Vega/Navi Rumors (Updated)

[MrTeal]

so You are telling that Big Vega Is slower than GTX1080 ?

I'm not telling you anything. You asked for the link and pasted it for you, along with some commentary that the demo that they're talking about seems to be from the December event given the picture posted.

If the die really is 520mm²+ I would be pretty shocked if it was slower than a GTX1080 though. That would be a pretty massive failure on AMD's part.

 

[cytg111]

plus : It does seem like GloFo have gotten their * together since the first polaris launch, clocks may surprise.
minus : If it is targeted HPC, theres no telling how it'll perform gaming wise comparing mm^2 to mm^2, it may suck major .....

 

[Glo.]

Guys, I will say to you all one thing. Stop the bickering, and have open mind. Vega can be either a Flop or kick Nvidia's ass.

Currently, based on what we have seen with architecture preview, the second option is more possible, and everything now is in the hands of software engineers at AMD to bring proper drivers for the architecture.

According to AMD themselves, Vega has at this point zero drivers, and is running on modified Fiji drivers. So the performance demoed by AMD is WORST POSSIBLE CASE scenario.

 

[happy medium]

If it is targeted HPC, theres no telling how it'll perform gaming wise comparing mm^2 to mm^2, it may suck major .....

That's a very good point.

 

[AtenRa]

plus : It does seem like GloFo have gotten their * together since the first polaris launch, clocks may surprise.
minus : If it is targeted HPC, theres no telling how it'll perform gaming wise comparing mm^2 to mm^2, it may suck major .....

It may, but Hawaii also was made for the HPC market as well as for games and turned out to be the best GPU of the last 4-5 years.

 

[happy medium]

Guys, I will say to you all one thing. Stop the bickering, and have open mind. Vega can be either a Flop or kick Nvidia's ass.

Currently, based on what we have seen with architecture preview, the second option is more possible, and everything now is in the hands of software engineers at AMD to bring proper drivers for the architecture.

According to AMD themselves, Vega has at this point zero drivers, and is running on modified Fiji drivers. So the performance demoed by AMD is WORST POSSIBLE CASE scenario.

I find it very very VERY hard to believe that after a year the gtx1080 will be on the market that AMD can't beat it easily with its fastest chip in May 2017.

AMD is not be competing with a gtx1080 but with a 1180 by the time it releases.

Quote: from Anandtech.

"Why? Well part of the reason is that we’re still months off from the launch of Vega. I believe it’s fair to say that by announcing a first-half of the year launch date when we’re already in 2017 is a strong indicator that Vega will not launch until later in that window, likely some time in Q2. So we’re still a good three to five months out from the launch of Vega, which means AMD doesn’t want to (or need to) release too many details this far out. Rather they can trickle out chosen details for maximum impact."

http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser

 

[crisium]

There must be a lot of HPC crap in there that I don't care about if that's big Vega. So 28nm Fiji is 4096 shaders and 4096-bit at ~600mm^2, and 14nm Vega is 4096 shaders and 2048-bit ~550mm^2? A new process and only ~10% smaller size?

Seems like a total waste of space for gaming is somewhere in there.

I guess AMD doesn't have the R&D to make both a GP100 and GP102 like Nvidia. Out with (most) of the irrelevant to gaming stuff.

But hopefully that's something larger than 4096 shaders in that ~550mm^2 die. And/or we're looking at a major architecture overhaul (i.e. if you downclocked Vega to 1050MHz, it would still demolish the Fury X).

 

[Phynaz]

There are two (2) Vega chips, which one was on that PC case ??? AMD didnt say which one they used.

And let me ask you the same question,

Are you really suggesting that Vega has lower perf/mm2 than Polaris 10 ??

I'm not suggesting anything. I'm observing.

BTW, did that person you told to sell their GTX 980ti and buy a RX 480 ever do it?

 

[Phynaz]

There must be a lot of HPC crap in there that I don't care about if that's big Vega. So 28nm Fiji is 4096 shaders and 4096-bit at ~600mm^2, and 14nm Vega is 4096 shaders and 2048-bit ~550mm^2? A new process and only ~10% smaller size?

Seems like a total waste of space for gaming is somewhere in there.

I guess AMD doesn't have the R&D to make both a GP100 and GP102 like Nvidia. Out with (most) of the irrelevant to gaming stuff.

But hopefully that's something larger than 4096 shaders in that ~550mm^2 die. And/or we're looking at a major architecture overhaul (i.e. if you downclocked Vega to 1050MHz, it would still demolish the Fury X).

A huge chunk of the die and functionality is geared towards HPC. That's where the $$$ is at.

Notice they have already announced the HPC cards but not the graphics cards. That tells you where the focus is.

 

[Phynaz]

So the performance demoed by AMD is WORST POSSIBLE CASE scenario.

Is there any proof of the above?

 

[PhonakV30]

It was running with Fiji driver.

 

[Bacon1]

Is there any proof of the above?

Its literally in the first part of the paragraph you cut out of your quote:

According to AMD themselves, Vega has at this point zero drivers, and is running on modified Fiji drivers. So the performance demoed by AMD is WORST POSSIBLE CASE scenario.

