AMD 7nm Vega 20 32GB GPU 3DMark Performance Benchmark Leaked

Hotwired · 30 Apr 2018 at 13:20

32GB of memory.

HBM memory at that.

Anyone in their right mind thinking this is a gaming card?

TheRealDeal · 30 Apr 2018 at 13:24

Rezident said:
wccftech so you know it doesn't need to be based in any fact or reality.

Read the article. It is based on fact as the run has been verified. What's not known is if there was anything misread.

Sure the site is not reliable but they are only reporting on a run of 3dmark that actually happened on the 24th April and is an amd chip of some kind.

D.P. · 30 Apr 2018 at 15:10

bru said:
Could you link to AMD's public announcement of Vega 20 please, because I couldn't find one only rumour and speculation.

This was mentioned in AMD's announcement at CES this year, but was also announced last year (before Vega 10 was even released0, and has been re-iterated recently with the financial report

sideways14a · 30 Apr 2018 at 15:27

This thing could be anything, but... if its clock speed is even a bit right then its going to be a monster that makes vega look like an ant.

but... its prob not

Vince · 30 Apr 2018 at 15:40

I think I shall start preparing the wallet

drunkenmaster · 30 Apr 2018 at 20:20

D.P. said:
vega 20 is just Vega 10 with the missing FP64 added back in, and some small refinements.

It is aimed at HPC and wont hit the consumer space.

I don;t see the need for all this speculation. AMD publicly announced all of these some time back.

It's only being talked about so far in HPC, but if it's capable of significantly higher clocks and 70% more performance in a similar power bracket, once production is up there is little to no reason not to release a gaming version.

The fairly natural assumption to draw from that is, Navi is close enough behind that a higher clocked Vega 10 on 7nm isn't worth producing for gaming cards. If it wasn't there would be no reason to produce 3 times as many cards and sell 2/3rds as Vega 64 v2.

4K8KW10 · 30 Apr 2018 at 21:20

KillBoY_UK said:
if its still "Vega" and the per clock figures is true then it would likely have to be a massive die with extra CU's which will still be bandwidth staved

GCN cannot scale beyond 4096 shaders........ someone said.

keybaud · 30 Apr 2018 at 23:09

For those that don't trust wccftech, they even link to the Videocardz article:

https://videocardz.com/76076/early-amd-vega-20-3dmark11-benchmark-result-emerges

bru · 1 May 2018 at 02:12

It all depends on the clock speed, because as it stands it isn't any faster than a Vega 56/1070.

3dmark 11 search (uncheck show valid results only)

Of course the details are hidden but they are in the Wccftech link

D.P. · 1 May 2018 at 06:50

drunkenmaster said:
It's only being talked about so far in HPC, but if it's capable of significantly higher clocks and 70% more performance in a similar power bracket, once production is up there is little to no reason not to release a gaming version.

The fairly natural assumption to draw from that is, Navi is close enough behind that a higher clocked Vega 10 on 7nm isn't worth producing for gaming cards. If it wasn't there would be no reason to produce 3 times as many cards and sell 2/3rds as Vega 64 v2.

It is talked about onyl for HPC because that is its only market. AMD have added the FP64 that was missing for HPS, but this has no value in gaming. AMD will have a different architecture in mind for gameras next year.

The whole clock speed stuff is nonsense because it is so obviously just an error with 3Dmark. And even if it was accurate you then have to make an assumption that some how clock speed can be increased, which is not realistic if the entire architecture was changed to allow much greater IPC. But that just highlights why it can;t possibly be realistic clock speed because Vega20 is not a whole new architecture, but Vega 10 with FP64 support. It woudln't even be called Vega if it was such a radical new design.

Kaapstad · 1 May 2018 at 07:02

TheRealDeal said:
Read the article. It is based on fact as the run has been verified. What's not known is if there was anything misread.

Sure the site is not reliable but they are only reporting on a run of 3dmark that actually happened on the 24th April and is an amd chip of some kind.

3dmark are not very good at getting it right.

I remember when they reported that I had a 3dmark score for a RX480 that would beat a Titan V lol, the bench software missed the fact that I was using 3 cards in crossfire.

4K8KW10 · 1 May 2018 at 09:14

Kaapstad said:
3dmark are not very good at getting it right.

I remember when they reported that I had a 3dmark score for a RX480 that would beat a Titan V lol, the bench software missed the fact that I was using 3 cards in crossfire.

The RX 480s are scaling very well. Must be close to 100% from the second card and over 50% by the third card.

TheRealDeal · 1 May 2018 at 10:48

Kaapstad said:
3dmark are not very good at getting it right.

I remember when they reported that I had a 3dmark score for a RX480 that would beat a Titan V lol, the bench software missed the fact that I was using 3 cards in crossfire.

If you read the article and from my own experience they have been pretty good at getting things correct with the last few releases. They were pretty spot on with the 480/580 and Vega. Mistakes do happen though but it's hard to believe that at this stage of the game it would misread Vega 64 so wrong. Maybe someone tricked it to do so. It's not just getting clocks wrong but everything pretty much. So I am to believe that it got clocks, vram, gpu Id and some other stuff as well.

I am open to this being a new AMD gpu but I don't know what. Vega 20 was rumoured way back to have over 6000 shaders so could just be that.

Schnippzle · 1 May 2018 at 10:56

Article of conjecture and assumptions. Why do mods allow this tripe? As Kaap said, unreliable data.

TheRealDeal · 1 May 2018 at 11:05

Schnippzle said:
Article of conjecture and assumptions. Why do mods allow this tripe? As Kaap said, unreliable data.

So a real 3dmark run is conjecture and tripe. It might just be a bad reading but it might not be either. We have had many runs of gpu's in the past like this that turned out to be spot on. It can't just be dismissed until proven fake. So assuming it's a real run then speculation on it's performance can be made. Off course nobody knows but what else are we to do in times where it's so slow in the gpu world.

stuie · 1 May 2018 at 11:07

But can it mine? Cause if not it wont sell

D.P. · 1 May 2018 at 11:31

TheRealDeal said:
If you read the article and from my own experience they have been pretty good at getting things correct with the last few releases. They were pretty spot on with the 480/580 and Vega. Mistakes do happen though but it's hard to believe that at this stage of the game it would misread Vega 64 so wrong. Maybe someone tricked it to do so. It's not just getting clocks wrong but everything pretty much. So I am to believe that it got clocks, vram, gpu Id and some other stuff as well.

I am open to this being a new AMD gpu but I don't know what. Vega 20 was rumoured way back to have over 6000 shaders so could just be that.

It isn't necessarily Bega64 though. It may well be Vega20, and it is reading the clocks wrong.

shankly1985 · 1 May 2018 at 12:42

3DMark on my Vega just reports the max rated clock speed 1630 in my case. This isn't correct has my GPU speed is between 1500/1600

TheRealDeal · 1 May 2018 at 13:33

D.P. said:
It isn't necessarily Bega64 though. It may well be Vega20, and it is reading the clocks wrong.

Yea it could well be that as well. All I am saying is it could well be Vega 20 or something totally different. What I really doubt is it's a Vega 64 with such a bad read this late in the day. Usually a bad read would get something wrong but pretty much everything points me to something not on the market yet.

Kaapstad · 1 May 2018 at 14:02

Here is a good example of 3dmark getting it very wrong.

Check out my Time Spy score for a single RX 480

The result is only marked as invalid because I used beta drivers.

https://www.3dmark.com/spy/117720

I was actually using 3 cards in crossfire but 3dmark totally missed this.