AMD RV790 & RV740 in 40nm in March 2009

[v_rr]

It looks like that RV740 and RV790 will be produced in 40nm and that the launch will be in March 2009.

http://forums.vr-zone.com/showthread.php?t=378647

[Shadowmage]

So did RV870 get delayed then?

[waver_01]

So did RV870 get delayed then?

delayed from what?

[largon]

The question is, whether RV790 is just a shrink or shrink with expand.
RV770 @ 40nm would push it down to around 190mm². Now, expand it some 40% and we'll have a chip that's the same size as RV770 but with 52TMUs and 220SPs (1100ALUs).

[roofsniper]

ati has been making two generations of cards off one chip for awhile now. hd2900 and 3870 are based off the same chip. 4870 and this which will be the 4900 will be based off the same chip. they are just fine tuning it adding a few more sps a die shrink and most likely a core and ram frequency boost. rv870 is set one year after rv770 and will be released in june 2009. nothing was delayed.

[Caveman787]

Lets just call it a stop gap just like with nvidia's gtx260 55nm and 280 55nm.

[Shadowmage]

delayed from what?

I heard that RV870 was going to be released Q2 2009.

Maybe the source confused RV790 with RV870?

[Shadowmage]

One interesting thing is that RV770 is very close to being pad-limited. Since pads don't scale as much as logic in a process shrink, one can assume that RV790 will likely feature more SPs or something else to prevent the chip from becoming pad limited.

However, they could also fix this by removing VDD/GND pads since the chip won't need as much power.

[largon]

Then again, all this silence and odd timing could be yet another cunning serving of Sandbagging à la AMD.

[spursindonesia]

Hmm, this is quite logical, since it can match nVidia toe to toe with their supposed upcoming GT212 cards, RV870 will be prepared to duel with GT300 family cards later on.

[jaredpace]

I heard that RV870 was going to be released Q2 2009.

Maybe the source confused RV790 with RV870?

It's still on time AFAIK.

RV790 March 09 (Q1)

RV870 June 09 (Q2)

[Brodholm]

Didn't they say that the was going to release the upgraded versions to Xmas first?

[roofsniper]

Didn't they say that the was going to release the upgraded versions to Xmas first?

well it would make sense. they said they would release a new card once a year and every 6 months after it is released they would make an upgraded one. 4870 was released at the end of june which would make sense with it at the end of December. so maybe that got delayed or who knows. maybe they are just saying 6 months after 4870x2 is released. which would be around march.

[ryboto]

Lets just call it a stop gap just like with nvidia's gtx260 55nm and 280 55nm.

Lets not call it a stopgap. Ati has been pretty consistent with process switches. I think it goes back as far as the R500 series. X1950pro was a 80nm shrink from 90nm(I think) around Q4/Q1, then there was R600 80nm Around Q2/Q3, then RV670 around Q4/Q1, Rv770 around Q2/Q3, and now we have RV790/740 around Q4/Q1. Granted, all of these have depended on process technology availability. R800 will come probably around Q2/Q3. No stop-gap, just predictable consistent release schedules.

[Helmore]

The question is, whether RV790 is just a shrink or shrink with expand.
RV770 @ 40nm would push it down to around 190mm². Now, expand it some 40% and we'll have a chip that's the same size as RV770 but with 52TMUs and 220SPs (1100ALUs).

RV770 on 40 nm. would be closer to 120 mm^2. We are talking about a die shrink from 55 nm --> 45 nm --> 40 nm. here. That would require a 128-bit memory bus and would thus be a good candidate for what RV740 is going to be. If RV790 has a 256-bit memory bus than it probably around 200 mm^2, or closer to what RV770 is right now if power consumption is around the same as the current RV770 and they include the sideport.

[largon]

Helmore,
Dunno how you got that 120mm² - it's way too low.
It would require a shrink worth of two full nodes. 55nm to 40nm is one single node (two half nodes, actually). One full node shrinks to about 75% of original.

