Vega/Navi Rumors (Updated)

[crisium]

So it will be cut the same as Fiji for the lower price part. As we could have guessed.

 

[Bacon1]

There are two GPUs in iMac Pro.

Radeon Vega 56, and Radeon Vega 64. You know perfectly well what those numbers mean.

Interesting thanks, where did you see those models?

 

[Head1985]

Its good that amd cut card reasonable.Not like Nv these days with totaly crippled cutdown cards(GTX1070)
The gap should be 10-15%.But i still want full vega

Edit:Btw cutdown card should still beat GTX1080 if full vega is atleast 15-20% faster than GTX1080.Thats good thing.
Only thing that can prevent it is if AMD uses some slow HBM2 on cutdown card and oc cant be possible.Lets hope both card will have same HBM2.

 

[Karnak]

Interesting thanks, where did you see those models?

Vega GPU

• Radeon Pro Vega 56 graphics processor with 8GB of HBM2 memory
• Configurable to Radeon Pro Vega 64 graphics processor with 16GB of HBM2 memory

https://www.apple.com/imac-pro/specs/

• Radeon Pro Vega
• 11 teraflops single precision
• 22 teraflops half precision
• 8GB or 16GB High Bandwidth Memory
• 400 GB/s memory bandwidth

https://www.apple.com/imac-pro/

 

[Despoiler]

Vega FE is a pro desktop card while the Vega in the iMac Pro is most probably a notebook type Mobile PCI-Express Module (MxM) . Apple could have even gone a step further and integrated the Vega GPU with HBM2 package on the iMac Pro motherboard . Moreover the iMac Pro is launching in December while Radeon Vega FE is launching in late June.

You seriously think they are putting an RX Vega in the iMac? The iMac that is designed for professional workloads? What comes first? The chicken or the egg? Form factor has absolutely nothing to do with anything. Chip, product stack, drivers.

 

[Glo.]

http://videocardz.com/newz/amd-releases-vega-gpu-die-shot

Is it me or is there 8 Shader Engines?

 

[Saylick]

https://videocardz.com/newz/amd-releases-vega-gpu-die-shot
/img snip
Is it me or is there 8 Shader Engines?

Looks like 8 to me.

Also, I wonder how much truth there is to WhyCry's "The only difference from the previous generation Compute Units is that NCU also has the 16-bit processing capability with doubled throughput" statement.

EDIT: Is it just me or does that die shot too basic to be a real die shot? Compared to the Fiji shot below, this one is almost conceptual in look and feel. I could have photoshopped in some rectangles with a color gradient to make it look cool and it would have looked close enough to that image. We can't even tell where the major functional units are located.

 

[T1beriu]

Yeah, it's not really a die shot. I would call it a heavily stylized die layout.

LE: Glo, remove the s in the https image link so the image can load.

 

[Head1985]

http://www.guru3d.com/news-story/amd-radeon-r9-fiji-die-shot-photo.html

 

[Krteq]

Huh, if it really have 8 SEs, then "over 2x peak [geometry] throughput per clock" will finally start to make some sense.

 

[Bacon1]

Here is Fiji's layout:

http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/4

 

[Glo.]

Huh, if it really have 8 SEs, then "over 2x peak [geometry] throughput per clock" will finally start to make some sense.

Over 2x Peak Geometry Throughput was quoted to be in configuration 4 Shader Engines vs 4 Shader Engines. So if Vega has 8 Shader engines that means over 4 times higher Geometry Throughput vs Fiji.

 

[majord]

better shot of Fiji for reference

https://www.flickr.com/photos/130561288@N04/29575201691/in/album-72157650403404920/

 

[ElFenix]

Its good that amd cut card reasonable.Not like Nv these days with totaly crippled cutdown cards(GTX1070)
The gap should be 10-15%.But i still want full vega

this is neither here nor there, but it's pretty obvious desktop 1070 is double cut so that mobile 1070 with the single cut can be clocked lower and maintain performance with its desktop counterpart (and provide plenty of salvage dies for both lines).

/endtangent

 

[Bacon1]

better shot of Fiji for reference

https://www.flickr.com/photos/130561288@N04/29575201691/in/album-72157650403404920/

Was that the guy that ripped his chip in half when trying to remove the heatspreader?

