Support NeoGAF

[artist]

Synapse Design is a company responsible for AMD’s chip floor designs. Synapse recently announced new GPU tapeouts, including two 28HPM silicons (this was discovered at Anandtech forums).

The first one is relatively large chip, 500 sq.mm+. The 28HPM is as powerful as 20nm SoC architecture, so this might be the new Radeon flagship GPU. This processor would run at 1GHz frequency, and according to the source, it will not feature HBM (High-Bandwidth-Memory).

The second GPU is 350sq.mm+ silicon. There is a good chance that this particular chip is actually Tonga. We did confirm that Tonga is almost as big as Tahiti, 350sq.mm sound reasonable.

I think they're implying new GCN GPUs (like Tonga) getting the new features;

UAV Ordering (Unordered Access View Ordering)

UAVs cannot guarantee that each frame will access pixels in the same order. In order to avoid artifacts, some sorting is required. AMD’s UAV Ordering is basically a copy of Intel’s PixelSync. It’s a special algorithm to solve three problems: order-independent transparency, complex scenes anti-aliasing (for example hair or fences) and shadows from transparent effects.

This technology will drastically change the complexity of the scene in newest games.

A good example of such algorithm being used is shown below.

Fast Conservative Rasterization

While UAV ordering is just a software algorithm, Fast Conservative Rasterization will be a hardware upgrade. Conservative Rasterization is nothing new. However current implementation is slow, and everyone who ever tried TressFX knows that very well. Disabling TressFX makes hair strands thicker. Since this technology is also using order-independent transparency, new features will considerably speedup the process, thus allowing the scene to be rendered faster and more complex.

http://videocardz.com/51021/amd-gcn-update-iceland-tonga-hawaii-xtx

 

[Bomber Bob]

Gamecube GPU?

 

[MysticX]

Gamecube GPU?

hehe was thinking the same thing

 

[Persona7]

Gamecube GPU?

http://www.amd.com/en-us/innovations/software-technologies/gcn

 

[gngf123]

Gamecube GPU?

Graphics Core Next

 

[dragonelite]

Dam was expecting we would get High bandwidth memory(Hybrib memory cube(?)) this year.

 

[mrklaw]

Same size at different process size sounds like more power = good news (the 350mm one)

The 500mm one is a bit wtf. And at the equivalent of 20nm? Would there be any precedent for that?

Edit: Titan is 551mm, r9290X is 438. So it is very large but not without precedent. Should pack a punch though.

 

[Dredd97]

Same size at different process size sounds like more power = good news (the 350mm one)

The 500mm one is a bit wtf. And at the equivalent of 20nm? Would there be any precedent for that?

Edit: Titan is 551mm, r9290X is 438. So it is very large but not without precedent. Should pack a punch though.

I think clearly they see a future in what Mark Cerny called 'fine grain compute' and the GCN is a big part of that, as their CPUs don't stack up to those if intel, they see a future in the gpu doing all the work the CPU traditionally did...

 

[belmonkey]

I think clearly they see a future in what Mark Cerny called 'fine grain compute' and the GCN is a big part of that, as their CPUs don't stack up to those if intel, they see a future in the gpu doing all the work the CPU traditionally did...

In such a future, would that at least result in simply needing to add a new GPU to a home PC to make it a gaming machine?

 

[artist]

Same size at different process size sounds like more power = good news (the 350mm one)

The 500mm one is a bit wtf. And at the equivalent of 20nm? Would there be any precedent for that?

Edit: Titan is 551mm, r9290X is 438. So it is very large but not without precedent. Should pack a punch though.

It would be the biggest GPU that AMD has built by a big margin.

Hawaii ~438mm2 (4% increase over previous biggest AMD gpu which was R600)

If this is true then it would mean a 15% increase over the previous biggest AMD gpu. And AMD is averse to big dies, that's not their mojo. (but I could see them returning to the strategy since the process nodes have slowed down)

 

[FLAguy954]

I can't wait for the Tonga GPU. It's definitely gonna be my next go-to GPU (if they can price it @ $250 or less).

 

[SneakyStephan]

What is the 28HPM thing?
500mm^2 die is good news for us, (assuming it doesn't release at geforce titan price)

gtx 580 was a 520mm^2 die and a proper high end card
With both amd and nvidia going with pretty small die midrange cards (gk104 etc) to fill their 'high end' bracket prices just skyrocketed

IF they make this their high end card then nvidia will be forced to actually compete again and we might get away from this doubling of prices from the past 3 years
I really hope the 350mm2 card gets the r9 280 sticker and a fitting midrange price

 

[Renekton]

What is the 28HPM thing?

One of TSMC's 28nm processes, seems to be for mobile devices.

 

[SneakyStephan]

One of TSMC's 28nm processes, seems to be for mobile devices.

But the article is talking about a 350 and a 500mm^2 die for desktop gpus?

