New details have come to light regarding the flagship HD 9970 part from AMD.

[WingZero30]

I wonder if there will be a 9990: x2 4096 SP (8192 SP)

Or better Quadfire 9990!!

 

[humbug]

Or better Quadfire 9990!!

32'768 SP's........ Bitcoin that!

 

[Twirlywoos]

No idea how relevant this is, but i'm looking at my Heaven Vally thread.

2x 7970 @ 1250 / 1700 / (2x 2048 SP's) = 4117

1x 7970 @ 1230 / 1794 / (1x 2048 SP's) = 2197

That's 95%+ GPU scalling

This spec also has +100% ROP's (64) / +34% Memory bandwidth (512Bit)

Added on are 16 Parallel Processing Units (no idea what that will do to performance)

And (add Shader Core optimisation?)

Its hard to say, but I am really looking forward to the end of this year....

Where is that AMD driver which is meant to give great Crossfire performance, reduced frame latency and stuttering? I read about it maybe a month ago being released at the same time as the 7990 but has it been released?

 

[Gregster]

Where is that AMD driver which is meant to give great Crossfire performance, reduced frame latency and stuttering? I read about it maybe a month ago being released at the same time as the 7990 but has it been released?

They are testing it (prototype driver) and it gives good results but not due for public release till June/July.

 

[humbug]

They are testing it (prototype driver) and it gives good results but not due for public release till June/July.

That ^^^^

 

[andybird123]

No idea how relevant this is, but i'm looking at my Heaven Vally thread.

2x 7970 @ 1250 / 1700 / (2x 2048 SP's) = 4117

1x 7970 @ 1230 / 1794 / (1x 2048 SP's) = 2197

That's 95%+ GPU scalling

This spec also has +100% ROP's (64) / +34% Memory bandwidth (512Bit)

Added on are 16 Parallel Processing Units (no idea what that will do to performance)

And (add Shader Core optimisation?)

Its hard to say, but I am really looking forward to the end of this year....

2 x 7970's have a total memory bandwidth of 384-bit x 2 = 768
so it is actually a 33% reduction in memory bus

bandwidth lost could be made up for with faster memory though - not sure what is in the pipeline between now and end of the year

 

[humbug]

2 x 7970's have a total memory bandwidth of 384-bit x 2 = 768
so it is actually a 33% reduction in memory bus

bandwidth lost could be made up for with faster memory though - not sure what is in the pipeline between now and end of the year

Does it work like that? does doubling a 384Bit bus from separate GPU's translate into one 768Bit bus?

 

[andybird123]

yes... right now, you have 1 GPU working on frame 1 and GPU 2 working on frame 2
they both get the full 384bit bus to play with while they work (VRAM amount is mirrored so you only get that once, but the bandwidth is dedicated to each chip so it counts twice)

when you move to 1 single GPU, if it were to try and work the same way (e.g. half the chip working on a single frame and the 2nd frame being worked on by the other half) then the whole GPU only gets the bandwidth once, so in this example, each frame only gets half the bandwidth

if that makes sense?

 

[humbug]

yes... right now, you have 1 GPU working on frame 1 and GPU 2 working on frame 2
they both get the full 384bit bus to play with while they work (VRAM amount is mirrored so you only get that once, but the bandwidth is dedicated to each chip so it counts twice)

when you move to 1 single GPU, if it were to try and work the same way (e.g. half the chip working on a single frame and the 2nd frame being worked on by the other half) then the whole GPU only gets the bandwidth once, so in this example, each frame only gets half the bandwidth

if that makes sense?

Yes it does

I did an experiment a while ago on the effects of Memory Bandwidth alone on performance. it turned out to be 0.5 to 1, so 10% increase = 5% is actuality.

 

[Boomstick777]

Here they are saying that AMD's next will still be based on 28nm - http://translate.google.com/transla...-and-hello-volcanic-islands-05122013/&act=url.

Taking all of these leaks with pinch of salt..

 

[Gregster]

Here they are saying that AMD's next will still be based on 28nm - http://translate.google.com/transla...-and-hello-volcanic-islands-05122013/&act=url.

Taking all of these leaks with pinch of salt..

Good find.

Hawaii also will not use the 20nm process, will continue the existing 28 nm process, the limitations of the process specification seems unlikely to do so violent, so what 4096SP what 512bit bus, we can only dream to meet the ... XD. Into the HSA heterogeneous architecture era seems to have been determined, the next step is to wait for AMD will bring out what kind of products to consumers.

Further information or detailed specifications, technical illustrations and data, we will continue to follow-up reports.

I didn't think they were ready to jump to 20nm just yet but was hoping I was wrong.

Maybe the fact that TSMC are working on 20nm got everyone excited but I believe it is solely intended for Apple at the present time and AMD/Nvidia to follow?

