more).

Just take a look what Intel is capable off with its pathetic APUs. Intel sells more APUs than entire AMDs CPU lineup.
You can find Intels APUs (iGPU only systems) from NUCs, AIOs, uSFF etc etc.

AMD is only selling to the DIY APU market, where they dont need to invest resources as they only sell the APU and thats it. But even in that market they dont invest much, you can clearly see this because the support and product mix is pathetic. Thats why they dont sell a lot of Desktop APUs, not because there is no TAM for it.

Perhaps in a couple of years when they will have more revenue and become more confident, they will increase OEM spending and we will see more APUs in desktop. But even them i highly doubt it, they will rather spend more for Laptops than desktop.
 

juergbi

Junior Member
Apr 27, 2019
12
14
41
There is a huge market for Desktop APUs, the problem is that AMD doesnt have the resources to create more dies and invest money, time an engineers to OEMs to create all those products that those APUs could be used on (AIOs, Ultra SFF and more).

Just take a look what Intel is capable off with its pathetic APUs. Intel sells more APUs than entire AMDs CPU lineup.
You can find Intels APUs (iGPU only systems) from NUCs, AIOs, uSFF etc etc.

Isn't that huge market mainly about business and other non-gaming desktops where Renoir is a great fit? For AIOs and Ultra SFF power efficiency is an important factor and hence a hypothetical Zen 2 backport to 12nm would be worse than Renoir.

OEMs could presumably also build AIOs with Renoir or Matisse and a dGPU on the same board (like a laptop) if they want to build a product with a more powerful GPU. However, this is obviously not needed to compete in the market with Intel iGPUs (so far).
 
  • Like
Reactions: Tlh97

naukkis

Senior member
Jun 5, 2002
702
570
136
It would be better than Zen+ at 12nm (3200G/3400G)
It will also free a lot of 7nm volume, no need to have that much higher perf/watt in desktop. A 12nm GloFo Zen 2 APU with better iGPU than current Vega 8 would be way better and a lot cheaper for Desktop use vs 4700G.
This would also benefit OEMs, as AMD could keep 7nm APUs for OEM use only.

Zen2 is design optimized for 7nm density. It can't be backported to lesser density nodes. AMD itself stated that even with 7nm they had to cut L1i in half to be able to fit bigger UOP-cache, and that 256-bit FMA in 12nm would have used too much power.

If backporting could be done with moderate cost AMD would have done it.
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
Isn't that huge market mainly about business and other non-gaming desktops where Renoir is a great fit? For AIOs and Ultra SFF power efficiency is an important factor and hence a hypothetical Zen 2 backport to 12nm would be worse than Renoir.

OEMs could presumably also build AIOs with Renoir or Matisse and a dGPU on the same board (like a laptop) if they want to build a product with a more powerful GPU. However, this is obviously not needed to compete in the market with Intel iGPUs (so far).

For desktop , even for AIOs and uSFFs , highest power efficiency is not that much of a concern. You can use 35W TDP SKUs.
A 12nm Zen 2 would have better power efficiency than 12nm Picasso and it will still be highly competitive against Intels 14nm SKUs (Core i3 10100 etc).

Renoir is a lot more expensive, this hypothetical 12nm ZEN 2 APU (4C 8T) would replace the Picasso 3200G/3400G models.
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
Zen2 is design optimized for 7nm density. It can't be backported to lesser density nodes. AMD itself stated that even with 7nm they had to cut L1i in half to be able to fit bigger UOP-cache, and that 256-bit FMA in 12nm would have used too much power.

If backporting could be done with moderate cost AMD would have done it.

Die size is important at 7nm due to wafer capacity, lower yields and higher wafer cost.
Those are not a concern at GloFos 12nm, they would have a little bigger die vs Picasso but still it would be cheaper than 7nm Renoir die.
 

naukkis

Senior member
Jun 5, 2002
702
570
136
Die size is important at 7nm due to wafer capacity, lower yields and higher wafer cost.
Those are not a concern at GloFos 12nm, they would have a little bigger die vs Picasso but still it would be cheaper than 7nm Renoir die.

If they just same design to bigger manufacturing process achievable clock frequency will be greatly reduced and resulting CPU would be much slower than 14nm optimized Zen1 design. AMD did tweak their Zen2-core to 7nm process for over a year.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
Instead of going through all the effort to practically reinvent the wheel to put a Zen2 design on GloFo 12nm, instead, do a Raven2/Dali style 7nm die with a single 4 core CCX and 4-6CU of Vega iGPU with only 8+4+4 PCIe lanes. It would be roughly 60% of the size of the existing Renoir die size, not require a drastic relayout of Zen2's core to address the different design rules of GloFo 14/12, and still net you a LOT more usable dies per wafer for a segment of the market that really doesn't care much about performance. Just converting 10% of the Renoir Wafers to that could supply a whole lot of low end mobile and desktop chips in the 6-25 watt range. Since there would be a considerable lead time on this, the yields on the Renoir wafers should be very mature, and the need for 4 core, 75% CU scavenging of dies would be greatly reduced.