AMD RV770 refresh -> RV790

http://www.tsmc.com/tsmcdotcom/PRLi...il&language=E&newsid=3222&newsdate=2008/11/17 ?

 

And at least according to CustomPC, AMD rep has already confirmed that AMD will bring 40nm gfx chip(s) during Q1

 

More than one? Are you talking about amd or including nvidia as well?

Last i heard amd's RV740 was going for an A12 revision to fix up minor problems. Power saving on A11 hadnt been so great compared to 55. There is a debate at the moment apparently whether to release a gddr5 version as it might exceed 75W and require a power connector which they don't want to do.

Have not heard about any other specific ati chips. The RV870 has got pushed back a few months apparently to make way for lower margin, probably higher volume products. The TSMC 40nm process has "opened up"(or "melted down" depending on which party you were) cause of everybody cancelling orders.

Both ati and nvidia are no longer constrained to choosing low volume products. In nvidias case this also meant delaying their high end chip and starting 40nm with multiple smaller parts. I guess this has forced their hand into releasing the 55nm GT200 and GX2 to fill time.

 

http://translate.google.com/transla...9&ct=result&prev=/search?q=RV775XT&hl=en&sa=G

 

Observations confirm most of what you're saying, but I reckon RV870's delay is not huge, as ATI apparently has little problem moving RV770 stock, RV740 is going to replace the low-end RV770 SKU (presumably) or make a new one, and RV870 still slots higher than RV770.

They'll have trouble moving R700XT/PRO (both X2 counterparts) after the GTX 285/295 arrive, so RV870 shouldn't be delayed that much.

 

it seems that r775pro would have gddr4 memory, if the above come true. I believe that there would be 1gb models too. Though the fact that it would be out in march means that 40nm chips could be delayed at least until june in my opinion.

πηγή

 

???

 

I doubt the RV775 die would be shrinked down in size/transistors, while having extra SIMD's at the same time. Also, why would they equip the RV775 XT with only 512 MB of VRAM?

EDIT: Also noticed that according to the number of TMU's, and by going with the current RV770 architecture, the stream processor figure should be 960.

 

Yep, and it doesn't comply with the rearranged RV730-core either. Yet another redesign or is AMD deviating from it's "one size fits all" route?

 

Try fake info?

 

rv770 is 10*16*5 = 800 and 10*4 = 40 (SP and TUs respectively).
rv730 is 8*8*5 = 320 and 8*4 = 32
So if this is true, it's probably:
12*14*5 = 840 and 12*4 = 48
At least it's symmetric, and maybe a better sp/tu ratio, and if it was so easy to change for rv730, it could be..

And regarding die size I guess they are at least cutting the inter connect.

 

My guess is no refresh is coming, and they will go directly to a new baby @40nm and DX11 compliant before winter ends.

This strategy of super-RV770 is too similar to the super-RV670 leaks strategy to believe it again :>

 

If I recall, ATi's architecture can be improved upon in multiple ways. Either by lengthening the SIMD's or by adding more. And I don't think the texture quads are bound to the number of SIMD arrays.

 

Texture units are bound to SIMD arrays since RV770 again.
This supposed RV775 just doesn't make any sense.
Why would they rebuild the RV770 to get what essentially is the same RV770?
+8 superscalars and +8 TMUs won't give them anything at all from the performance point of view.
Plus they'll need to go with rather strange number of ALUs per multiprocessor (70) while in the past they've always tried to have a multiple of 40 (80x4, 80x10, 40x8, 40x2 etc) and that will give them what exactly? A need to rebuild their mature and fast R6x0+ driver codebase? To get +5% of performance of RV770? That's just stupid.
I'm guessing that RV775 is just a new revision of RV770 meant to combat a new revision of GT200 (GT200b), nothing more.
RV740 and RV870 are the real new chips and both will arrive in Spring'09 on 40nm.

 

Pretty sure you can't have a shader array length of 14, since those correspond to pixel quads (so needs to be divisible by 4).

I doubt that this accounts for much, and given that obviously a major achievement for rv770 was to get higher perf / more units on a given size, I doubt there's a lot of further optimization possible. Also, the use of 1400Mhz memory would mean a return of gddr4 - not saying it's impossible, but all in all I say fake.

 

Bugger. Thought that was a pretty neat aspect of R600.

 

I do recall some talk a while ago regarding the redundancy in RV770 and it actually having something like 840 SPs. Maybe yields have improved to a point where they can now safely enable 40 more SPs without wrecking yields. Well I guess that would make since for the SPs since 840 is only 5% more than 800, but 48 TUs would be a 20% increase...

 

homerdog, if that was the case, it wouldn't be a new chip or a new spin, as it would still be rv770, not rv775.

 

I don't mean to say all they did was reclaim redundancy. I doubt they would build 20% redundancy into the TUs. But is it a possibility that they re-enabled some of the SPs and 'engineered in' some more texturing power? If they dropped the SidePort perhaps they had some die space to fill up. IIRC RV770 was pad limited.

Of course this is all assuming those specs are legit, which from what I gather is a long shot.