AMD: R9xx Speculation

[Jawed]

Hecatoncheires might be even further out. I think I already said after Computex that S.I. comes after N.I.

One is a shrink
One is a hybrid
One is a new architecture

Is MCM/side-port based architecture still in play?

If so, that would be "new architecture".

If one was cancelled then it seems it was the shrink (along with 32nm). In my view that would then leave the hybrid and new architecture.

The hybrid has the ALU changes that we've been hearing more about lately, i.e. VLIW-4 with no dedicated T. That's "relatively simple". SIMDs get smaller but there are more of them. e.g. 30 SIMDs, 48 ROPs, 384-bit. Perhaps this chip is "Evergreen completed".

The new architecture, if MCM/side-port is important, is a bigger change. MCM/side-port might also be argued to be dependent on memory system architecture changes. Changes that I reckon are needed for good tessellation performance, and better support for non-texture reads/writes.

If "new architecture" doesn't have anything to do with MCM/side-port, it still sounds like more work than changing the ALUs. I still think setup-rasterisation needs an overhaul.

 

[Silent_Buddha]

So you guys are suggesting that AMD's margin on CPUs is actually currently lower than their GPUs? That's a first. What does a Cypress chip go for?

It wouldn't be all that surprising. When you have 4 core Phenoms (and salvage 3 core parts) selling for ~100 USD and lower in bulk that isn't going to do much for your margins. Especially considering, I believe, Phenom is still larger than Core iX chips.

I don't think AMD currently has a direct or semi-direct competitor chip to Intel in any segment that isn't larger and thus costlier (presumeably) to produce. They've also been losing share in the server/workstation space to Intel which is going to erode magins also.

It also wouldn't surprise me if margins for Evergreen products are quite high for consumer chips. Especially the 5870 and 5850.

As for Nvidia, perhaps Fermi is finding a lot of traction in the HPC and professional markets. If they shift most of the allocations for Fermi to those segments that's going to boost their margins quite nicely.

And as someone mentioned, them basically leaving the chipset market is going to help overall margins. While AMD absorbing more of the chipset business will erode their own margins, although I'm sure the increased revenue is probably worth it.

Regards,
SB

 

[eastmen]

I wonder if they are planning on improving crossfire scaling. It seems to scale much worse than sli.

 

[Psycho]

I wonder if they are planning on improving crossfire scaling. It seems to scale much worse than sli.

Links?
And not those *only* comparing 1gb cards to 1.5gb cards in 2560x1600x8 - could very well be frame buffer limitations.
Look here for instance: http://www.hardware.info/nl-NL/articles/amdnampoZGCa/Clash_of_the_Titans_3way_SLI_GTX_480_test/13 Notice how the CF scaling sometimes goes down when resolution goes up - normally we would expect the opposite (less cpu limit), so it's probably framebuffer limited (unless it's actually CF-bridge bandwidth limited?)

 

[Kaotik]

I wonder if they are planning on improving crossfire scaling. It seems to scale much worse than sli.

Latest comparison I've seen put 2-way CFX to around 71-72% scaling, 2-way SLI to 78%. However, 3-way CFX scaled a lot better than 3-way SLI on that review (something like 29% over 2-way SLI vs 49 or 59 or something over 2-way CFX)

 

[DuckThor Evil]

Latest comparison I've seen put 2-way CFX to around 71-72% scaling, 2-way SLI to 78%. However, 3-way CFX scaled a lot better than 3-way SLI on that review (something like 29% over 2-way SLI vs 49 or 59 or something over 2-way CFX)

That doesn't sound right. It implies that 3-way crossfire scales perfectly or even better to 2-way lol. Link?

 

[Kaotik]

That doesn't sound right. It implies that 3-way crossfire scales perfectly or even better to 2-way lol. Link?

http://www.hardware.info/nl-NL/articles/amdnampoZGCa/Clash_of_the_Titans_3way_SLI_GTX_480_test/13

(Yes, the drivers are outdated)
And I remembered the percents a bit wrong, here's the correct row:
GTX480 > GTX480 SLI: 78.7% scaling
GTX480 > GTX480 3-way SLI: 128.2% scaling
GTX480 SLI > GTX480 3-way SLI: 27.2% scaling

HD5870 > HD5870 CFX: 70.9% scaling
HD5870 > HD5870 3-way CFX: 151.8% scaling
HD5870 CFX > HD5870 3-way CFX: 46.9% scaling

edit:
http://www.benchmarkextreme.com/Articles/GTX 480 Radeon 5970 CPU Bottleneck/P1.html
There's newer drivers for nVidia, which indicate around same scaling numbers as the older drivers

 

[DuckThor Evil]

Wow those Ati scaling numbers are absolutely increbible. Thanks for the link.

