AMD RV770 refresh -> RV790

Memory running 256bit 1400MHzx2 = will give 89.6GB/s bandwidth using GDDR4, it seems impressive! but with high latency of GDDR4 will yield lower performance :(

Their wasn't been so much performance improvements comparing Radeon HD3870-GDDR4 vs. Radeon HD3850-GDDR3.

My Radeon HD4850 Iceq4 memory running 256bit GDDR3 @1200MHzx2 = gives me 76.8GB/s bandwidth. GDDR4 @1400MHz will be slightly better!
 
(it rumors that AMD has ordered all GDDR5 chips that a memory manucturer produces in Q1 and Q2 next year). Probably RV775XT will be branded as Radeon HD5870. Real products are expected to be available late January, just after the new version of GTX200 Series.

AMD will launch RV775Pro in March 2009, while RV870 based on 40nm won’t be out before October 2009. No news indicates AMD will design single-card and dual-core products again, but why not?http://en.expreview.com/2008/12/12/amd-rv775-ready-to-give-nvidia-a-punch.html
Click to expand...


Let me get this right:
*12 SIMD
*48 TMU's
168 *5D = 840 streams
900MHz GPU
GDDR5 up to 1400MHzx4 = 179.2GB/s bandwidth "that's overkill"
*55nm
*927M Transistor count
*246mm2

The only way to believe that RV775 is smaller then RV770 is do to improve 55nm tech "advanced optimization"
 
Shtal said:
Let me get this right:
*12 SIMD
*48 TMU's
168 *5D = 840 streams
900MHz GPU
GDDR5 up to 1400MHzx4 = 179.2GB/s bandwidth "that's overkill"
*55nm
*927M Transistor count
*246mm2

The only way to believe that RV775 is smaller then RV770 is do to improve 55nm tech "advanced optimization"
Click to expand...


They also list it at slightly fewer transistors than RV770. I just assumed they imply AMD cleaned up/slimmed down the same basic design a bit.
 
homerdog said:
I do recall some talk a while ago regarding the redundancy in RV770 and it actually having something like 840 SPs. Maybe yields have improved to a point where they can now safely enable 40 more SPs without wrecking yields. Well I guess that would make since for the SPs since 840 is only 5% more than 800, but 48 TUs would be a 20% increase...
Click to expand...
RV770 has 10 SIMDs.
You can't have 840 SPs in RV770 because +40 SPs in 10 SIMDs would mean +8 superscalars and 8 isn't dividable by 10. So it's either 800 or 850 or 900.
It's the same for TUs -- they probably have 50 of them in there with 1 per SIMD for redundancy.
 
RV775 seems like Fud, too many weird architectural changes. Not saying it isn't possible, just not likely for a new revision/respin.
Sounds like RV740 is out late Q1/early Q2.
I don't think RV870 has been delayed past late Q2/early Q3.
 
DegustatoR said:
RV770 has 10 SIMDs.
You can't have 840 SPs in RV770 because +40 SPs in 10 SIMDs would mean +8 superscalars and 8 isn't dividable by 10. So it's either 800 or 850 or 900.
It's the same for TUs -- they probably have 50 of them in there with 1 per SIMD for redundancy.
Click to expand...

Ah. So reaching 840 SPs by enabling redundancy is not a possibility, and they actually (probably) have 25% redundancy in the TUs (is that normal?). Thanks.
 
RV730 SIMDs are only 40 wide, but the quad rendering is a pretty strong point against the 14*5 SIMDs it would have to be.
 
The Inq claims to have had a chat with Chris Hook and gotten some insights into AMD's roadmap.

http://www.theinquirer.net/inquirer/news/020/1050020/ati-s-hd-49x0-coming-soon-ish

ATI won't be launching anything in January, that's more or less guaranteed. A refresh to the current range of cards will be launched over the course of 1H-09, but not as early as January (CeBIT sounds like as good a time as any). That will include a delayed RV740 derivative and a suped-up RV770, which should receive the HD 49xx moniker. GDDR3 and GDDR5 will remain the memory bits and pieces, but ATI has dropped GDDR4 for good... too expensive and too much like GDDR5, said the Cap'n. Now the juicy bit is that AMD has narrowed down the 40nm parts launch to the first half of 2009 too, so between the derivatives and the shrinks there will be plenty of options on the market in 2009.

Last, but certainly not least, the 58xx series cards will be reserved for Q4 2009 and will pair up with Microsoft's DirectX 11 launch. So fanbois and fangoils, you'll have to wait a while.
Click to expand...

Interestingly they claim to know more but will respect their vow of secrecy. I thought the Inq didn't do NDA's? :LOL:
 
trinibwoy said:
Interestingly they claim to know more but will respect their vow of secrecy. I thought the Inq didn't do NDA's? :LOL:
Click to expand...