 

[happy medium]

It was running with Fiji driver.

Says AMD........I remember them telling us something about Polaris performance per watt also last time around. We know how that turned out.

I also remember a AOTS demo which made them look great, but we all know how that turned out.

I would be very careful jumping on the AMD bandwagon.

 

[AtenRa]

I'm not suggesting anything. I'm observing.

If you only observing you dont judge others predictions

BTW, did that person you told to sell their GTX 980ti and buy a RX 480 ever do it?

That was for the early DOOM Vulkan horrendous Maxwell performance, it was not a prediction for Polaris 10

 

[Phynaz]

Its literally in the first part of the paragraph you cut out of your quote:

Show me a quote from AMD where they say "the performance demoed by AMD is WORST POSSIBLE CASE scenario."

 

[zinfamous]

I think Vega will come in between 1080-Titan level performance and will spur a 1080ti reaction from NVidia. The slow roll Vega release explains Nvidia's silence on 1080ti. Looks like 2017 is the year for slowass, delayed releases. Also, I don't know how I feel about another 8gb card coming on the scene. AMD used to be ahead of the curve in terms of Vram, but lately they are lagging behind with their high end offerings. I'd rather have 10-12gb of DDR5X than 8gb of HBM2 since the gaming performance will be the same anyway. We've seen DDR5X perform just fine despite my early reservations. People trying to drive new 4k/144hz panels with 8gb might encounter some insecurities in the next year or two to say the least.

The big Vega is "supposed" to be 16gb of HBM2. that was mentioned in a good link yesterday and several times in the past. No, I'm not going to search.
Part of me thinks that the delay here is availability of those 8gb stacks or whatever is needed. Instinct is already set to release, "whenever," with 16GB, right?

 

[Phynaz]

That was for the early DOOM Vulkan horrendous Maxwell performance, it was not a prediction for Polaris 10

You flat out said a 480 would be faster than a 980ti multiple times.

But whatever, this is kind off topic here, so I'm going to drop it.

 

[PhonakV30]

Says AMD........I remember them telling us something about Polaris performance per watt also last time around. We know how that turned out.

I also remember a AOTS demo which made them look great, but we all know how that turned out.

I would be very careful jumping on the AMD bandwagon.

You can't judge Vega's Performance base on Fiji driver. with that , You can't say like this :

I find it very very VERY hard to believe that after a year the gtx1080 will be on the market that AMD can't beat it easily with its fastest chip in May 2017.

AMD is not be competing with a gtx1080 but with a 1180 by the time it releases.

 

[happy medium]

The big Vega is "supposed" to be 16gb of HBM2

That's for the professional market , not for us gamers.

 

[zinfamous]

That's for the professional market , not for us gamers.

link?

 

[AtenRa]

You flat out said a 480 would be faster than a 980ti multiple times.

But whatever, this is kind off topic here, so I'm going to drop it.

I have said that IF we get 2.2x higher density with 14nm over 28nm THEN a 232mm2 Polaris 10 could be faster than GTX 980Ti.

Unfortunately we didnt get 2.2x higher density with Polaris 10 over 28nm so it turned out we only got Polaris 10 being faster than GTX 980 at lower TDP.

 

[Azix]

AMD mentioned volta in their recent ad. with that chip size and HBM, expecting significantly better than 1080. This is like AMD if AMD had released 290x first instead of 7970 on the high end.

 

[PhonakV30]

Show me a quote from AMD where they say "the performance demoed by AMD is WORST POSSIBLE CASE scenario."

according to AMD, Vega was running with Fiji driver.does that represent Worse scenario ? Who knows.a lot of things in Vega has been changed.

link?

MI25 = 16 GB
Vega = 8 GB

 

[Glo.]

There must be a lot of HPC crap in there that I don't care about if that's big Vega. So 28nm Fiji is 4096 shaders and 4096-bit at ~600mm^2, and 14nm Vega is 4096 shaders and 2048-bit ~550mm^2? A new process and only ~10% smaller size?

It is most likely because big Vega has 8 Shader Engines in the GPU rather than 4 like Fiji had. You need to arrange it somehow.

Small Vega will most likely have 6 SE's with 48 ROPs.

 

[zinfamous]

according to AMD, Vega was running with Fiji driver.does that represent Worse scenario ? Who knows.a lot of things in Vega has been changed.



MI25 = 16 GB
Vega = 8 GB

I know MI25 has been reported to carry 16GB HBM2, but I haven't seen a specific piece from AMD that says a consumer dGPU with Vega will only ever top out at 8 GB. (By the way, MI25 is Vega, so that differentiation is a bit confusing, until we can start calling things 490, 580, 590, whatever )

Vega can go up to 32 GB. It can go into any card that AMD wants to ship for whatever purpose. Obviously it depends on the need/market/cost they are targeting, but where has it been said, specifically, that a consumer Vega card will only go up to 8GB? I've seen it mentioned for nearly a year now that the top Vega dGPU should expect 16gb--again, probably more rumor than anything, but repeated several times; so please point me to a definitive report where the biggest consumer Vega will never have more than 8 GB of HBM2.