AMD to launch 20nm and a brand new architecture for GPUs this year

[3DVagabond]

No way. Titan is a premium product in a market that actually has participants with real purchasing power but it's a limited market. It's cost comes down not only to premium performance, but costs, and satisfactory die yields. A $1000 Titan graphics card with the GPU is one less GPU that doesn't go into a CAD/Workstation or Tesla processing board that sells for much more. On the upside, it provides an outlet for Nvidia's most powerful current GPU for more than GPGPU.

A die shrink to 20 nm would mean such a powerful chip for AMD doesn't have to cost no where near as much, assuming yields are plentiful. AMD would have no need to sell the top end Volcanic Islands for $1000.

When was the last time the top GPU from nVidia or AMD wasn't used in both pro and consumer lines? Trust me, they aren't losing Tesla sales to put gk110 into a Geforce card. There apparently was a time they didn't have enough gk110's to fill Tesla orders and that's why Titan took so long to be released. That hasn't been the case for a while now though. You can pop onto Newegg and buy a K20 right now, if you want. As far as workstations, I don't think they even make a GK110 Quadro, do they?

Forget about the chip costing less so it'll be cheaper. They are pricing them according to performance. Look at GK104. Still more expensive than Tahiti even though it costs less to produce.

Smaller Chip + Less VRAM + Cheaper PCB = Higher Price.

 

[Skurge]

Has AMD ever had a 70-100% performance increase?.....I dont think so...Bit of an OTT statement on Titan being worthless too....sigh, we go through this every time a new GPU is on the horizon....

Is that a serious question?

I hope not.

9700-> X800, X800-> X1800, 3870->4870, 4870->5870

 

[AnandThenMan]

AMD would have no need to sell the top end Volcanic Islands for $1000.

No need? Need meaning profit, sure they have a need and will sell the flagship card for grand if the market will tolerate it.

 

[Rvenger]

AMD tried $550 and got denied

Wrong, They were flying off the shelves until the GTX 680 was released, then it the crickets came to town afterwards.

 

[BallaTheFeared]

No Colonel Sanders, you're wrong!

AMD wasn't able to produce the cards, it's like Titan they might be "sold out" but that doesn't mean a lot are being sold.

 

[Silverforce11]

Is that total marketshare including mobile? Because NV pretty much won 9/10 (or more!) laptop designs this gen. Its an utter loss in that sector for AMD.

Let's assume GCN 2.0 isn't like the switch from VLIW4 to GCN 1.0 for the driver team, and they are more on the ball with this switch. Twice the transistors in the same TDP or die space should in theory at least give 60% or more performance gains. Its going to depend on the TDP targeted.

7970 ghz vs 6970 is a sizeable performance leap, for less die space as well as a bigger focus on compute (which some of you have forgotten):

The gap is wider in newer games than older ones, that will be the key point. ie. 5870 vs 4870 at launch wasn't a huge leap but once newer games pushed the hardware that gap was indeed huge.

GCN 1.0 vs GCN 2.0 should be a fairer comparison in terms of the compute focused architecture and its gaming performance efficiency. Seeing the gains with the 7790 in efficiency should be a clear indication of the kind of performance they can deliver at similar TDPs. An optimist would thus expect >60% for the same power use. Possible? Heck yes, when its a big node jump with a tweaked architecture, anything less would be failing.

 

[toyota]

Is that total marketshare including mobile? Because NV pretty much won 9/10 (or more!) laptop designs this gen. Its an utter loss in that sector for AMD.

Let's assume GCN 2.0 isn't like the switch from VLIW4 to GCN 1.0 for the driver team, and they are more on the ball with this switch. Twice the transistors in the same TDP or die space should in theory at least give 60% or more performance gains. Its going to depend on the TDP targeted.

7970 ghz vs 6970 is a sizeable performance leap, for less die space as well as a bigger focus on compute (which some of you have forgotten):
http://tpucdn.com/reviews/AMD/HD_7990/images/perfrel2_1920.gif

The gap is wider in newer games than older ones, that will be the key point. ie. 5870 vs 4870 at launch wasn't a huge leap but once newer games pushed the hardware that gap was indeed huge.

GCN 1.0 vs GCN 2.0 should be a fairer comparison in terms of the compute focused architecture and its gaming performance efficiency. Seeing the gains with the 7790 in efficiency should be a clear indication of the kind of performance they can deliver at similar TDPs. An optimist would thus expect >60% for the same power use. Possible? Heck yes, when its a big node jump with a tweaked architecture, anything less would be failing.

sorry but 50% improvement for the 7970 ghz over the 6970 is not all that impressive this late. it took well over a year of driver improvements and a bump in clockspeeds to make the 7970 seem like decent jump over the 6970. and the 7970 ghz uses more power than the 6970 to finally make it pull away too.

5870 had a bigger lead than that over the 4870 from day 1 while using the same or less power too.

