Anand talk R580

[Unknown Soldier]

R580 Details Emerge

After speaking with several AIBs today, we have some confirmation that Foxconn already has design kits for ATI's R580 socket. Pipe and clock information for R580 is still not firm, so we will not comment on that at this moment. However, your ATI video card will most likely come with a socket flip chip from ATI. Unlike the NVIDIA socket prototype we commented about yesterday, the R580 socket is geared specifically for a PCIe graphics adaptor, rather than a motherboard-housed GPU socket.

However even with a dedicated socket, graphic card memory would still need to be soldered onto the board. The narrow tolerances of GPU memory would not allow for a DIMM-like solution. ATI's newest R520 GPU already advertises support for a wide range of memory products, including memory that has not even been announced yet.

Read More: http://www.anandtech.com/video/showdoc.aspx?i=2574

 

[soylent]

I would guess, that if this is true, this means we are about to see a big step ahead in memory bandwidth(gddr4 or xdr) or less emphasis on bandwidth intensive things(more reliance on cache and compression?). How else would something like this "work" in practice?

It would be kind of useless unless you could reuse the same board with at least the next generation of GPU. Features alone would make very few people upgrade if performance did not improve at all.

 

[klacik]

R580 Details Emerge

After speaking with several AIBs today, we have some confirmation that Foxconn already has design kits for ATI's R580 socket. Pipe and clock information for R580 is still not firm, so we will not comment on that at this moment. However, your ATI video card will most likely come with a socket flip chip from ATI. Unlike the NVIDIA socket prototype we commented about yesterday, the R580 socket is geared specifically for a PCIe graphics adaptor, rather than a motherboard-housed GPU socket.

However even with a dedicated socket, graphic card memory would still need to be soldered onto the board. The narrow tolerances of GPU memory would not allow for a DIMM-like solution. ATI's newest R520 GPU already advertises support for a wide range of memory products, including memory that has not even been announced yet.

Read More: http://www.anandtech.com/video/showdoc.aspx?i=2574

interesting info.. thanks for posting this..

i'm very interested in the physics acceleration on GPU.. hope it'll be used in the future and will not be another "useless" feature just because ATI came first with it.. i don't like the idea of dedicated cards/chips for physics

 

[Hellbinder]

I would guess, that if this is true, this means we are about to see a big step ahead in memory bandwidth(gddr4 or xdr) or less emphasis on bandwidth intensive things(more reliance on cache and compression?). How else would something like this "work" in practice?

It would be kind of useless unless you could reuse the same board with at least the next generation of GPU. Features alone would make very few people upgrade if performance did not improve at all.

I dont really understand the few comments like this i have read. Pumping in a new GPU with the same memory even 24 months later will benefit you. You will still upgrade features and increase pipelines etc.

 

[Nite_Hawk]

I dont really understand the few comments like this i have read. Pumping in a new GPU with the same memory even 24 months later will benefit you. You will still upgrade features and increase pipelines etc.

True enough. A 9700pro still has as much memory throughput as todays midrange cards. It could be nice to have better shader performance and an updated feature set while retaining the original board.

The other nice aspect of this is that if you already have a reasonable GPU (say a x800) and want to go from AGP to PCIE (or whatever the next standard is), you can bring your current GPU over to the new system. Having socketable GPUs probably would have been really useful 18 months ago when people started switching.

Nite_Hawk

 

[Megadrive1988]

Vendors tell us R580 is some ways away, so don't expect anything between now and CeBit. However, now that ATI's XBOX 360 project is winding down, expect the company to utilize more resources for R580, Crossfire and the Xpress chipset in the coming weeks

this doesn't really make sense to me. we heard that ATI was taking Xbox360 enginneers away in late 2004 for other projects. why would Xbox 360 be winding down *now* for ATI? Xenos is old as far as ATI engineering is concerned. I would imagine that R580 was being designed in parallal with Xenos and R520.

 

[Dave Baumann]

XBOX hasn't necessarily wound down either. Still work to be done.

 

[nelg]

XBOX hasn't necessarily wound down either. Still work to be done.

Can you elaborate?

 

[AlNom]

65nm?

 

[kemosabe]

XBOX hasn't necessarily wound down either. Still work to be done.

Surely you're not referring to anything that requires significant hardware (vs. software) engineering resources? Obviously ATI would want to cooperate closely with TSMC and MSFT to maximize R500 yields, etc, but how much manpower can that require, especially now that Xbox360 is just a month from hitting the shelves?

As for R580, what are they going to do between now and CeBit on a chip that's "in house and working"?

 

[soylent]

I dont really understand the few comments like this i have read. Pumping in a new GPU with the same memory even 24 months later will benefit you. You will still upgrade features and increase pipelines etc.

As far as I understand texturing performance and anti-aliasing is mostly memory bandwidth bound and shaders much less so.

Most generations aren't a big leap in tech. So wouldn't you just get some more shading power and every other generation or so a bunch of cool new features with little improvement in MSAA or AF performance other than that based on increasing efficiency? As long as usage of shaders, especially when it comes to shaders that do a lot of computation on little data, is increasing quickly this is all fine and well of course.

 

[Ailuros]

As for R580, what are they going to do between now and CeBit on a chip that's "in house and working"?

"In house and working" doesn't mean or suggest that it's ready for release right now either. It takes it's time from pre-production prototypes to reach mass-production silicon; even moreso nowadays that chips have become as complex as they have.