[spursindonesia]

On 40 nm, i think ATi can easily hit around 1600 SP 80 TMU 24 ROP specced chip with die size budget beetween 260-300 mm^2, if they can keep their success in designing a pretty dense transistorwise GPU chip.

[Helmore]

Helmore,
Dunno how you got that 120mm² - it's way too low.
It would require a shrink worth of two full nodes. 55nm to 40nm is one single node (two half nodes, actually). One full node shrinks to about 75% of original.

R600@80nm. is 408 mm^2
RV670@55nm. is 192 mm^2
And they are pretty much the same chip, they added some features and cut the memory controller in half, but the amount of transistors is roughly the same.

[Sly Fox]

Looks like we're in for yet another year of complete ATI dominance.


Too much?

[largon]

Yep, my earlier figures are wrong. My late night algebra is sometimes off...
But these should be about right; a full node optical shrink from 55nm to 40nm reduces die size to ~55% of original.
256mm² @ 55nm -> 140mm² @ 40nm

edit:
Need 40nm process kgate/mm² figures...
:\

[LordEC911]

On 40 nm, i think ATi can easily hit around 1600 SP 80 TMU 24 ROP specced chip with die size budget beetween 260-300 mm^2, if they can keep their success in designing a pretty dense transistorwise GPU chip.

Your specs are too low or your diesize is way too large.
40nm's main advantage is that increased density. AMD/ATi already did a great job with RV670/RV770 density and I'm sure they will do it again.

R600@80nm. is 408 mm^2
RV670@55nm. is 192 mm^2
And they are pretty much the same chip, they added some features and cut the memory controller in half, but the amount of transistors is roughly the same.

R600 was ~420mm2 with 720m.
RV670 had UVD ondie, DX10.1 logic, power controllers.

[munim]

One interesting thing is that RV770 is very close to being pad-limited. Since pads don't scale as much as logic in a process shrink, one can assume that RV790 will likely feature more SPs or something else to prevent the chip from becoming pad limited.

However, they could also fix this by removing VDD/GND pads since the chip won't need as much power.

Can you explain your post for the uninitiated? What is "pad limited" etc?

[spursindonesia]

Your specs are too low or your diesize is way too large.
40nm's main advantage is that increased density. AMD/ATi already did a great job with RV670/RV770 density and I'm sure they will do it again.

.........

Mate, hence the word "easily" was put there. IMHO, the shader and TMU unit might not always only need to be increased in number, but if their per unit capability is expanded, that would be pretty beneficial too.

[shiznit93]

Looks like we're in for yet another year of complete ATI dominance.


Too much?

dominance?

GTX260 costs the same as 4780, is just as fast in real games people actually play instead of crap like call of juarez (minimum framerates often get ignored), overclocks better, uses less power, has better drivers (ati is always releasing hotfixes, still no game profiles), better looking anisotropic filtering, is quieter and cooler. ATI just now got a north american partner with lifetime warranty.

Obviously Ati does it with a much smaller gpu and nvidia's margins on the GTX are crap but from a consumer point of view the GTX series is just as good in price/performance and imo better when you consider reliability and support.

I guess you are going to bring up the 4870x2 but I don't consider multi-gpu viable.

[Sly Fox]

dominance?

GTX260 costs the same as 4780, is just as fast in real games people actually play instead of crap like call of juarez (minimum framerates often get ignored), overclocks better, uses less power, has better drivers (ati needs a hotfix for everything, still no game profiles), is quieter and cooler. ATI just now got a north american partner with lifetime warranty.

Obviously Ati does it with a much smaller gpu and nvidia's margins on the GTX are crap but from a consumer point of view the GTX series is just as good in price/performance and imo better when you consider reliability and support.

I was just joking. It was a parody of something Shintai had said. Which wasn't that bad of a prediction, but did end up not being all that true.

You'll see I actually have a 260, not a 4870.