 

[Sir_Crocodile]

TSMC are taking so long for their die shrink to become usable to amd, going big is pretty much amd's only option for a significant power boost.

 

[Renekton]

But the article is talking about a 350 and a 500mm^2 die for desktop gpus?

Don't look at me lol, doesn't make sense to me either.

http://www.tsmc.com/english/dedicatedFoundry/technology/28nm.htm

 

[CTLance]

Interesting.

Random thought of the day: Man, I don't want to know the yield on a gargantuan chip like that. Since yield affects prices...

 

[artist]

One of TSMC's 28nm processes, seems to be for mobile devices.

And the bone ..

 

[KojiKnight]

TSMC are taking so long for their die shrink to become usable to amd, going big is pretty much amd's only option for a significant power boost.

This. There really are only 2 ways to get more power, and that's to shrink the process or grow the die. Growing the die has it's own risks (more power draw, more heat, more chance for flaws to significantly negatively impact runs, more expensive in general), but if the die shrinks aren't ready, it's this or be left behind.

 

[Locuza]

And the bone ..

I think this is wrong after all:
http://www.hotchips.org/wp-content/uploads/hc_archives/hc25/HC25.10-SoC1-epub/HC25.26.121-fixed-%20XB1%2020130826gnn.pdf

There was an Interview with AMD about Hawaii, someone asked if Hawaii is build on HP or HPM, since HPM should be more efficient, but AMD statet HPM is only for mobile.
So maybe HPM is really not well suited for High-Performance Components at that level.

 

[SneakyStephan]

Interesting.

Random thought of the day: Man, I don't want to know the yield on a gargantuan chip like that. Since yield affects prices...

Gtx 580 was known to have atrocious yields ( also had a 520mm die) and the production cost of the card was 120 dollars (can't remember if that included the cooling or not or if it was just the pcb+vram+die)

What did higher yields with much smaller dies get us last gpu gen? A big price hike

 

[mrklaw]

This. There really are only 2 ways to get more power, and that's to shrink the process or grow the die. Growing the die has it's own risks (more power draw, more heat, more chance for flaws to significantly negatively impact runs, more expensive in general), but if the die shrinks aren't ready, it's this or be left behind.

Where do nvidia go though? Titan is already > 500mm, they can't go much bigger. This could be good if AMD can pull it off - they are pretty competitive at 340mm, so at 500 they should have a good performance advantage. And then when die shrinks do come, they can cut price/power consumption/heat etc

 

[artist]

I think this is wrong after all:
http://www.hotchips.org/wp-content/uploads/hc_archives/hc25/HC25.10-SoC1-epub/HC25.26.121-fixed-%20XB1%2020130826gnn.pdf

There was an Interview with AMD about Hawaii, someone asked if Hawaii is build on HP or HPM, since HPM should be more efficient, but AMD statet HPM is only for mobile.
So maybe HPM is really not well suited for High-Performance Components at that level.

Hmm, let me see if I can find that interview.

Where do nvidia go though? Titan is already > 500mm, they can't go much bigger. This could be good if AMD can pull it off - they are pretty competitive at 340mm, so at 500 they should have a good performance advantage. And then when die shrinks do come, they can cut price/power consumption/heat etc

Say hello to Maxwell, look at GM107.

 

[Dredd97]

In such a future, would that at least result in simply needing to add a new GPU to a home PC to make it a gaming machine?

Yes, that's true, but AMD do love their APU processors so I guess this is that taken to an extreme maybe?

 

[Locuza]

Hmm, let me see if I can find that interview.

Sorry, it wasn't an Interview, it was an AMA-Thread:
http://www.tomshardware.co.uk/forum/id-1863987/official-amd-radeon-representatives/page-2.html#

Somewhere it should be.

Edit:
Found it:
http://www.tomshardware.co.uk/forum/id-1863987/official-amd-radeon-representatives/page-4.html#11878925

 

[artist]

Sorry, it wasn't an Interview, it was an AMA-Thread:
http://www.tomshardware.co.uk/forum/id-1863987/official-amd-radeon-representatives/page-2.html#

Somewhere it should be.

Edit:
Found it:
http://www.tomshardware.co.uk/forum/id-1863987/official-amd-radeon-representatives/page-4.html#11878925

Thanks.

 

[McHuj]

There was an Interview with AMD about Hawaii, someone asked if Hawaii is build on HP or HPM, since HPM should be more efficient, but AMD statet HPM is only for mobile.
So maybe HPM is really not well suited for High-Performance Components at that level.

High performance usually means very high clock. GPU's don't have very high clocks. Qualcomm's snapdragon manages 2+ GHz on 28nm HPM. Xbone manages 1.75 GHz for it's CPU.

A 800MHZ-1GHz GPU would probably do fine on that process and it may be necessary as for a very large chip the leakage power would be significant (and HPM tries to address that specifically).