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

 

[andybird123]

Here they are saying that AMD's next will still be based on 28nm - http://translate.google.com/transla...-and-hello-volcanic-islands-05122013/&act=url.

Taking all of these leaks with pinch of salt..

much as I expected - internally there is news floating about regarding the next 2 lineups and this is getting merged when it's leaked to the public, much like how GTX780 rumours first came up and everyone was disapointed when the final specs on the 680 were released

 

[bru]

Hmmm the 7970 is already 352mm² so if they are going to stay on 28nm then just how big will the core be?

I reckon if they stay on 28nm then it will be a huge core, if they do go for 20nm then it will be about the same size as the currant chip, but I doubt that it will be available by the end of the year, knowing the usual delays at TSMC.

 

[Marine-RX179]

I reckon if they stay on 28nm then it will be a huge core, if they do go for 20nm then it will be about the same size as the currant chip, but I doubt that it will be available by the end of the year, knowing the usual delays at TSMC.

But there was a new about TSMC being ahead of their own roadmap:
http://www.xbitlabs.com/news/other/..._Is_Ahead_of_Its_Own_20nm_Roadmap_Report.html
...so I guess we could at least get our hopes up?

 

[bru]

Well yes here's to hoping

 

[illuz]

TSMC are 2 months ahead of production of 20nm. I guess this adds a little flame to the rumour, fingers crossed!

 

[humbug]

Here they are saying that AMD's next will still be based on 28nm - http://translate.google.com/translate?sl=auto&tl=en&js=n&prev=_t&hl=en&ie=UTF-8&eotf=1&u=http%3A%2F%2Fchinese.vr-zone.com%2F62871%2Fbyebye-sea-islands-and-hello-volcanic-islands-05122013%2F&act=url.

Taking all of these leaks with pinch of salt..

I very much doubt it will be on 28nm, that would give it a 750mm^2 core.

They use the same slide and came to the same specifications conclusion, and then said blanketley its 28nm, not 20nm.

The slide does not indicate what the process is, but I think most sensible people would and have come to the conclusion that it must be 20nm, for 28nm would be impossible with those specs.

 

[Kaapstad]

Yeah, your probably looking at £500+ But hell if its twice the 7970 on one chip its worth it.

This is the bit that stops me from taking any of these rumours seriously. Even if the GPU is twice as powerful as a HD 7970 it would be unlikely that AMD would simply double up on the HD 7970s specs add a 512bit bus and leave it at that. The only people who are likely to do this are the people who spread rumours.

When this GPU does appear I would expect it to have an entirely new layout compared to a HD 7970.

 

[humbug]

This is the bit that stops me from taking any of these rumours seriously. Even if the GPU is twice as powerful as a HD 7970 it would be unlikely that AMD would simply double up on the HD 7970s specs add a 512bit bus and leave it at that. The only people who are likely to do this are the people who spread rumours.

When this GPU does appear I would expect it to have an entirely new layout compared to a HD 7970.

Like adding 16 Parallel Processing Units?

@ 20nm the DIE would be roughly the same size as the 7970.

Its unbelievable in the performance gain it would achieve, but in terms of technicality its completely believable.

AMD may mess about and make the 20nm with less than 4096 SP's (3000 perhaps?) a smaller DIE size than the 7970.

Or they may just be thinking, "we will not get caught out by Nvidia again, just go all in for it"

The thing is Nvidia have such a huge and dogged fan following, the only way they can gain back much needed market share is just to throw everything in, go maximum and hope its enough to muller whatever Nvidia do and make a clear undeniable difference between them.

Its what I would look to do.

If AMD haven't learnt anything from messing about by now they are idiots.

Time will tell

 

[Kaapstad]

Like adding 16 Parallel Processing Units?

@ 20nm the DIE would be roughly the same size as the 7970.

Its unbelievable in the performance gain it would achieve, but in terms of technicality its completely believable.

AMD may mess about and make the 20nm with less than 4096 SP's (3000 perhaps?) a smaller DIE size than the 7970.

Or they may just be thinking, "we will not get caught out by Nvidia again, just go all in for it"

The thing is Nvidia have such a huge and dogged fan following, the only way they can gain back much needed market share is just to throw everything in, go maximum and hope its enough to muller whatever Nvidia do and make a clear undeniable difference between them.

Its what I would look to do.

If AMD haven't learnt anything from messing about by now they are idiots.

Time will tell

I am not disputing that AMD could come up with something huge (I think and hope for everyones sake that they do) I think if they are going to 20nm a complete redesign is also in order. Anything to do with the HD 7970 in GPU terms is ancient history and things have moved on. I am sure we are going to see an entirely new and probably more efficient layout.