News AMD Announces Radeon VII

Page 19 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
@railven I don't think GCN has a future beyond Vega20. So maybe IP blocks can be recycled in Navi, or maybe not. But as a larger part of Navi's future design, I do not think that AMD will specifically be building on anything GCN.

I don't know how much to read into it, but if you look at earlier roadmaps you get a general sense of their plan:

AMD-GPU-roadmap-900x507.jpg


I'm assuming that Navi is still pretty heavily based on GCN, but it's pretty clear that they need a new architecture or that what they have would need some major overhauls because it has some clear limitations. Being on 7nm might allow them some advantages in terms of process to catch up to NVidia, but from what we've seen of Radeon VII (even though it's designed for a different market) it's pretty clear that GCN is tapped out or very close to it in the current form.

A lot of people took the Next-Gen to mean a new architecture that's been developed from a fresh start rather than being another iteration of GCN. It's pretty likely that by the time they were designing Polaris, AMD would have enough understanding of its limitations to start designing a new GPU architecture.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
@Mopetar nobody at this point knows if Navi is the stepping-off point to a new uarch. The tile-based Navi rumours seem to indicate that there is at least a partial departure from GCN designs. Or maybe, as you say, Navi is going to be GCN 6.0 .
 

PeterScott

Platinum Member
Jul 7, 2017
2,605
1,540
136
I'm assuming that Navi is still pretty heavily based on GCN, but it's pretty clear that they need a new architecture or that what they have would need some major overhauls because it has some clear limitations....

I also expect Navi to be another GCN tweak.

I think AMD has a decent general purpose GPU architecture, but NVidia really has step up on efficiency tricks. Better culling of triangles or things like that, then they work with Devs to implement things that work well with their efficiency tweaks.

In games where the efficiency tweaks don't work as well, AMD pulls ahead on brute force, as in competing products (rx580vs1060, Vega64vs1080) AMD usually has more transistor than NVidia, and does better at raw compute.

I also got the impression the first time a saw that architecture roadmap, that Navi was pretty much and extension of Vega, and Next-Gen was the big change.

But basically I think AMD needs some efficiency tricks that work, and that should work with a GCN architecture. It will be interesting to see what Navi brings.
 
  • Like
Reactions: krumme

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
Yea that's why I said I don't think Navi is coming this year, it will be Vega on 7nm.
It looks like Navi vs the Nvidia 3000 series both on 7nm at the end of 2020 at best.

Navi could start at the lower segments than Vega II ($699) , for example a 200mm2 navi at 7nm could be the foundation for sub $499 cards within 2019.

NAVI confirmed for 2019 ,

From AMDs Q4 2018 earnings

As we entered 2019, we are preparing to launch our strongest product portfolio ever. In gaming, we will launch our high end Radeon 7 GPU in February, followed by our next generation Navi GPUs later in the year.
 
Last edited:

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
I also expect Navi to be another GCN tweak.

I think AMD has a decent general purpose GPU architecture, but NVidia really has step up on efficiency tricks. Better culling of triangles or things like that, then they work with Devs to implement things that work well with their efficiency tweaks.

In games where the efficiency tweaks don't work as well, AMD pulls ahead on brute force, as in competing products (rx580vs1060, Vega64vs1080) AMD usually has more transistor than NVidia, and does better at raw compute.

I also got the impression the first time a saw that architecture roadmap, that Navi was pretty much and extension of Vega, and Next-Gen was the big change.

But basically I think AMD needs some efficiency tricks that work, and that should work with a GCN architecture. It will be interesting to see what Navi brings.

Raja didn't get the bonus.

Its obvious Vega technology didn't pan out as expected. All the fancy stuff in the white paper didn't work. Its actually a pretty huge problem. If not a minor disaster for amd. Millions of Vega dies out there - especially the apu with a huge need of that technology because of bandwidth constrains. They were saved by mining and later zen.

Now we get navi and it's about time the stuff starts to work if navi is gcn based as we asume. But its 1 year before retirement. Is it actually worth it to make drivers for it?
 

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
In the latest q&a with Lisa with among other AT she gets away with saying Vega is a success because the technology is used in many places.

She shouldn't get away with that nonsense but should be confronted. Yes it got the name Vega but it's hardly any different from Polaris. It's just spin. Vega works but it's slow and inefficient and as a project evolved over 4-5 years it was a disaster.
 