Nothing that a nice fat cheque addressed to VNU Publishing wouldn't change. Let's just hope it doesn't bounce on them... ;)

As for the rumor about a RV775/RV790 sticking to 55nm and "just" 840 shader units, i'd call it BS.
Most likely covert AMD FUD, designed to conceal a 40nm-based, 1000/1200+ stream GPU, or a new, de-coupled shader clock on 800 stream processors...
 
DegustatoR said:
RV770 has 10 SIMDs.
You can't have 840 SPs in RV770 because +40 SPs in 10 SIMDs would mean +8 superscalars and 8 isn't dividable by 10. So it's either 800 or 850 or 900.
It's the same for TUs -- they probably have 50 of them in there with 1 per SIMD for redundancy.
Click to expand...

There's a nice die shot here in the middle of the page showing the RV770 layout and you can see the individual groups.

Anyone want to comment on this thread at chiphell a month ago: claims a 99% yield, hints at a variation on the TSMC 55GP process.
 
rjc said:
claims a 99% yield, hints at a variation on the TSMC 55GP process.
Click to expand...
Haha. Wait, that was a joke, right? :) Maybe 99% yield *for the shader core* when you include coarse redundancy as in the HD4830, but for the overall chip? Uhhh, no.

EDIT: Also, probability that you could enable fine-grained ALU redundancy and therefore indirectly make branching granularity lol-ridiculous (i.e. 17/18 instead of 16): exactly, precisely and unambiguously ZERO. Those guys really want to make the rasterizer cry...

EDIT2: Just to be clear, I think ATI's redundancy mechanism is very good, I just don't think these claims make any sense in their context.
 
The RV740 ES use the same GDDR5 that the HD4870 is using, so if they go for this expensive VRAM RV740 has to need the bandwidth. Maybe because its more than just a shrinked RV730.
 
how does GDDR5 production ramps up? that GDDR5 from the 4870 should/would now become the "slow" GDDR5. Probably it deserves it being a 128bit GPU in the territory of the previous 256bit ones (RV670, G92)
 
Arun said:
Haha. Wait, that was a joke, right? :) Maybe 99% yield *for the shader core* when you include coarse redundancy as in the HD4830, but for the overall chip? Uhhh, no.
Click to expand...

Does defy belief. The poster's other information(also has posted info about the ati 40nm sample chips) seemed quite serious, strange to put a joke like that in there and not to admit to it when challenged in the thread that followed. Way back in march they were claiming 71% for the current RV770, anyone got any more recent figures?

What about the claim in the thread that the chip is using the TSMC 55nm GT process? Would it give them enough to last 6 more months on the RV770 design?
 
rjc said:
Does defy belief. The poster's other information(also has posted info about the ati 40nm sample chips) seemed quite serious, strange to put a joke like that in there and not to admit to it when challenged in the thread that followed.
Click to expand...

Read that thread like a month ago. I think by "yield" he meant 99% of the chip could run at the target frequency without a voltage bump over the current 4870, not the usual meaning of yield.
 
bearmoo said:
Read that thread like a month ago. I think by "yield" he meant 99% of the chip could run at the target frequency without a voltage bump over the current 4870, not the usual meaning of yield.
Click to expand...

That interpretation works much better. Interesting no one seemed to pick up on it in the thread, everybody including myself just assumed the naive version ie 2 to 3 bad chips per wafer - all hopeless optimists at heart i guess. ;)

Would fit better with the idea of the chip transferred from 55GP to the 55GT process for performance. Given the timing of the post assuming the information was roughly current and the chip didn't need another respin looking perhaps at february-march for release at best?

This chip would put a bigger gap to the RV730 below it, so i guess they will be pushing hard to try and get RV740 out not long after...say april-may?

If above is true really hard to see RV870 before late Q3
 
w0mbat said:
The RV740 ES use the same GDDR5 that the HD4870 is using, so if they go for this expensive VRAM RV740 has to need the bandwidth. Maybe because its more than just a shrinked RV730.
Click to expand...
Even RV730 could use GDDR5's extra bandwidth, it's really hampered by its 32GB/s...

Jawed
 
Jawed said:
Even RV730 could use GDDR5's extra bandwidth, it's really hampered by its 32GB/s...
Click to expand...
Probably only hampered a little more than the 4850 is, relatively speaking, so moving to 3.6 GHz GDDR5 and leaving core clock the same would probably only add 5-15%, depending on the game.

Of course, if ATI can crank up the clock for RV740 while adding GDDR5, it's going to be one awesome 128-bit card.
 
You must log in or register to reply here.
Back
Top