 

[Cookie Monster]

http://videocardz.com/newz/amd-releases-vega-gpu-die-shot

Is it me or is there 8 Shader Engines?

This looks more like an artist's representation of the die but still cool nonetheless.

 

[HurleyBird]

It appears to be a real die shot, you can even see the sort of structures you would expect to be there in the dark areas just barely visible. It's definitely been manipulated though to highlight the functional units in a way that destroys most of the details. It's also possible that they just pasted everything on top of the die shot of a different GPU.

It's interesting that the 4096 shaders still take up so much of the die, despite fiji having the same number yet being made on 28nm. It looks like the NCU is substantially larger than previous generations.

There are also some amateurish mistake, like the darkened line right above the bottom row of functional units.

 

[HurleyBird]

The more I think about it, and look at it, the more I think they just pasted some structures on top of a different die. The compression doesn't help (does anyone have an uncompressed image?) but the dark structures just look subtly off in relation to the highlighted areas, and the darkened line mistake I mentioned feels like something you would get if you copied and pasted something from somewhere else, rather than working within the bounds of the base image.

Easiest way to confirm would be to take previous die shots of AMD GPUs to see if things line up.

 

[Krteq]

Was that the guy that ripped his chip in half when trying to remove the heatspreader?

It seems so

https://www.flickr.com/photos/130561288@N04/28925674133/in/album-72157650403404920/

Anyway, I'm still not convinced about legitimacy of that Vega "die shot".

 

[JDG1980]

Is it me or is there 8 Shader Engines?

The 4 shader engine limit was one of the biggest bottlenecks on Fiji, and Vega was supposed to remove that limit. I know the Linux driver code floating around out there still has the same stats as Fiji (including only 4 SEs), but that never made much sense to me - why would they stick with a design they knew was a flawed compromise? (Raja Koduri admitted that Fiji involved various tradeoffs because they were at the limits of die size for 28nm.)

 

[.vodka]

This looks more like an artist's representation of the die but still cool nonetheless.

That's exactly it. It could be mostly accurate, similar to Polaris' diagram or not.

 

[Mopetar]

Looks like 8 to me.

Also, I wonder how much truth there is to WhyCry's "The only difference from the previous generation Compute Units is that NCU also has the 16-bit processing capability with doubled throughput" statement.

Almost none. You don't go from Polaris that puts off heat comparable to a small sun at 1400 MHz to a
Vega part that has more shaders and runs at or above 1600 MHz without some serious changes under the hood.

One thought that I haven't seen brought up is whether or not AMD sacrificed IPC for speed. They did it with bulldozer so it's not outside the realm of possibility and Vega has the kind of huge bump to clocks that make you wonder what had to die on the table in order to achieve them. So far everyone has assumed at least some minor IPC gains, but what if they traded performance per clock for more speed? Pure speculation on my part, but certainly possible.

 

[tential]

So, what do people think about Vega 11 in 2018? Doesn't that feel strange?

 

[richaron]

So, what do people think about Vega 11 in 2018? Doesn't that feel strange?

Makes sense when they're still milking Polaris. They've had every GPU on the shelves for multiple "generations" since GCN started, I'd be surprised if anyone expected different. Guess that's just AMD protocol for getting their return on investment.

 

[Excessi0n]

One thought that I haven't seen brought up is whether or not AMD sacrificed IPC for speed. They did it with bulldozer so it's not outside the realm of possibility and Vega has the kind of huge bump to clocks that make you wonder what had to die on the table in order to achieve them. So far everyone has assumed at least some minor IPC gains, but what if they traded performance per clock for more speed? Pure speculation on my part, but certainly possible.

AMD explicitly stated that Vega is designed for both increased clockspeed and increased IPC in the Vega architecture preview that they gave to journalists at the start of the year.

Way I see it, AMD will have failed miserably if this thing doesn't at least match the 1080Ti. Since we already know how many TFLOPs it puts out, that means they need to pull off a large increase in actual performance per FLOP. Their great execution with Ryzen leaves me hopeful that they can pull it off, but the leaks and presentations so far seem underwhelming. Hopefully it's just driver issues or slower engineering sample silicon. Fingers crossed...