Karnak

Senior member
Jan 5, 2017
399
767
136
In the latest q&a with Lisa with among other AT she gets away with saying Vega is a success because the technology is used in many places.

She shouldn't get away with that nonsense but should be confronted. Yes it got the name Vega but it's hardly any different from Polaris. It's just spin. Vega works but it's slow and inefficient and as a project evolved over 4-5 years it was a disaster.
Because she is right. Just look at the latest earnings report, highest datacenter GPU revenue of all time for AMD. How is that a disaster?

It's only underwhelming as a pure desktop gaming GPU where Vega is "clocked to the moon" just to reach a certain amount of performance to compete. Becase the Vega architecture is build for HPC and not gaming which is most likely the case with Navi. But besides that Vega works very well. Whether it's as a HPC GPU for datacenters or as an integrated graphics for the APU's.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
In the latest q&a with Lisa with among other AT she gets away with saying Vega is a success because the technology is used in many places.

She shouldn't get away with that nonsense but should be confronted. Yes it got the name Vega but it's hardly any different from Polaris. It's just spin. Vega works but it's slow and inefficient and as a project evolved over 4-5 years it was a disaster.

4-5 years?? Vega launched on August 14, 2017. And its not inefficient at all if its run the way it was designed to be run. AMD has stated that both Polaris and Vega were designed based on request from specific customers. And those customers wanted very low power GPU's. When both Polaris and Vega are run at lower clocks, their efficiency goes WAY up. But, in their non-OEM guise, their power usage is high because AMD cranks the clocks up, and runs the core voltage higher so that they don't have to bin every GPU (Which is why undervolting AMD GPU's works so well). But to AMD, performance is more important than power usage, which makes sense. Because power usage does not matter to the vast majority of buyers. Using less power doesn't make your game run better. And even nVidia's power consumption has gone way up with their RTX cards.

Vega and Polaris are very different. Yes both are GCN based GPU's, but saying its hardly different is hyperbole. The two GPU's are very different, compare the dies side by side and its readily apparent.
 

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
AMD line item reports by sector, not product line. In the Q&A Lisa Su did give a rough estimate of 7.5% of quarterly revenue came from GPU data center which works out to ~$100M.
That's a third just to cover 7nm port for vega. Only upfront fixed process cost for single die. Fixed process cost. You really haven't started covering fixed cost...
Go look at NV.
 

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
4-5 years?? Vega launched on August 14, 2017. And its not inefficient at all if its run the way it was designed to be run. AMD has stated that both Polaris and Vega were designed based on request from specific customers. And those customers wanted very low power GPU's. When both Polaris and Vega are run at lower clocks, their efficiency goes WAY up. But, in their non-OEM guise, their power usage is high because AMD cranks the clocks up, and runs the core voltage higher so that they don't have to bin every GPU (Which is why undervolting AMD GPU's works so well). But to AMD, performance is more important than power usage, which makes sense. Because power usage does not matter to the vast majority of buyers. Using less power doesn't make your game run better. And even nVidia's power consumption has gone way up with their RTX cards.

Vega and Polaris are very different. Yes both are GCN based GPU's, but saying its hardly different is hyperbole. The two GPU's are very different, compare the dies side by side and its readily apparent.
Who oem wanted Vega?
In datacenter it can run nearly okay. But the 100m quarter doesn't cover cost at all not to say profit. Consumer needs to use it too and here it's not running as it was designed. It's running as a Polaris. It's broken.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
That's a third just to cover 7nm port for vega. Only upfront fixed process cost for single die. Fixed process cost. You really haven't started covering fixed cost...

Those are Q4 2018 results. Most of AMD's pro GPU sales in that time period were Mi25s - Vega10. Vega20 revenues won't hit in a big way until this year. Keep your eyes peeled.

Who oem wanted Vega?

Presumably, the system integrators selling Mi25-based compute clusters to AI/deep learning customers. AMD has apparently made a lot more money off Vega10 in the pro graphics arena than they have in the consumer market, which is pretty interesting when you consider how many of those RX Vega64s went to miners at inflated prices.
 

Dribble

Platinum Member
Aug 9, 2005
2,076
611
136
I think it's pretty likely Navi will basically be Polaris @ 7nm + DDR5X/6 support, and support for whatever the latest HDMI format is.
 