 

[fbuffer]

http://www.hardware.info/nl-NL/articles/amdnampoZGCa/Clash_of_the_Titans_3way_SLI_GTX_480_test/13

(Yes, the drivers are outdated)
And I remembered the percents a bit wrong, here's the correct row:
GTX480 > GTX480 SLI: 78.7% scaling
GTX480 > GTX480 3-way SLI: 128.2% scaling
GTX480 SLI > GTX480 3-way SLI: 27.2% scaling

HD5870 > HD5870 CFX: 70.9% scaling
HD5870 > HD5870 3-way CFX: 151.8% scaling
HD5870 CFX > HD5870 3-way CFX: 46.9% scaling

edit:
http://www.benchmarkextreme.com/Articles/GTX 480 Radeon 5970 CPU Bottleneck/P1.html
There's newer drivers for nVidia, which indicate around same scaling numbers as the older drivers

Pretty close to the same percentages from Xbits recent review : Four Radeon HD 5870 vs Three GeForce GTX 480

summary of 1st half of benchmarks are posted here and here.. out of about 70 different benches ATI had better scaling in 60ish with NV better in 10 or so.

 

[eastmen]

well i stand corrected. Thank you guys. But still lets get closer to a 100% per card added

 

[ECH]

Pretty close to the same percentages from Xbits recent review : Four Radeon HD 5870 vs Three GeForce GTX 480

summary of 1st half of benchmarks are posted here and here.. out of about 70 different benches ATI had better scaling in 60ish with NV better in 10 or so.

They have a tendency to use super sampling for Adaptive transparency for ATI while only using mulitsample transparent AA for nvidia in their reviews. Check out this review's testbed. Therefore, would expect lower frame rate results on CF and CFX. So if their review mirror each other I have to ask what he used for control panel settings because they shouldn't be the same.

 

[Gipsel]

And NI gets int24 multiplication in the ALUs, not only uint24 like Evergreen (which is still not usable from IL, isn't it?).

@Jawed:
I just had a short look on the Cat 10.7 and sometimes small miracles appear to happen. There are now 3 new IL instructions:
IL_OP_FMA, IL_OP_U_MUL24, and IL_OP_U_MAD24 (didn't found the text equivalent of fma, but for the 24bit instructions they are umul24 and umad24), the ones supported by Evergreen (NI will get also get int24 versions). Currently I have no possibility to check if they are working, but they are there (I'm just looking at the driver files). Let's hope Cat10.8 brings support for the still missing IL_OP_U_MUL24_HIGH instruction.

 

[Jawed]

@Jawed:
I just had a short look on the Cat 10.7 and sometimes small miracles appear to happen. There are now 3 new IL instructions:
IL_OP_FMA, IL_OP_U_MUL24, and IL_OP_U_MAD24 (didn't found the text equivalent of fma, but for the 24bit instructions they are umul24 and umad24), the ones supported by Evergreen (NI will get also get int24 versions). Currently I have no possibility to check if they are working, but they are there (I'm just looking at the driver files). Let's hope Cat10.8 brings support for the still missing IL_OP_U_MUL24_HIGH instruction.

Ooh, tasty.

One of the things I'm curious to play with is the FMA instruction. Because it can only issue to 4 of the 5 lanes it slightly simplifies the compiler's task when dealing with very high ILP (e.g. in matrix multiplication where 64-way ILP is normal). In theory this could lead to lower register allocation and more streamlined operation scheduling.

I wish AMD would allow GSA/SKA to compile against the installed driver as well as the driver history that it contains. The months of disconnect between the drivers and GSA/SKA is really pretty awkward.

 

[Jaaanosik]

So which GPU will be the first on 28nm?

Globalfoundries May Deliver First 28nm Chips in the First Half of 2011

Unfortunately, it is unknown which graphics processors will be made in Dresden, Germany, but manufacturing of rather complex graphics chips using thin 28nm-HP process will be a benefit for ATI, provided that it is indeed among the first clients for the node. What is unlikely is that AMD starts to use a brand-new HKMG fabrication technology for a wide range of products and especially very complex GPUs for high-end market.

 

[Squilliam]

So which GPU will be the first on 28nm?

Globalfoundries May Deliver First 28nm Chips in the First Half of 2011

Hmm probably the laptop parts given their higher ASPs for a given performance level and much tighter power requirements. It'll also fit in nicely with the back to school Laptops July (Q3) in they can ramp up production in time and get samples out to laptop makers (again on time).

 

[Robert Varga]

So which GPU will be the first on 28nm?

Globalfoundries May Deliver First 28nm Chips in the First Half of 2011

/

...

 

[rpg.314]

So which GPU will be the first on 28nm?

Globalfoundries May Deliver First 28nm Chips in the First Half of 2011

IMHO, it's too close to call.

 

[Alexko]

Should we start to take bets?

 

[AlexV]

Should we start to take bets?

On whether or not Girlfriend Foundries actually delivers in something other than ppts/pdfs? Yes please.

 

[Squilliam]

Hypothetically speaking, if they release on 40nm in 2010 for their R9xx chips, would they do a 28nm pipe cleaner r9xx next year and follow with R10xx on 28nm or would they migrate R9xx onto 28nm and keep the same model names?