AMD RV770 refresh -> RV790

My only doubt is that if the refresh is 2x faster, then what will they do with RV870, considering that it is their goal to increase perf by ~2x every gen and their die size is limited by their design choices.
Maybe RV870 has no RBEs and all that freed-up die space is given over to ALUs+TUs and back-end operations are done programmably.

I'm thinking 32 RBEs in RV790 would take a lot of space...

Jawed
 
That, in my view, would be a 55nm refresh, say with CrossFireX Sideport deleted and higher core/memory clocks.

Anyway, right now, I'm leaning towards a 40nm monster, if there really is a refresh for RV770. 2x performance, across the board... If RV730 can get approaching 2x performance in its refresh, why not RV770?

Jawed

Actually I was thinking of this in terms of 40 nm. Using a 128-bit memory interface (albeit with much faster GDDR5) it would be very similar to the R600 -> Rv670 refresh. Abeit in this case rather than going for as much power and die size savings as possible, they could instead go for a modest performance increase along with a modest power use decrease. Add to that a small die to offset the increased cost of doing business at 40 nm.

IMO - That would be an ideal "refresh."

I find it fairly amazing that virtually no information has been leaked with regards to a possible Rv790 if rumors are to be believed that sample chips have been received.

Regards,
SB
 
Maybe RV870 has no RBEs and all that freed-up die space is given over to ALUs+TUs and back-end operations are done programmably.
All that reworking MCs, L2 and AA in the ROPs for just one generation to use? Possible, but not very economical from where i'm sitting.
 
All that reworking MCs, L2 and AA in the ROPs for just one generation to use? Possible, but not very economical from where i'm sitting.
Alternatively, 32 colour writes per clock? What percentage of frame render time is spent doing colour writes?

I'm just thinking, with D3D11 not explicitly creating a "programmable output merger" stage, but instead doing all that funny business with pixel shaders being able to read/write "render targets" and comprehensive access of compute shaders to memory...

But yeah, you do have a point.

Jawed
 
With the quite solid info on the rv740 specs and die size around 100mm2, I can't help thinking that a simple doubling of that would make perfect sense for the 790, ie:
32 RBE, 1280 SP, 64 TU, 256 bit @ 190-210mm2.

Anyone dare to comment on the die size scaling from such a doubling? I think we can expect the chip to be just big enough to support the 256bit bus (and power usage), but would a double rv740 take more or less the double size?

I guess 18 clusters instead of 16 would make a better balance, but they probably don't want it bigger than necessary, and the balance would be closer to G2xx this way.
 
TBH at any decent clocks, chip like that should pretty much trash by clear advantage every single GTX out there excluding the 295
 
TBH at any decent clocks, chip like that should pretty much trash by clear advantage every single GTX out there excluding the 295

Then again, "RV790" won't be selling for quite a while, now will it ? ;)
I'd expect the GTX 260's, 280's, 285's and 295's to last for a relatively brief period, as they're not really cost competitive with RV770 and/or RV790.
 
Actually I was thinking of this in terms of 40 nm. Using a 128-bit memory interface (albeit with much faster GDDR5) it would be very similar to the R600 -> Rv670 refresh. Abeit in this case rather than going for as much power and die size savings as possible, they could instead go for a modest performance increase along with a modest power use decrease. Add to that a small die to offset the increased cost of doing business at 40 nm.

IMO - That would be an ideal "refresh."
Hmm, I guess while we're playing whack-a-mole, 1.6GHz+ GDDR5, for about 100GB/s+ would do the trick :smile:

Jawed
 
Hmm, I guess while we're playing whack-a-mole, 1.6GHz+ GDDR5, for about 100GB/s+ would do the trick :smile:

Jawed

Doesn't GDDR5 top out at ~1.1GHz currently ?
Placing expensive -and non-existent- 1.6GHz GDDR5 on a mainstream product sounds preposterous... :smile:
 
Maybe RV870 has no RBEs and all that freed-up die space is given over to ALUs+TUs and back-end operations are done programmably.
Are RBEs render back ends? If yes, what's the diff b/w them and ROPs?

With the quite solid info on the rv740 specs and die size around 100mm2, I can't help thinking that a simple doubling of that would make perfect sense for the 790, ie:
32 RBE, 1280 SP, 64 TU, 256 bit @ 190-210mm2.

I'd suspect that there will be more SPs than that
 
Doesn't GDDR5 top out at ~1.1GHz currently ?
Placing expensive -and non-existent- 1.6GHz GDDR5 on a mainstream product sounds preposterous... :smile:

At least Samsung and Qimonda have 1250MHz (5GHz) in production at the moment afaik
 
Doesn't GDDR5 top out at ~1.1GHz currently ?
Placing expensive -and non-existent- 1.6GHz GDDR5 on a mainstream product sounds preposterous... :smile:
Q2, 6Gbps GDDR5 will be available:

http://www.hynix.com/inc/pdfDownload.jsp?path=/datasheet/Databook/1Q09_Databook_GraphicsMemory.pdf

4 chips would provide 512MB only, though, but 96GB/s. I think having only 512MB of memory is actually a more compelling argument against "HD4890" being a 40nm, 128-bit, cost-reduced refresh of HD4870 - 1GB cards are now widely seen as the best choice in this segment of the market.

But, if I understand this correctly, a 1GB HD4890 with a 128-bit bus could be constructed using the clamshell mode of GDDR5. This has 2 memory chips per 32-bit channel, doubling capacity.

Also, Hynix has announced 7Gbps GDDR5 for 2009H1 availability:

http://www.hynix.com/gl/pr_room/new...&SearchWord=&SELECT_DATE=&menuNo=02&m=01&s=01

Jawed
 
You're describing the old mechanism there for dual-rank/clamshell. GDDR5 actually has a x16 mode, so each device has half the bandwidth, but you double up on the number of devices.
 
Last week another RV790 sample was tested. Here is interesting that the memory was clocked with 975 instead of 900 MHz.
By the way Qimonda IDGV1G-05A1F1C-40X memory (1024 MiB), too.

Source: Hardware-Infos
 
From half a dozen different threads on chiphell it appears the RV790 is in (or just about to enter) production. Is a 55GT version of the RV770. Guessing from the timeline cebit in march looks like a good date for the launch.

Also the 40nm RV740 appears to be having some problems, not sure if it is still the power issue reported previously - 40nm not providing the power savings they were expecting - or something else. Pretty sure they were doing an A12, not sure if this implies a A13. Generally 40nm TSMC is struggling, on the order front(ie half what they expected), financial ("pay free" days for workers, redundancies, delaying and reducing equipment orders) and usual technical difficulties for a new process.

Have heard a nvidia 40nm chip is doing ok though, either GT216 or GT218 not 100% which. The graphics market is half the size it was last year, nothing much is certain, new designs are prone to delay or cancellation at any time.
 
Last edited by a moderator:
So, 10 or 20% higher clocks, otherwise the same specs?
Quite believable actually.
GT200-216 level, aprox. GTX295 level for 4890X2.
And RV870 in 3-4 Q as the next chip for this segment.
GT212 will probably be out earlier than RV870 but will probably end up being slower and without DX11 of course (more ore less the same way it was between RV770 and G92).
GT300 will be out later -- and the big question is will it be able to coup with RV870X2 this time around?
Plus GT212 vs RV870 situation isn't looking too good for NV so i won't be that surprised if Charlie was right and GT212 got cancelled in favor of some GT3x0-chip which was pushed closer in the roadmap.
 
Back
Top