Vega/Navi Rumors (Updated)

[Krteq]

Well, they used a lot of duct tape there

 

[MrTeal]

All those dual GPU cards are far above 300W ...

I can't speak for all of them, but my 5970s and my 7990 don't draw significantly more than 300W at stock clocks. If you tried OCing the 5970s VRM temps would go nuts and efficiency would suffer badly but at stock they were actually quite efficient.

 

[ShintaiDK]

From the link

Still not reading? Let me bold it for you.

None of the fans are blocked on the pictures. And there are 3 of them. Its way better than the average case.

And why would AMD setup something that would throttle

 

[Glo.]

There's no confirmation that Vega will only have 2048bit bus ... only PCGH saying that Vega 10 has to be 2048 because it's 16 GB.

It isn't only SemiAccurate post:

2048 bit memory bus is the configuration for every GPU that has been spawned by Greenland design. It has to be this way, because the CPU itself will also have access to the HBM memory, if AMD would want complete heterogenous unit.

Also note: 512 GB/s bandwidth is achievable only with 2048 bit memory bus.

There is no, 16 GB memory stacks, HBM2 is limited to 32 GB memory size on 4096 bit memory bus.

That is why MI25 is dual Vega GPU, if it has 32 GB of HBM.

 

[HardStyler3]

None of the fans are blocked on the pictures. And there are 3 of them. Its way better than the average case.

And why would AMD setup something that would throttle

to hide its performance from nvidia that they cant react fast enough why would they play with open cards lel

 

[ShintaiDK]

to hide its performance from nvidia that they cant react fast enough why would they play with open cards lel

That's about the last one in the world they can hide performance from. Anyone, even Intel with its own fabs can only hide the price of a product. Zen, Skylake-EP, Cannon Lake, Vega etc is by this time known by the competition. AMD even mentioned unreleased Skylake-EP products at SC16 in a blunder.

 

[Head1985]

I am sure Nv already testing vega anyway.They always know everything about upcoming AMD cards.

 

[nathanddrews]

Vega performs as expected when you look at Polaris.

So you're saying that Vega performs very well? AKA better than OC 1080? Just want you to be clear about what you're saying.

 

[dacostafilipe]

2048 bit memory bus is the configuration for every GPU that has been spawned by Greenland design. It has to be this way, because the CPU itself will also have access to the HBM memory, if AMD would want complete heterogenous unit.

What are you talking about? The CPU access the memory over PCIe or a special interface connected to the GPU, not the HBM directly.

Edit: I'm talking about dGPUs here, not the slide you posted that talks about APUs.

That is why MI25 is dual Vega GPU, if it has 32 GB of HBM.

IF it's a 32GB GPU, it could also be a 4096bit bus ... why not? Remember Fiji?

With the < 300W envelop it makes more sense to be a single GPU board.

 

[Mercennarius]

What's most important is Vega's pricing...if it trades blows with the 1080 while being $150+ less then its a home run for AMD.

 

[Glo.]

What are you talking about? The CPU access the memory over PCIe or a special interface connected to the GPU, not the HBM directly.

Edit: I'm talking about dGPUs here, not the slide you posted that talks about APUs.



IF it's a 32GB GPU, it could also be a 4096bit bus ... why not? Remember Fiji?

With the < 300W envelop it makes more sense to be a single GPU board.

Because its basically the same chip. Greenland was only part of an APU. The CPU needs to have access to HBM2, if it is APU.

Vega is the same family - has been developed after Greenland was developed, and based on the latter chip. Its basically the same chip, but without the CPU.

 

[PhonakV30]

So GTX 1080 will have short Life If Price is lower than GTX 1080.Time to launch new Pascal or Volta.

 

[happy medium]

I think this confirms that the MI25 is big Vega. 4096 shaders = Big Vega

"
Before we dive into the press-release, there's a few things interesting. In terms of the VEGA GPU being used in this platform. The Mi25 is the model that will be fitted with VEGA. It is listed as <300 Watt but that is board power. Secondly AMD has not shared any performance or spec slides on VEGA just yet. However if you dig deeply into the slides you'll notice a twofold of server clusters:

• The K888 G3 with Radeon Instinct offers 4x Radeon Instinct MI2 @ a listed 100 Teraflops
• The Falcon Radeon Instinct cluster with 16x Radeon Instinct MI25 @ 400 Teraflops

Simple math gets us to the mention 25 TFLOPS per GPU. And earlier rumor stated this: "It's mentioned that this GPU has 64 Compute Units, multiply that with 64 shader processors per cluster and you'll get to 4096 shader processors. It will offer up to 24 TFLOP/s 16-bit (half-precision) floating point performance."


http://www.guru3d.com/news-story/am...ct-including-vega-based-mi25-accelerator.html

 

[Piroko]

This thread is derailing once again and as usual it's the same names who do it. One of them is particularly spiteful between his posting history and his sig...

