AMD: R9xx Speculation

[Gipsel]

I think it's quite likely these will also be VLIW-4 and have the architecture changes of Cayman (whatever those are...). Why? [..]

Let me add a few arguments against it.

First, VLIW4 is a new architecture which AMD may want to test in a market with the lowest risk for them, the enthusiast segment, where they had no single GPU offer up to now. If it doesn't work so well one may loose the halo, but one doesn't exactly has that effect right now either.

If that is not convincing, I try another fact. The driver enumerates quite some unknown ASICs. As I said already there are currently 6 (six!) with the VLIW5 architecture but only 2 with the VLIW4 one.
After Cypress we have three VLIW5 chips, then two VLIW4 and then again three VLIW5. Granted one of them can be a placeholder like a generic generation enumerator, but it actually fits quite well with Barts, Turks, and Caicos on the discrete side and Sumo, Wrestler, and Trinity on the APU side for the VLIW5 architecture. By the way, only the VLIW4 ASICs are going to be double precision capable.

 

[mczak]

Let me add a few arguments against it.

First, VLIW4 is a new architecture which AMD may want to test in a market with the lowest risk for them, the enthusiast segment, where they had no single GPU offer up to now. If it doesn't work so well one may loose the halo, but one doesn't exactly has that effect right now either.

Hmm I don't think it would really be much of a risk for the Turks/Caicos markets neither, as those aren't really that performance sensitive (in contrast to market Barts is positioned). So if performance isn't quite that good who cares?

If that is not convincing, I try another fact. The driver enumerates quite some unknown ASICs. As I said already there are currently 6 (six!) with the VLIW5 architecture but only 2 with the VLIW4 one.
After Cypress we have three VLIW5 chips, then two VLIW4 and then again three VLIW5. Granted one of them can be a placeholder like a generic generation enumerator, but it actually fits quite well with Barts, Turks, and Caicos on the discrete side and Sumo, Wrestler, and Trinity on the APU side for the VLIW5 architecture.

Ok that is more convincing. I certainly would think APU parts are VLIW5 (they are in the pipeline for quite a while already) but 6 of them is too many... Honestly though I can't see why it would make sense for AMD to continue with VLIW5 parts - if VLIW4 is deemed better/more efficient for the high-end why wouldn't this also apply to low-end?

By the way, only the VLIW4 ASICs are going to be double precision capable.

So AMD is quite stubborn in insisting only high-end needs this...

 

[Mat3]

R600 VS ability was an order of magnitude faster than R580's five Vec5 VS engines...

You mean eight?

 

[ECH]

Is it not better to develop profiles for games using VLIW4 ASIC?

 

[Mintmaster]

You mean eight?

My bad. Yeah, that proves my point even further. R600 definitely used more than one SIMD for vertex shading.

 

[Squilliam]

You know these things?

Barts did not exist on the roadmap in 32nm. Barts in 40nm turned up before the 32nm cancellation.

Ok then. Who is Barts' daddy? Where did Barts come from?

 

[OpenGL guy]

Ok then. Who is Barts' daddy?

Homers, of course!

 

[AlphaWolf]

Homers, of course!

Eat my shorts.

 

[UniversalTruth]

Maybe AMD can push 5970's against GTX580? After all, that card was devised to fight the 512CC GTX380 way back when.

Ok, but I don't think they can do it. Because that will be a lost battle for the R5970. At least because of its so weak tesselation unit and the fact that it will have to fight against a "new" generation card from NVidia.

R6970 should be really good. At least as fast as R5970 in normal games and faster in heavy tesselated areas. It's time for AMD to return that single chip peformance crown. I hope the new architecture is good enough. In other forums like SA the guys predict that Cayman XT will be faster than GTX580 and Cayman Pro being as fast as GTX580.

 

[Lightman]

This quote from Xbitlab article is quite interesting:

A next-generation tessellator from AMD will be implemented in the Cayman GPU (Radeon HD 6900). AMD claims it to be three times as fast as the Cypress’ tessellator. We guess AMD’s engineers will focus on increasing the performance of the tessellator proper in the next implementations, possibly on optimizing the hull shaders. We can expect the future architectures, Southern Islands, Hecatonchires, etc., to bring some innovations into the very design of the tessellation pipeline, like what Nvidia offers in its Fermi architecture where each large array of stream processors has a dedicated tessellator for optimal data threading.

(bolded by me)
Have they been briefed about Cayman and spill little beans here or do they make things up?

Anyway article worth reading!

 

[caveman-jim]

This quote from Xbitlab article is quite interesting:


(bolded by me)
Have they been briefed about Cayman and spill little beans here or do they make things up?

Anyway article worth reading!

If there was a briefing there would likely be an NDA to go with it, so people who had been at the briefing wouldn't be able to confirm or deny the details.

 

[jaredpace]

If there was a briefing there would likely be an NDA to go with it, so people who had been at the briefing wouldn't be able to confirm or deny the details.

The probably mean 2.6x-ish times faster under optimum conditions

 

[ZerazaX]

http://www.inpai.com.cn/doc/hard/137168.htm

Google translate is bad but some interesting tidbits:

-Cayman is a different architecture from Barts (which is mostly like Cypress)
-Cayman is clearly put in the enthusiast GPU range, and there is a "New Strategy"
-Second page goes over a core change, says the ratio of DP to SP went from 5 to 4, so VLIW4 seems pretty much confirmed?

Interesting stuff

 

[Mianca]

Interesting stuff

Most interesting indeed -espcially that slide that (indirectly) states that Cayman is a ~ 360mm² chip - that would be some major win given the rumored performance levels ...

 

[Picao84]

- What new strategy? Where the supposed new strategy starts on that chart is right after R600, so its the CURRENT strategy?
- Where is Antilles? There is Hemlock there, so why not Antilles?

 

[Alexko]

That's Hiroshige Goto's stuff, it originates from here as far as I can tell: http://pc.watch.impress.co.jp/docs/column/kaigai/20101101_403541.html

Don't know what it's doing on a Chinese site.

 

[chavvdarrr]

- What new strategy? Where the supposed new strategy starts on that chart is right after R600, so its the CURRENT strategy?
- Where is Antilles? There is Hemlock there, so why not Antilles?

Do we know if Antilles is 2xCayman?

 

[no-X]

I'd say this slide indicates it...

 

[Kaotik]

Not to mention the drivers said already 2 or 3 months ago that Antilles are using "NI CAYMAN"

 

[Robin B]

What do you think about this, no cayman this year ?

http://translate.google.no/translat...co.jp/docs/column/kaigai/20101101_403541.html