 

[Will Robinson]

Hmmm...a 20 nano card would nicely go with the new platform I'm building later this year.
I expect AMD won't abandon the $500 customer base and we will find their fastest single GPU card in that price range while they use 2xGPU once again for the ultra performance $1000 class ala 5970,6990 and 7990.

 

[raghu78]

sorry but 50% improvement for the 7970 ghz over the 6970 is not all that impressive this late. it took well over a year of driver improvements and a bump in clockspeeds to make the 7970 seem like decent jump over the 6970. and the 7970 ghz uses more power than the 6970 to finally make it pull away too.

GCN was a new architecture for AMD and took time to polish and extract good performance with drivers. Also HD 6970 was very poor in tesselation and compute performance. GCN is a compute powerhouse. performance in the latest DX11 games in most cases is close to 80% higher on HD 7970 Ghz wrt HD 6970 especially with MSAA. older dx10 and dx9 games show lesser improvement as they depend on raw FLOPS. but newer DX11 games depend on an improved architecture for tesselation performance, compute performance. all these improvements take die area and power.

http://gamegpu.ru/action-/-fps-/-tps/far-cry-3-blood-dragon-test-gpu.html

http://gamegpu.ru/action-/-fps-/-tps/crysis-3-test-gpu/graficheskaya-chast.html

http://gamegpu.ru/action-/-fps-/-tps/tomb-raider-test-gpu-v-2.html

http://gamegpu.ru/action-/-fps-/-tps/battlefield-3-ehd-game-test-gpu.html

GCN 2.0 will be based on the foundations of GCN . so it will have the advantages of a mature architecture. drivers will be mature. It would not be surprising if a HD 8970 on 20nm is 60 - 70% faster than HD 7970 Ghz.

 

[Silverforce11]

sorry but 50% improvement for the 7970 ghz over the 6970 is not all that impressive this late. it took well over a year of driver improvements and a bump in clockspeeds to make the 7970 seem like decent jump over the 6970. and the 7970 ghz uses more power than the 6970 to finally make it pull away too.

5870 had a bigger lead than that over the 4870 from day 1 while using the same or less power too.

As I've said, it depends on the games being benched, with a lot of older titles or less graphically demanding games, the gap is close. 4870 > 5870 is an evolution of similar architectures, thus its expected to have a greater performance leap with a node shrink.

6970 > GCN with its compute focus (revolution) while being a smaller chip, 50% is fine. It's ~60% on that TPU chart.

GCN 1 to version 2 is an evolution, with a node shrink. Drawing similar parallels, for similar die space or TDP, GCN 2.0 should be >60%.

Edit: Also, release 7970 had a massive OC headroom and scaling which the 6970 did not, even at release I recall seeing OC vs OC numbers and it was ~70-75%.

 

[3DVagabond]

Hmmm...a 20 nano card would nicely go with the new platform I'm building later this year.
I expect AMD won't abandon the $500 customer base and we will find their fastest single GPU card in that price range while they use 2xGPU once again for the ultra performance $1000 class ala 5970,6990 and 7990.

I don't think they are pushing to be first on 20nm so they can cut us a deal. 28nm hasn't been a money maker for AMD (I know it's not all because of their graphics division.). They've had to cut prices and offer amazing bundle prices to grab sales. I'm sure they are banking on being first to 20nm so they once again can increase ASP's.

Let's hope this time they can release a more refined product. Although, with them releasing Malta before it's drivers were ready for prime time, there's certainly no guarantee they won't release yet another half baked product (Obviously they wanted it out before nVidia's upcoming refresh and rebrands). I understand them wanting to be first to market, but if they look at nVidia, waiting until they have a more polished product seems to make a bigger impression of quality and value than first to market does.

 

[raghu78]

I don't think they are pushing to be first on 20nm so they can cut us a deal. 28nm hasn't been a money maker for AMD (I know it's not all because of their graphics division.). They've had to cut prices and offer amazing bundle prices to grab sales. I'm sure they are banking on being first to 20nm so they once again can increase ASP's.

Let's hope this time they can release a more refined product. Although, with them releasing Malta before it's drivers were ready for prime time, there's certainly no guarantee they won't release yet another half baked product (Obviously they wanted it out before nVidia's upcoming refresh and rebrands). I understand them wanting to be first to market, but if they look at nVidia, waiting until they have a more polished product seems to make a bigger impression of quality and value than first to market does.

AMD seem to have given up on a 28nm refresh due to their strained financials, most of which is directly attributable to a CPU division which is bleeding the company. also the fact that they are handling two huge product launches in the PS4 and Xbox Next means they had to prioritize their already strained resources.

