NVIDIA GF100 & Friends speculation

So, Napoleon is suggestion that the GTX580 is using a revision of the GF100 core.. and GF110 and GF114 are pin-to-pin compatible.

Trinibwoy, this is becomming more and more interesting

 

Hmh? It was 560 & 460 are pin-to-pin compatible, not GF110?

 

Imo the device-IDs of D13U-SKUs are to different to be just GF100 A4s or B1/2/... .
Its probably more a redesign like G70 to G71, but without process change.

 

Predicted clocks etc:


Cooler details:





Nvidia GeForce GTX 580 cooler seen in detail | TechConnect Magazine

 

GTX 580 vs GTX 480 OC:

GTX 580 @ 772/1544/2004MHz ?

http://www.chiphell.com/forum.php?mod=viewthread&tid=133004&extra=page/=1&page=4

GTX 480 @ 925/1850/1950 MHz with 260.99 WHQL

http://www.forum-3dcenter.org/vbulletin/showthread.php?p=8355798#post8355798
Feature tests: http://www.forum-3dcenter.org/vbulletin/showthread.php?p=8285416#post8285416

Feature tests of "GTX 580" look normal for 512SPs@772MHz, with a bit improved color fill (higher BW), but graphics tests seem significant faster than on the OC GTX 480.

Also the feature test 1 shows, that there are only 64 texture addressers.

 

So we're looking at a ~10% clock increase (both chip and mem) and an additional SM for about a 15% performance increase all in all over GTX480, preferably with the same or lower power envelope?
If that's only a respin of GF100 (I always thought this might make sense) I'm wondering however what took them so long?
Edit: if it's not a respin, any ideas what could be changed?

 

Reduced die-size?
Display and video features: HDMI 1.4a, improved video processing, single-card 3D Vision Surround?
G80 like TMUs with doubled TFs?
Significant better performance/watt (Napoleon@Chiphell said 120% performance at 80% consumption -> 50% increase in performance/watt)?

 

That would be nice. The transistor density of GF100 is not that much smaller than that of GF104, however, so I would be surprised if there are miracles there.

Ah yes the usual suspects. I was more wondering about changes in the shader core.

Interesting idea. Would be a bit funny - they abandoned that design for G9x and now it would be back. Might just be what's needed to get that quoted 20% performance improvement I keep hearing...

Couldn't the perf/watt improvements be achieved with a "simple" respin, i.e. no functional changes? I always thought it's not really the architecture but just the implementation was broken (well compared to GF104 at least, it might still be overall less efficient than Evergreen). It's interesting though this 120% performance figure comes up - that's slightly more than what you would expect from the additional SM plus the 10% clock boost.

 

+1 SM + higher core clock should result in +17.6% performance. Maybe they'll use full-speed FP16 TMUs (like in GF104). Or the "+20%" is just rounded value.

 

At best! In practice that would probably be more like +15% tops. You'd need something significant on the architectural side to reach +20%.

 

Maybe they increase the base-clock (ROPs, L2) >20% over GTX 480?
But they idea with FP16-TMUs seem very likely, even GTX 460 did not show a bigger increase against GTX 465.

 

It would fix another strangeness of the Fermi lineup: the chips with higher tex/alu ratio AND less bandwidth (per TMU) can do full speed FP16, whereas GF100 can't. Made no sense to me, except if you think of GF100 as primarily a compute device. I guess though any performance increase this could have will be quite depending on the application (need to use 4 channel FP16 textures in the first place...).

 

If GF110 is still half-rate DP capable they might be able to do a Tesla upgrade (Tesla 2060/2080).
With ~50% increased performance/watt it would be possible to get near to 700 GFLOPs DP.

If Cayman is really 1920 4D he should be able to push out >700 TFLOPs DP in a <=225W version (~750MHz) and >1 TFLOPs DP as dual-gpu-board with <=300W (~ 650MHz).

 

I also think a new model of Tesla will be offered at higher $$$. More pure profit for Nvidia.

But with no ECC there will be no HPC for Cayman.

 

Charlie's take: THE DESPERATION AT Nvidia has reached silly proportions, with the paper launch of the GTX580 pulled in from 'Cayman Day' to 'Investor Conference Call Day', both holidays in November. I wonder how many reviewers will overlook the fact that there won't be cards this year? We told you about the GTX580 a few weeks ago and now a few more details have come out. The short story is that the chip behind the GTX580, the GF110, is nothing more than a bug fixed GF100. If you think this should be called the GF100b, you are right. If you think performance will underwelm, you are right too.

 

Weird, I thought he said GF100 was unfixable?

 

It's the long predicted "b" respin rebadged as a new product line. Whether it can find a market and a profitable price (for Nvidia) given it's HPC targets, large die size, yields, etc, is another question.

It will be interesting to see when it finally manages to get onto shelves, rather than just an announcement at an investors conference. Things must be serious if Nvidia decides to use the GTX580 launch months ahead of shipping to bury bad news, rather than as a spoiler against the Cayman launch.

 

So it was fixed then?

 

Redesigned/replaced. That's separate from it being a viable product.