Paratus

Lifer
Jun 4, 2004
16,613
13,295
146
What I find interesting is AMD has continued to compete at the high end with with same basic design for three generations at three different nodes.

4096/256/64 - HBM @ 28nm, 14nm, and 7nm.
 
  • Like
Reactions: Head1985

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
Its 5 years since Maxwell launched.
https://www.anandtech.com/show/7764/the-nvidia-geforce-gtx-750-ti-and-gtx-750-review-maxwell
We get Radeon 7 excactly 5 years after. Can it do some of the same tricks or does Radeon 7 mean we have to wait 7 years?

The 750 "Maxwell" is not the same as the 'Maxwell" that made up cards like the 980Ti. Not sure what you mean by tricks, but the Vega 64 was already faster than a 980Ti. I am not sure what your point of Vega II coming five years after the 750 means? Yes nVidia had faster cards out two and a half years ago, but they were Pascal, not Maxwell.
 

stuff_me_good

Senior member
Nov 2, 2013
206
35
91
GCN came out like 2012, so it's like 7 years old. :eek: Damn that makes me sad. Yet another bulldozer from AMD. :( Will they ever learn to make new architecture every 5 years? Probably not.
 

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
The 750 "Maxwell" is not the same as the 'Maxwell" that made up cards like the 980Ti. Not sure what you mean by tricks, but the Vega 64 was already faster than a 980Ti. I am not sure what your point of Vega II coming five years after the 750 means? Yes nVidia had faster cards out two and a half years ago, but they were Pascal, not Maxwell.
Nv had tricks that made Maxwel arch far more efficient and also faster than prior gens. Some of the same tricks was in Vega hardware but never was made to work. They worked out the gate on Maxwell.

Thats why Vega is a bulldozer for gaming purpose. If we get a 7nm port without these things fixed it's incredible. I simply can't fathom it.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
GCN came out like 2012, so it's like 7 years old. :eek: Damn that makes me sad. Yet another bulldozer from AMD. :( Will they ever learn to make new architecture every 5 years? Probably not.

Why should AMD design a whole new architecture all the time when nVidia doesn't either? Each revision of GCN has changes that make it an iterative design. nVidia does the same thing. The only real difference comes down to naming. AMD adds a revision number after GCN. For instance Tahiti was GCN 1.0, and Vega is GCN 1.5. nVidia doesn't name their architectures this way. They simply use a new name for each iterative generation. Pascal for instance was very similar to Maxwell. It wasn't a new design, it was an iterative set of changes.

And I don't think you can compare Vega to bulldozer. Bulldozer was almost entirely designed by software, there wasn't any hand tuning done by a person. And it went out and tried an entirely different architecture with its shared FPU's and modules.

Vega on the other hand may use more power than an nVidia chip, but thats because it was designed for lower power use cases, so when AMD cranks up the clocks, the power usage spikes. If AMD would be ok with lower the clocks like 10%, power usage drops significantly. The same goes for Polaris. Polaris was designed for Sony and MS, and at the lower clocks it runs at for them, it uses very little power. But you crank it up to 1300Mhz and power consumption balloons out. This doesn't make them bad designs. It just means AMD is stretching the envelope that they were designed to run in. AMD has stated if it wasn't for OEM's paying for the design of these chips, that they may not exist at all because of their financial situation at the time.
 
  • Like
Reactions: kawi6rr

Qwertilot

Golden Member
Nov 28, 2013
1,604
257
126
That down clocking thing is kind of half true. The power draw gap is obviously exaggerated by AMD shipping cards at frankly dumb clocks in order to try and keep up better performance wise.

That doesn’t mean they’re actually perf/watt competitive - look at what happens in gaming notebooks. That’s a realy lucrative market and AMD essentially nowhere.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
That down clocking thing is kind of half true. The power draw gap is obviously exaggerated by AMD shipping cards at frankly dumb clocks in order to try and keep up better performance wise.

That doesn’t mean they’re actually perf/watt competitive - look at what happens in gaming notebooks. That’s a realy lucrative market and AMD essentially nowhere.

I think that has more to do with OEM's being locked to Intel. Most Ryzen laptops have Vega GPU's in them, but most laptops are Intel. AMD is in the majority of Macs sold however. Not gaming, but still a lot of machines.
 
  • Like
Reactions: happy medium