Anand talk R580

[Pete]

Patrick, oops. Yeah, that's sensible, but you'll have to account for scaling to a 512b DDR memory bus.

Chal, I suppose so.

 

[Hubert]

Imagine a dual socket card with two R580s cores. Would use system memory of course, because there would be no room on PCB for any.

 

[Joe DeFuria]

I think what Dave meant was there's two possible situations for the refresh part. Situation A is R580+GDDR3. Situation B is R580+GDDR4. He indicated Situation A is more likely than Situation B.

No, Dave was saying that given Situation A (R590 + whatever memory) or situation B (R580 + Gddr4), situation B is more likely for the part after R580+GDDR3.

 

[BRiT]

No, Dave was saying that given Situation A (R590 + whatever memory) or situation B (R580 + Gddr4), situation B is more likely for the part after R580+GDDR3.

Oh hrm... I didn't get any reference to R590 in his post at all. Must have been lost in the non-threaded view and non-quick-reply nature of the board... Going back and re-reading it all. I can see it now.

Event A: R580+GDDR3
Event B: R580+GDDR4
Event C: R590+GDDR?

Likely: Event A -and- Event B
Unlikely: Event C

Thanks DeFuria!

 

[YeuEmMaiMai]

we will still have the same problems that we do today since the sockets may change as new generations come out. Rendition thought of this years ago with their vertite series and even had a demo board out.

 

[_xxx_]

Anand has some new info:

http://www.anandtech.com/video/showdoc.aspx?i=2596

the advisory claims that RV535 will utilize an 80nm process.

 

[no-X]

It corresponds to this. So:

RV505 - cheaper RV515 (80nm or HM???)
RV515 - 4-1-1-1, no ring bus, 90nm
RV516 - RV515 from UMC
R520 - 16-1-1-1, 90nm
R521 - R520 from UMC
RV530 - 4-1-3-2
RV531 - RV530 from UMC
RV535 - RV530(?) on 80nm
RV540 - ?
RV560 - 8-1-3-2/1
R580 - 16-1-3-1, 90nm
R590 - R580 on 80nm ?

not sure, if the 8-1-3-x part is RV560 or RV540, Daves guesses are that 8-1-3-x is RV560...

edit: the question is, whether the 80nm parts will be slower or faster, than 90nm parts... Does the 80nm process use low-k?

 

[Dave Baumann]

Get rid of the parts that are incremented by 1, they are erroneous, IMO. RV535 and RV540 could be the same, but that would be a guess.

 

[Arty]

no-X's revised list:

RV505 - cheaper RV515 (80nm or HM???)
RV515 - 4-1-1-1, no ring bus, 90nm
R520 - 16-1-1-1, 90nm
RV530 - 4-1-3-2
RV535/RV540* - RV530 on 80nm
RV560 - 8-1-3-2/1
R580 - 16-1-3-1, 90nm
R590 - R580 on 80nm

 

[Ailuros]

ISV did not have R5xx prototypes in 2004 at all, actually Ati was not able to ship working samples before the soft ground issue was fixed - pre-fix chips was simply to flaky.

From the day where they decide to start volume production to the day they are actually ready to ship, it is about three months. It is very hard to say how long it takes from a “workingâ€￾ prototype until release, if they need a another silicon respin it could take close to six months before they are able to actually ship in volume, but a new metal layer steppings might have very little impact on the time table.

Spring 2004 was an obvious typo; it should have read Spring 2005.

 

[Arty]

There is also another bit of info originating from Anand today; its not just X1300 that doesnt need a master card for CrossFire, X1600 also doesnt need a master card. Anandtech is down so I cant post the link or specifics. :|

edit: http://www.anandtech.com/video/showdoc.aspx?i=2594

Looks like they are confident on this one:

"With the news that Radeon X1600 and X1300 cards will no longer need a mastercard to operate in Crossfire mode, Crossfire boards like the AT8 will certainly become a more viable alternative to SLI in the near future."

http://www.anandtech.com/mb/showdoc.aspx?i=2597&p=1

 

[ANova]

The R580 probably won't need a master card either.

 

[KimB]

The R580 probably won't need a master card either.

Er, how do you expect it to pass the data between the two cards? It can't do it over the PCIe bus, as there's not enough bandwidth for a high-end solution.

 

[SugarCoat]

The R580 probably won't need a master card either.

thought the inq said ATI was doing the RD580 chipset specifically to host that ability?

Er, how do you expect it to pass the data between the two cards? It can't do it over the PCIe bus, as there's not enough bandwidth for a high-end solution.

since the 80 drivers cant Nvidia cards work without the bridge as well? Im quite sure i saw tests and the difference with and without the bridge was negligible. I could be wrong.

 

[Bouncing Zabaglione Bros.]

Er, how do you expect it to pass the data between the two cards? It can't do it over the PCIe bus, as there's not enough bandwidth for a high-end solution.

There may be a bridge, just you won't need special master/slave cards - they'll all be able to interoperate.

 

[KimB]

thought the inq said ATI was doing the RD580 chipset specifically to host that ability?

But it still needs to get the data between the cards somehow.

since the 80 drivers cant Nvidia cards work without the bridge as well? Im quite sure i saw tests and the difference with and without the bridge was negligible. I could be wrong.

Only on low-end SLI hardware (I believe just the 6200 and 6600 non-Pro). It may be less of an issue for the x16 SLI motherboards (where both slots work at x16), but most of the SLI motherboards out there operate the slots in x8 mode when you have two graphics cards plugged in. There just isn't remotely enough bandwidth for SLI across the PCIe bus in this situation.

 

[KimB]

There may be a bridge, just you won't need special master/slave cards - they'll all be able to interoperate.

Well, how is this going to work with motherboards that have different slot spacings?

Are they going to have a cable instead of a bridge? Or ship their cards with multiple-size cables? Or have motherboard manufacturers bundle their bridge for their Crossfire motherboards (thus making Crossfire impossible for any current dual-slot motherboards)?

 

[Bouncing Zabaglione Bros.]

Well, how is this going to work with motherboards that have different slot spacings?

Are they going to have a cable instead of a bridge? Or ship their cards with multiple-size cables? Or have motherboard manufacturers bundle their bridge for their Crossfire motherboards (thus making Crossfire impossible for any current dual-slot motherboards)?

Either or. The point is you won't need to buy a master or slave card - there will only be one type of card and it can act in Crossfire mode with any other of these cards.

I think a cable can be one compromise size. After all, it's unlikely that slots will be more that a few centimeteres apart. I haven't seen any Crossfire or SLI boards that don't have their graphics slots next to each other.

 

[ANova]

Er, how do you expect it to pass the data between the two cards? It can't do it over the PCIe bus, as there's not enough bandwidth for a high-end solution.

In a similar fashion to SLi I'm sure. The new chipsets which support dual x16 lanes will also be available.

 

[no-X]

But it still needs to get the data between the cards somehow.

Er, how do you expect it to pass the data between the two cards? It can't do it over the PCIe bus, as there's not enough bandwidth for a high-end solution.

Why to transfer composited image back to the card? The chipset could have own RAMDAC/TMDS.