AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

[Man from Atlantis]

hmm that great overclock numbers make me think rumours about nvidia's leaving hotclocked SPs is real.. even a factory overclocked card can do %45 better lol, ppl should reach 1400-1500MHz with msi tf iii, asus 3slots and gbt 3fans cards vs..

also feeding the conspiracy theory i dont know but isn't there a rumour that Tahiti's CU count is factor of 3 ?

 

[silent_guy]

AFAIK 0,85/1,17 volts are (one of?) the default configuration(s) for HD 7970 - at least that's what the tools say.
HT4U did some multimeter action on their sample: http://ht4u.net/reviews/2011/amd_radeon_hd_7900_southern_island_test/index17.php

Interesting. Talk about going beyond the usual canned benchmark review.

So 1.17V looks like a regular operating voltage, not just some overclock configuration. That explains the higher than expected power consumption at load. Maybe it's a bit higher due to power gating to compensate for a small voltage drop across the gating transistor?

Or, mind blown, it is HPL after all? Still a remote option, though. Just look at those clocks...

 

[Man from Atlantis]

this is what he did..

http://kingpincooling.com/forum/showpost.php?p=18959&postcount=28

ok before someone cry i will be nice again

1.6vcore (pretty sure its needed)
1.86dram (not sure if its needed)
1.06+ vddci (not sure if its needed)
+ overmodding of a bunch of other stuff (not sure if its needed)

-120C ~ -180C ~ depending on resolution (not sure if its needed)

heat is a walk in the park, maybe 1/2 of fermi easy splashin jobs on KPC tek 9 slim

 

[TKK]

There are no 36 CU designs.

But maybe there are 40 CU designs...

 

[Dooby]

"Da Original" ? I smell shens.

 

[Wirmish]

Also, does Tahiti secretly have 36 CU's after all? 2304 shaders on the "cancelled products" would suggest such

Rage3D

Eric: The CUs themselves, the whole thing is very scalable, you can disable CUs one after another and you can improve yield that way.
Let's say you have one bad CU, well it's not a 7970 anymore, but maybe it's a 7950? We will scale CUs, don't know if we'll scale ROPs.

Tom: We have redundancy there anyway.

Eric: Yeah, we have redundancy there, and I think we will be scaling the memory but I'm not 100% sure.

Tom: Yeah, it's pretty sure.

36 CU = 1 CU-Array (4 CU) for redundancy ?

 

[3dilettante]

It's not clear from that exchange if they meant there is redundancy for the CUs or for the ROPs.
I can see there being redundant ROPs, since they are no longer per-channel and are fixed-function blocks organized in compact arrays.

A die shot would give the answer, but none has been shown that I know of.

 

[OlegSH]

36 CU = 1 CU-Array (4 CU) for redundancy ?

How it will split across two of ACE? Looks like a BS to me

 

[MBAMath]

32 CUs, not 36, no hidden unit for future use. Got that straight from horse's mouth.

 

[Raqia]

http://www.chiphell.com/thread-336984-1-1.html

Looks like there's a lot of headroom in this process:

http://news.softpedia.com/news/28nm-Next-Generation-GPUs-45-Faster-232468.shtml

and they've gone the route of limiting power consumption for now. AMD may need to squeeze in a few new model names between the 7970 and 7990, the latter of which will presumably be the dual gpu card.

 

[Alexko]

I suppose they could always release a 7980 or whatever, but I think they're more likely to just let partners make overclocked 7970s: same basic effect, much less hassle.

 

[Dooby]

Give me a 36CU, 1400mhz 7970U for £450-500, and I'd be (finally) happy with this card. 50(ish) GPixel and 200(ish) GTexels are nice round numbers.

 

[Kaotik]

36 CU = 1 CU-Array (4 CU) for redundancy ?

Usually redundancy has been by extra units within "the big unit", not complete extra unit?

 

[RobertR1]

32 CUs, not 36, no hidden unit for future use. Got that straight from horse's mouth.

I might have to drop my GTX580 for this quickly.

Can't wait to see BF3 benchmarks of the Atomic RX.

 

[Dave Baumann]

It's not clear from that exchange if they meant there is redundancy for the CUs or for the ROPs.
I can see there being redundant ROPs, since they are no longer per-channel and are fixed-function blocks organized in compact arrays.

A die shot would give the answer, but none has been shown that I know of.

It's ROPs. Also why there are no half RB configurations of Cayman.

 

[Man from Atlantis]

AMD HD 7950 won't show up on January 9th

 

[UniversalTruth]

AMD HD 7950 won't show up on January 9th

Very interesting what their explanation is.
If it is because of the bad yield, then they should have more salvage ready parts rather than "fully-enabled" (question mark here) R7970s.
Or R7970 is the salvage part and the fully-enabled is canceled or pending launch even later, while R7950 could be simply what everyone thinks R7930 should be. /sarcasm tags

 

[Shtal]

http://www.freeimagehosting.net/ffb04

http://www.freeimagehosting.net/7f831

http://kingpincooling.com/forum/showthread.php?t=1553

 

[trinibwoy]

Very interesting what their explanation is.
If it is because of the bad yield, then they should have more salvage ready parts rather than "fully-enabled" (question mark here) R7970s.
Or R7970 is the salvage part and the fully-enabled is canceled or pending launch even later, while R7950 could be simply what everyone thinks R7930 should be. /sarcasm tags

It doesnt make sense but Guru3D corroborates fudo's story and they claim AMD as the source.

http://www.guru3d.com/news/amd-radeon-hd-7950-delayed-to-february-week-1/

 

[Rangers]

It doesnt make sense but Guru3D corroborates fudo's story and they claim AMD as the source.

http://www.guru3d.com/news/amd-radeon-hd-7950-delayed-to-february-week-1/

Such is the annoyingness of people complaining about "paper launches". Now because of them we dont get the 7950 info for nearly another month, for no good reason. Just so that people wont complain about a "paper launch".

Card would have available early February either way, only difference is now we wont know more about it for longer.

Longtime paper launch supporter here.