It's fair game for AMD to write "up to 32GB HBM" as long as there's 8-Hi stacks planned by the memory foundries. That does not mean that it'll be released with 32GB if availability isn't there. A 25 TFlop/s single chip FP16 solution is by far the most obvious match for all rumors.

According to PCGH the performance roughly matches 1080 OC models in that preview, which is already totally fine if the price is right.
Now if they can pick up more performance until release then this will start to really put some pressure on the current lineup of Nvidia and that is about the best thing that can happen to us consumers. It might pressure Nvidia into releasing a 1080Ti that is barely cut down from full GP102 at a price that isn't stupid.

 

[MrTeal]

It doesn't seems like Corsair Carbide

Yeah, that case appears to be an Air 540, while the small white one is definitely a 240. The google translation of the hardware.fr article isn't crystal clear, but it sounds like the picture you posted is of a different system doing network training.

Finally, end with a clear Radeon Instinct prototype was already on display last week and was busy train network. however impossible to observe the map more closely than do offer our cliché, AMD having taken care to conceal any opening of the case that getting into.

 

[happy medium]

So GTX 1080 will have short Life.

Well if the Vega 10 is released in May, the gtx1080 will be out a year.
To be honest , I don't think Vega 10 will be much faster than a gtx1080 unless you playing Doom Vulcan @ 4k.

Vega 10 should be competing with Titan XP and the 1080ti, just like Fury x vs the gtx980ti /Titan X.

 

[nathanddrews]

What's most important is Vega's pricing...if it trades blows with the 1080 while being $150+ less then its a home run for AMD.

IIRC, the Fury X and Fury both launched at approximately the same price as their performance competition at the time: the 980Ti and 980, so I wouldn't expect AMD to undercut the 1080 with Vega unless it performs worse. And of course NVIDIA can fire back with the 1080Ti (or 2080?) and drop the 1080's price to undercut Vega as, by that time, they will have the home field advantage (established production, economies of scale, etc.).

Either way, I look forward to the competition.

 

[dacostafilipe]

...

Comparing APUs and dGPUs again ... I'm out here ...

 

[Head1985]

Well if the Vega 10 is released in May, the gtx1080 will be out a year.
To be honest , I don't think Vega 10 will be much faster than a gtx1080 unless you playing Doom Vulcan @ 4k.

Vega 10 should be competing with Titan XP and the 1080ti, just like Fury x vs the gtx980ti /Titan X.

this 4096Sp vega is GP104 competitor/7970 successor.
I really dont know why people expecting 7970 successor matching GP102.That is another tier SKU 400-500mm2.
7970 also compete only vs GK104/GTX680 and not with GTX780TI/Big kepler.
This vega is just similar size as GP104 SKU(if 4096sp is correct)

 

[happy medium]

I guess this would be the Vega lineup vs Nvidia?
What do you guys think?

Vega 10 full chip A.K.A FuryX Vs Titan XP
Cut Vega 10 A.K.A Fury Vs 1080ti 10gb
Full Vega 11 A.K.A 490X vs gtx1080
cut Vega 11/ A.K.A. 490 vs gtx1070

 

[happy medium]

This vega is just similar size as GP104 SKU(if 4096sp is correct)

SO there is a bigger Vega? Vega with more than 4096 sp's?
I haven't seen those rumors.

 

[sirmo]

I was always under the impression that Vega 10 was the smaller, and Vega 11 was bigger. It would make little sense to me if AMD prioritized the bigger GPU first. They need a high end GPU that competes with 1070 and gives 1080 run for its money. With Vega 11 being the enthusiast GPU.

 

[ultima_trev]

Big Vega will compete with GTX 1070 and GTX 1080.
Small Vega will pretty much be a GTX 1060 competitor, replacing the Polaris line.

 

[Despoiler]

None of the fans are blocked on the pictures. And there are 3 of them. Its way better than the average case.

And why would AMD setup something that would throttle

The use of tape, limited details, largely NDA'd event, still beating a 1080 while handicapped. I wonder what the reason could be? This isn't too hard. Why don't you tell us? I know you can do it!

 

[sirmo]

Big Vega will compete with GTX 1070 and GTX 1080.
Small Vega will pretty much be a GTX 1060 competitor, replacing the Polaris line.

I don't see why they need to replace Polaris.. it's been fairly succesful in its price segment. It's not typical for AMD to replace an architecture less than a year since its inception. They have specifically even said that Polaris was mainstream first while Vega was high end and performance.