Hardware rumour mill gossip

[kemosabe]

I know it's been rehashed on other threads, but anyone know of a compelling reason why a 24-pipe R520 spring release is out of the question?

We've been led to believe that R420 was just a stop-gap solution, but it seems most people still don't expect to see a new architecture from ATI before the end of 2005.

 

[digitalwanderer]

Not true, I'm expecting an R5xx part early next year.

 

[Polarbear53]

Hmmm, with the time frame, makes it sound like 32 pipelines for xbox 2 might just be feasible. One question though, if a chip is basicly the same throughout, do twice as many pipelines equal ~ twice the performance?

 

[BRiT]

*cough*april*cough*

 

[Luminescent]

Hmmm, with the time frame, makes it sound like 32 pipelines for xbox 2 might just be feasible. One question though, if a chip is basicly the same throughout, do twice as many pipelines equal ~ twice the performance?

All else equal, double the pipes would mean double the throughput in situations where application performance is determined by the time it takes to complete operations corresponding to those pipes.

 

[KimB]

One question though, if a chip is basicly the same throughout, do twice as many pipelines equal ~ twice the performance?

It can, but it needs to be programmed for. To get twice the performance you would need to not be memory bandwidth bound, and that means using lots of math ops in proportion to the amount of data that needs to be input (texture reads) and output.

In a console, I definitely can see the wisdom of shooting for a massively parallel design without a huge amount of bandwidth, a design that would quite possibly be killed in the desktop space. In a console, after all, the programmers will be able to find ways around such limitations, as they need only develop for that one platform.

 

[Sabastian]

EDIT:Bah, never mind. I should go lay down now.

 

[nutball]

*cough*april*cough*

April launch, availability ... July/August then?

 

[PatrickL]

If ATI has a scheduled release for R520 by the end of 05 Q2, i wonder if they will release a R480 before. I don't think it is needed and that could just add more confusion without real benefit.
The only reason for them to do that would be they found a way to produce same R420 at a lower cost for them.

 

[KimB]

If ATI has a scheduled release for R520 by the end of 05 Q2, i wonder if they will release a R480 before. I don't think it is needed and that could just add more confusion without real benefit.

Except that I would expect nVidia to be releasing a new high-end GeForce 6800-based product within, oh, six months (probably dubbed the 6900).

 

[kenneth9265_3]

If ATI has a scheduled release for R520 by the end of 05 Q2, i wonder if they will release a R480 before. I don't think it is needed and that could just add more confusion without real benefit.

Except that I would expect nVidia to be releasing a new high-end GeForce 6800-based product within, oh, six months (probably dubbed the 6900).

That sounds logical....

 

[LeStoffer]

We've been led to believe that R420 was just a stop-gap solution, but it seems most people still don't expect to see a new architecture from ATI before the end of 2005.

ATI have more than suggested that they want a 0.09 process before SM 3.0 is affordable enough to ship.

So the questions are:
1) When is 0.09 ready for a huge 200-300 (+/-) million transitor chip at TMSC?
2) Will ATI jump into a brand new process with their highend chip first (which have been against their policy so far)?

Q3 2005.

 

[Dave Baumann]

Best guage to when they expect it is when they anoounce R480.

 

[overclocked]

Btw has IBM something similar to 110nm/and or are they using their 90nm process just for their own CPU´s etz?

 

[PatrickL]

If they announce a R480 chip, there should be at least six months before R520 i guess. I doubt AIB want to launch new cards every 3 months.

 

[kemosabe]

What time frame would they use for R480 if not the pre-Christmas retail season? Wasn't it Huddy himself who insinuated in that leaked developer presentation that they would have 512Mb boards this year? Don't reckon they'd just slap more memory onto the current X800 cards, so unless the roadmap has changed I'd speculate a refresh in late fall, which would put R520 late spring/early summer 2005. Are there no further refinements (i.e. low-k, etc) to TSMC's 0.11u process that might allow for R520 to be fabbed at that node? If NV40 was feasible at 0.13u (albeit we know little about whether IBM just absorbed any eventual excessive costs), is it unreasonable to expect that 0.11u will have improved/matured enough to allow ATI to deliver a SM3.0+ flagship part in nine months? Everything I've read about 0.09u suggests it might not be ready for prime time in 2005 at any fab.

 

[karlotta]

my best guess is preview around Oct 5th, onsale two weeks after. Speed bumbed 512mb, the X700 is next week....
R520 is .09 and is a febuary preview.

 

[kemosabe]

Looks like Samsung is getting closer to delivering 700MHz GDDR3, and I would assume the fall refreshes will need it.

My wild guess would be R480 (~575/650; 256/512Mb; MSRP $499/599) launch next month, availability in mid-November; R520 announcement at CeBit and availability in May/June.

But what happens to ATI's high-end lineup when R480 shows up? The decision to provide the vanilla XT (500/500) is likely allowing for better yields of 16-pipe R420 cores, so I would expect this part to drop to $399 (using 2ns GDDR3) to compete with the 6800GT and the XT PE to suffer a premature death. X800 Pro drops to $349.

 

[karlotta]

OY, what happened to Micron.. my "buddy" up in Boise was all over the 800+gddr3.. and that was last spring . He doesnt talk to anybody anymore after the PE release.... but i havent seen any Micron Ram for any card? Is it all going to the Xbox?

 

[kemosabe]

Darn good question, especially since Micron made a fuss about having co-developed the spec with ATI. I haven't heard a peep about whether they are actually producing any GDDR3 memory or who it's going to. Maybe something went sour with their ramp and this might help explain the high-end availability issues at the IHVs.