AMD RV770 refresh -> RV790
- Thread starter w0mbat
- Start date
Jawed
Legend
This confusion is as thorough as RV770's 
RV740 is 40nm and should be the first 40nm chip. One rumour said RV740 was supposed to have been a November/December/January (take your pick) launch which was pushed back by TSMC and that the implication is that RV790 is also 40nm because it would have fitted in nicely as a February/March/April launch, after RV740.
That does make a fair amount of sense, since I'm expecting RV740 to perform like HD4850 but with a 128-bit bus, GDDR5 and a real puzzler of a question over its RBE count/configuration. RV740 would be a little monster.
That would lead into RV790 also being a monster, and the only way to not choke it is to give it more RBEs or up the Zs per RBE. In which case GTX285 would be dust.
Trouble is, I think RV790 is just a boring, timid, refresh on 55nm.
Jawed
RV740 is 40nm and should be the first 40nm chip. One rumour said RV740 was supposed to have been a November/December/January (take your pick) launch which was pushed back by TSMC and that the implication is that RV790 is also 40nm because it would have fitted in nicely as a February/March/April launch, after RV740.
That does make a fair amount of sense, since I'm expecting RV740 to perform like HD4850 but with a 128-bit bus, GDDR5 and a real puzzler of a question over its RBE count/configuration. RV740 would be a little monster.
That would lead into RV790 also being a monster, and the only way to not choke it is to give it more RBEs or up the Zs per RBE. In which case GTX285 would be dust.
Trouble is, I think RV790 is just a boring, timid, refresh on 55nm.
Jawed
Maybe they threw out the CFX-port, added two or one additional cluster and did some floor plan optimization to end up at 205mm^2 at 55nm.
Back in October, there were some rumors about a 205mm² 1.5TFLOPs(what would be RV790s ~850MHz @880-960SPs) GPU:
http://translate.google.ch/translat...rv870-neue-details&sl=de&tl=en&hl=de&ie=UTF-8
Back in October, there were some rumors about a 205mm² 1.5TFLOPs(what would be RV790s ~850MHz @880-960SPs) GPU:
http://translate.google.ch/translat...rv870-neue-details&sl=de&tl=en&hl=de&ie=UTF-8
Taping out an entire new chip on the same process, with ~10% improvement in clocks and alu's, sound mighty unlikely to me.
Instead, doing the same effort for a smaller node seems more plausible to me.
Jawed
Legend
That's just a reckless guess, ignoring routing. As to scheduling, I still haven't worked out how this is configured across the die - i.e. how much scheduling is within a cluster and how much is outside...
Jawe
RV790 is 55nm no?
Well, TBH, the specu posted earlier makes me think Jawed may be along the right track. If they have the samples and are telling the world about the clocks, why wouldn't they say that it has x more alu's if it had? But then why haven't they said that it has same either. Of the two, I find the earlier more probable, though I can't figure out why rv790 would be on 55nm.
And since tsmc's 40 nm process leaks madly, it would seem reasonable that a 40 nm chip won't have more clocks. After all, IIRC rv670 and rv770 have the same clocks too, to keep the power in check.
And since tsmc's 40 nm process leaks madly, it would seem reasonable that a 40 nm chip won't have more clocks. After all, IIRC rv670 and rv770 have the same clocks too, to keep the power in check.
Jawed
Legend
Jawed
Legend
For what it's worth, the rumours are for more than one 40nm chip. Has "RV720", RV710's replacement, ever been mentioned? If not, then doesn't that make RV790 likely to be the other chip an 40nm?
205mm2, say 16 clusters, 1280 ALU lanes, ~2.2 TFLOPs with a 256-bit bus and no change to the RBEs.
Jawed
205mm2, say 16 clusters, 1280 ALU lanes, ~2.2 TFLOPs with a 256-bit bus and no change to the RBEs.
Jawed
The 3 leadoff 40nm chips are - RV740, GT216 and GT218. Not sure what else comes after, amd also needs something to replace their igp, also i swear i read last week somewhere jsen huang let it slip ion would dx11 before the end of the year(will spend rest of today looking for this link
).The rumor i heard was ~290mm2, that was last year. Wasn't sure if the extra size was due to the change from 55GP to GT or they had added extra units. Sorry is not 40nm, the vr-zone link said 1.3V above TSMC's 1.2V max for 40nm...exceeded on the first chip, doesnt seem very probable.
Is interesting planning to try and put out a US$300 part, that seems way high without a lower priced volume part. Maybe they did something to the MC/RBE's so it only rally works well with GDDR5, samsung expects GDDR5 to be 20% of the market this year.
i swear i read last week somewhere jsen huang let it slip ion would dx11 before the end of the year(will spend rest of today looking for this link
This? http://blog.laptopmag.com/nvidia-ceo-sounds-off-on-netbooks-ion-platform-tegra-and-mids
JHH said:
Could be a simple slip of the mind. He would obviously have DX11 on the brain these days.
alu count seems reasonable, but memory bandwidth seems to have very little growth compared to rv770, when it was already texturing limited in certain games.
This makes me wonder, could amd decide to implement some edram (ie xenos like) stuff on future stuff (may be not rv790/740 but on rv8xx perhaps or may be even later), that would relieve a lot of bandwidth requirements. It could be on die if necessary and since display resolutions don't grow as fast as moore' law, the cost could be manageable/negligible in not too distant future.
Even though the resolutions don't grow rapidly anymore, 2560x1600 requires already damn lot of it, can't see it as really reasonable solution anymore, unless they go for "real tiler" (which to my understanding would cut the mem requirements for edram)