 

[SugarCoat]

True enough. A 9700pro still has as much memory throughput as todays midrange cards. It could be nice to have better shader performance and an updated feature set while retaining the original board.

The other nice aspect of this is that if you already have a reasonable GPU (say a x800) and want to go from AGP to PCIE (or whatever the next standard is), you can bring your current GPU over to the new system. Having socketable GPUs probably would have been really useful 18 months ago when people started switching.

Nite_Hawk

Thats assuming they dont adopt an Intel train of thought where anytime they change something on their processors you need to run out and buy a new motherboard for it. Something that will show usefullness for at least 2 years would be incredible, but i doubt it. They're in business to make money, not lose it. More or less what i would expect is a host board to support a few range of cores, another board for the mid range of cores, and yet another for the lower range should costs work out. In the end, i still think we'll be replacing PCB's yearly for major core changes. Maybe even minor ones. All this may allow is that any average joe can buy, say, a Pro core and put it in their PCB, and later if they want to they can buy an XT core to replace the Pro. The goal to be that you just got higher performance without replacing a 400 or 500 dollar card. I dont see someone having the same board for an R580 as the R600, i think that is wishful thinking.

Basically, a main PCB with a socket for each core set per year seems like what they might try.

 

[Zaphod]

However, your ATI video card will most likely come with a socket flip chip from ATI.

Heh. Rendition and Micron tried, with less than astounding success, to launch an open socketed standard for GFX applications way back in 1998. Socket-X, if I remember correctly. Time must indeed be cyclic or something.

 

[Pete]

Most generations aren't a big leap in tech. So wouldn't you just get some more shading power and every other generation or so a bunch of cool new features with little improvement in MSAA or AF performance other than that based on increasing efficiency?

The X1800XL improved on AA performance relative to the similarly-clocked X800XT.

 

[kemosabe]

"In house and working" doesn't mean or suggest that it's ready for release right now either. It takes it's time from pre-production prototypes to reach mass-production silicon; even moreso nowadays that chips have become as complex as they have.

Well it has often been argued that the R520 issues would not delay R580 development. Since R520 was intended to launch in the spring and there is working R580 silicon (suffering from no major reported issues), one is left to assume that the only reason it's not scheduled to launch as per historical timeframes for refreshes is only because of R520's late arrival (not that this doesn't make business sense). I'm betting ATI will be prepared to launch as early as December depending on what NVDA does about its G70 successor, otherwise we would logically have to conclude that R580 is just as late. :neutral:

 

[Geo]

Well it has often been argued that the R520 issues would not delay R580 development. Since R520 was intended to launch in the spring and there is working R580 silicon (suffering from no major reported issues), one is left to assume that the only reason it's not scheduled to launch as per historical timeframes for refreshes is only because of R520's late arrival (not that this doesn't make business sense). I'm betting ATI will be prepared to launch as early as December depending on what NVDA does about its G70 successor, otherwise we would logically have to conclude that R580 is just as late. :neutral:

Oh, I dunno. It seems rather unfair to me to judge R580 "refresh" timing relative to previous launch by the same standards as X850. Plus add in the move towards simultaneous (or nearly) availability, and assume (yes, I know) that a mid-March launch means mid-March availability, rather than a December launch of X850 that had availability when? February? It felt that way anyway.

But at any rate, factor both in and mid-March doesn't look so "delayed".

 

[Sxotty]

Well it has often been argued that the R520 issues would not delay R580 development. Since R520 was intended to launch in the spring and there is working R580 silicon (suffering from no major reported issues), one is left to assume that the only reason it's not scheduled to launch as per historical timeframes for refreshes is only because of R520's late arrival (not that this doesn't make business sense).

I would venture that it makes little business sense actually. If r580 was able to be launched on tuesday then ATI should do so.

The research for the r520 is a sunk cost, they have already spent the money and saying "we should delay the r580 so the r520 can recoup its developments costs" would be idiotic. If ATI does that then they will be paying the price for it at a later time.

If the r580 is the killer everyone seems to imagine then why not bring it out and let it mop up the gtx and rake in the profits?

 

[Jawed]

Let's see how many damn X1800XTs and X1600XTs land in the shops, first. ATI is bullish about quantities.

With a bit of luck the prices will take a serious hit. With a 512MB GTX due soon we'll have "price parity" if not feature parity (nor performance parity, since it's looking like months before ATI sorts performance). The market in the run up to Christmas is going to tell us lots about how aggressively the two sides are going to fight between now and Vista - if there's only one more generation/refresh + refresh (R590,560,540) left to fill what could be another 15 months (though I doubt it).

I have viewed R520 as the gentle intro of R5xx for some time, with R580 being "the real deal". I'm just annoyed everything's slowed up (but more annoyed that the slow-up prolly put the final nail in the coffin of an AGP R520). Ah well.

Jawed

 

[Dave Baumann]

Contrary to some report R580 had taped out even when R520 was going through its numerous spins; R580 did, of course, also come back with exectly the same issue. Having found the problem and rectified it ATI are nowgoing through the motions in a matter of priority order - R520 was fixed first and its final revision is ramped, and should finally squirt out as a product at the end of this month, RV530 is obviously th next priority and the R580 will be last. The fix for R580 will be the same as R520, and it probably won't many other potential issues will have already been known quantities as the quads are brought over from RV530 - I would suspect that R580 is finished by now, or going through its final revisions, it then becomes a question of when it goes into production.