AMD - Piledriver chitchat

[DanceswithUnix]

I do wonder if that is a limitation of the process that the GPU can't be as big.

The Interposer is silicon, made using usual foundry techniques but with only a few metal layers (SemiAccurate claim on something like a 65nm process to keep costs down). So while we are used to top of the range GPUs being as big a lump of silicon you can make, now that limit is applied to the interposer. Once you put the HBM memory around the outside, there just won't be space for a massive GPU only space for a pretty big one. It is true that a fast GDDR5 interface uses a lot of space, as driving high current out of a pin doesn't scale like logic does.

That would also explain Nvidia talking up 3D memory (despite actually going down the same 2.5D route that AMD are using) as putting memory on top of the GPU keeps you at the sort of top die sizes that Nvidia like to use.

Edit to add: The SA article is rather good: http://semiaccurate.com/2015/05/19/a...ks-hbm-memory/

[watercooled]

What Nvidia have shown uses an interposer exactly the same as AMD. I think there would be thermal dissipation issues for memory stacked on top of a GPU die consuming hundreds of watts. The memory itself is the 3D part - memory dies stacked on top of one another and an interface die.

I don't think passive interposers are bound by the same reticle sizes as logic dies, but you could be right in that there are limits for interposer size. However I'm not sure if that would be the case here as being forced down in GPU die size would be taking a step backwards and offsetting the benefits of HBM were it a factor.

As I said it could be absolutely nothing.

[scaryjim]

... being forced down in GPU die size would be taking a step backwards and offsetting the benefits of HBM were it a factor. ...

Not if the die-space savings from the simpler memory interface were greater than the die-space loss from being more limited in die-size (although that seems pretty unlikely). Presumably if HBM does require a reduction in die area due to physical constraints of making an interposer, AMD have waited until they are in a position to mitigate that limitation somehow (higher density libraries, perhaps?) - I can't see them bringing a new tech to the market on a substantially compromised product...

[DanceswithUnix]

Well the interposer will be limited by the same absolute upper bound that a silicon process has, but then AMD don't usually go that big so that is probably fine.

From what the SA article says (probably pretty boring if you already read Anandtech, there is a lot in common) an interposer with just a few metal layers is quite cheap to make so the usual yield problems of making something huge shouldn't be so bad.

[Xlucine]

I've just noticed something on the HBM slides: http://images.anandtech.com/doci/926...ctor_575px.jpg

The core die is smaller on the (possibly hypothetical) HBM GPU vs the 290X? It could just be that the images aren't to scale or are plain mocked-up, but if true maybe it implies either a second, smaller HBM GPU or that HBM saves a large amount of die space on the flagship?

The latter seems less likely though; memory IO does take up a large amount of die space but considering the rumoured increase in shader count, an overall reduction in die size seems unlikely.

It's not to scale - the ">70mm" line is half the length of the 110mm line, rather than the ~60% it ought to be. If they were to scale, the whole pcb footprint would only be 56mm square!

[watercooled]

Well spotted, I didn't even think to check that.

[watercooled]

980Ti benchmarks starting to appear? http://videocardz.com/55566/nvidia-g...nce-benchmarks

Not sure if legit obviously.

[CAT-THE-FIFTH]

Seems many on forums are not happy with Apple using the R9 370X(which is basically an HD7770) in the new MBP with retina display. Tried to explain it to people due to better OpenCL support and actual performance in things like Sony Vegas,to no avail I suspect.

Its hilarious that AMD can do no right!!

[Biscuit]

Its actually probably not because of that, its more likely AMD dropped their pants and basically gave the parts to Apple for next to nothing.

[watercooled]

For a much smaller company I'd probably agree that's a likely reason, but it would seem less so for Apple based on their history - they don't really do short-term cost-saving moves, and swapping out a major part for one generation of cost savings would be a pretty short-sighted move. Plus they've been using AMD for the iMac and Mac Pro for some time now, so it's not all that unusual.

In terms of efficiency, Maxwell is good in terms of gaming but don't forget that isn't really the main use for GPUs in this system; as CAT said it's more for professional use and AMD's efficiency is actually pretty good there.

[CAT-THE-FIFTH]

Its actually probably not because of that, its more likely AMD dropped their pants and basically gave the parts to Apple for next to nothing.

I doubt it. More AMD parts support OpenCL 2.0,and Adobe CS has moved over mostly to OpenCL now and there are more AMD certified cards by Adobe than Nvidia ones for the Windows version. Final Cut Pro uses OpenCL too. There are even videos posted in this thread talking about their work toegether.

Nvidia has relatively poor OpenCL performance in comparison,since they are still trying to push CUDA. Examples:

http://images.anandtech.com/graphs/graph8460/67225.png
http://images.anandtech.com/graphs/graph7764/61465.png

It just shows how indoctrinated people are now with all the Nvidia advertising and negative PR from sites,that even a win is made into a negative.

A few benchmarks off a Russian website:

http://www.efxi.ru/file/gtx970_vs_gtx770.files/008.gif

That uses Sony Vegas 13 and PressReleaseProject.veg, OpenGL plugin Red Giant Magic Bullet Looks and OpenCL plug REVisionFX Twixtor. Time is in minutes.

Another OpenCL plugin:

http://www.efxi.ru/file/gtx970_vs_gt....files/001.gif

Red Giant Magic Bullet Denoiser II 1.4

Yet another one in Adobe PP:

http://www.efxi.ru/file/gtx970_vs_gt....files/004.gif

ReelSmart Motion Blur 5.0.3

Another one in Adobe PP:

http://www.efxi.ru/file/gtx970_vs_gt....files/001.gif

Digital Anarchy Beauty Box Video 3.0

Look at where the R9 285 and R9 290 are when it comes to OpenCL.

[CAT-THE-FIFTH]

First full picture of Fiji:

http://www.overclock.net/t/1557597/a...#post_23959158

Its very short but uses two 8 pin PCI-E power connectors.

[Bonebreaker777]

Heard that 390x will be still using Hawaii (even if re-re-branded). Is it true?

Fiji will receive similar designation as the Titan from nVidia.

[watercooled]

We don't know yet, but that's one version of the speculation.

[CAT-THE-FIFTH]

The first test of the A10 7870K:

http://www.computerbase.de/2015-05/a...870k_im_detail

[CAT-THE-FIFTH]

The Inside Story of Lattice-CSC, #1 Supercomputer on the Green500
:

http://www.top500.org/blog/the-insid...-the-green500/