Bonaire HD 7790 which provides more than 30% performance at the same 85w TDP as HD 7770 proves that AMD GCN is very competitive with Kepler from perf/watt. AMD could have released a HD 8970 with 2560 sp and a improved front end (4 ACE, 4 tesselator engines, 4 rasterizers) and 48 ROPs all on a 384 bit memory bus at 420 - 440 sq mm. the performance per sp of Tahiti is not as good as Bonaire or Pitcairn. AMD could have easily got a 25 - 30% performance improvement by removing bottlenecks and increasing shader count.

If HD 8970 is on 20nm the Titan's performance leadership is over. if its on 28nm still AMD should compete with GTX 780 aka Titan LE as they did in the past (GTX 570 vs HD 6970).

 

[3DVagabond]

AMD seem to have given up on a 28nm refresh due to their strained financials, most of which is directly attributable to a CPU division which is bleeding the company. also the fact that they are handling two huge product launches in the PS4 and Xbox Next means they had to prioritize their already strained resources.

It would have been dumb for them to have released new 28nm models and then replace them a few months down the road with 20nm VI. Also, it doesn't seem like nVidia had anything earth shattering coming. Although, the 700 series will mean they should gain the upper hand throughout the range perf/$. AMD will have to adjust pricing to compensate. It shouldn't be for more than a few months though and then, assuming good execution, AMD will be back in the catbird seat. Possibly for quite some time. nVidia's hand might prove to be a bit weak with only a cut down Titan and reshuffling their old models down a notch. Not much of a refresh, in reality. It'll be cheap R&D wise though.

True they had the console parts, and those were/are large projects. The fact they are planning a new 20nm lineup first though means they haven't put their discrete gpu product on hold because of it.

 

[Pottuvoi]

Good for AMD.

Seems like both Nvidia and AMD had to push back the true next generation GPUs from 2013 to 2014.
So no buy from me for this year from either manufacturer.

 

[boxleitnerb]

I think one thing you can give AMD is that their GPU division does a great of making use of new processes. So I'm not worried at all about them "rushing" to 20nm.

This has nothing to do with AMD but with TMSC. If the process just isn't economically viable yet for 300+ mm2 GPUs, AMD can do nothing about it but wait.

 

[2is]

Yeh except that makes no sense what so ever

It makes perfect sense, particularly if they are getting good yields.

 

[3DVagabond]

Good for AMD.

Seems like both Nvidia and AMD had to push back the true next generation GPUs from 2013 to 2014.
So no buy from me for this year from either manufacturer.

Depends on what you call true next gen. 20nm GCN2.0 should be a "true next gen" improvement. It appears as though they cancelled their 28nm refresh and moved 20nm forward. I don't know whether or not it caught nVidia by surprise, but it sure surprises me.

 

[blastingcap]

Competition is good for consumers. I approve. But I am also nervous about what AMD will try to price it at. All of this assumes the rumors are true, of course, and that it's not paper-launched on Dec. 31 or something.

 

[VulgarDisplay]

Depends on what you call true next gen. 20nm GCN2.0 should be a "true next gen" improvement. It appears as though they cancelled their 28nm refresh and moved 20nm forward. I don't know whether or not it caught nVidia by surprise, but it sure surprises me.

Didn't one of AMD's exec's say they were going to get back into making the GPU game an all out brawl with Nvidia? Constantly fighting for the performance crown.

 

[Rikard]

Competition is good for consumers. I approve. But I am also nervous about what AMD will try to price it at. All of this assumes the rumors are true, of course, and that it's not paper-launched on Dec. 31 or something.

Yes, I am looking forward to all the "It is AMD's fault that I cannot afford a 780!" threads

 

[f1sherman]

NO TAPE OUT YET =

AMD to launch 20nm and a brand new architecture for GPUs sometime next year

/fixed

 

[Silverforce11]

NO TAPE OUT YET =

AMD to launch 20nm and a brand new architecture for GPUs sometime next year

/fixed

Its overly optimistic to expect a launch this year. If TSMC is correct and they are launching 20nm production soon, AMD would have to have a perfect tapeout leading to ramping immediately to fit a Q4 launch.

Likelihood of TSMC succeeding with a new node on time? Almost nil.

 

[f1sherman]

6 months for tape out, and 6 months for setting up production lines

I'm guessing that's 12 months... with A0 silicon!

 

[Ibra]

LOL. AMD's lunatics must be desperate to believe that AMD will launch 20 nm this year. XDR2 memory once again?

 

[raghu78]

Its overly optimistic to expect a launch this year. If TSMC is correct and they are launching 20nm production soon, AMD would have to have a perfect tapeout leading to ramping immediately to fit a Q4 launch.

Likelihood of TSMC succeeding with a new node on time? Almost nil.

we don't know any details. the tapeout might have already happened. TSMC 20nm seems to be ahead of internal targets. don't be surprised to see a 20nm HD 8970 in October or November.

http://www.xbitlabs.com/news/other/..._Is_Ahead_of_Its_Own_20nm_Roadmap_Report.html