Official GT200 Reviews/Benchmarks Thread

Would it have been launched a week earlier, noone would have called the GTX260 "good value", that's for sure...
Interesting there's no mention of memory clock, however. At the price point it's going to be I guess it can't really be faster than the memory of the 9800GTX (1100Mhz), but if it's slower the 9800GTX+ probably won't be faster than the 9800GTX and even less of a threat to a 4850...

 

, and thanks.

I finally looked at Xbit's About page, and I see that, while they're Russian and partly based in Russia, they seem to be an English-only site, so no Russian home page for early #s.

 

http://www.digit-life.com/articles3/video/gt200-part1-p1.html

 

Looking at the numbers from xbit, it really isn't all that much better than the 8800U. Drivers or should they have simply waited for the die shrunk version? I'm thinking the latter as it would have allowed them to make the drivers even that much better.

 

http://www.bit-tech.net/hardware/2008/06/24/nvidia-geforce-gtx-280-architecture-review/1

Bit tech review is up. This review looks at synthetic performance (AA performance, GS performance and other various shader tests etc).

According to this, GS performance is very good. Power consumption is pretty impressive i might add.

 

Seems you can make your own "G200B-GTX 280":
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=6609594&postcount=358
3D-Voltage to 1.03V from 1.18V with Nibitor
-> 680MHz OC still possible, but 65W lower consumption over default and lower noise(65% is enough, while default goes up to 100%).

Sure, it may varying from card to card.

 

Apart from clock increases, the GT200b should also bring lower consumption and overall shorter dimension card. That and if priced right, I will pick it up. That is if I can manage to hold off from the 4870.

 

IMHO, G200b will suck as much power as they can throw out of the box (so ~G200) - but it will ship with clocks as high as possible.

 

what about GT200 on 45 nano ? can it be done ?

 

Yes it can be done.. question is.. Will we see it before 1h 2009?

 

What will be interesting is what happens if there's only space for a 256-bit bus, necessitating the use of GDDR5.

With half the bus width this chip will also have half the ROPs.

Will half the ROPs, presumably at considerably greater clocks than GTX280, be enough?

Alternatively, of course, NVidia could change the ratio of ROPs to units of bus width, i.e. put twice as many ROPs into each partition.

Jawed

 

Precisely. I shake my head everytime I read a so-called "pundit" or "enthusiast" speak of a future shrunken GT200 using a 256-bit bus and GDDR5. Sure that addresses the bandwidth aspect of the equation, but as you mention, NV's ROP partitions are tied into memory controller channels (1:1) so unless NV's got some "super ROPs" in the works that can achieve twice the current workload, it just isn't going to happen.

 

But are they?

If NVidia re-architects for GDDR5 then perhays they can double them?

Alternatively, maybe the ROPs are just waiting for more bandwidth. I've long argued that GDDR3 is too slow per ROP for NVidia's architecture (even more so for RV770) - there's a load of performance just going unused. So it may turn out that changing to GDDR5 unleashes the ROPs sufficiently that halving their count with 20-30% clock boost and despite only having 80% or so bandwidth (i.e. less bandwidth in total) still results in at least GTX280 performance.

Jawed

 

Yes

I certainly won't rule it out of the realm of possibility, but I think that's a lot to ask for. Perhaps each memory channel could service two ROP partitions in a future 256-bit GT200-derived GDDR5 SKU? Although I suspect this would lead to an increase in latency without a rearchitecting of the memory controller. Perhaps we'll see a 40/45nm GT300 (or whatever you care to call the successor to GT200b) featuring a ring bus

I can't find a reason to disagree with this. NV's got such huge Z fill rate and AA sample rate there's just no way they can feed those ROPs with GDDR3.

I see what you're getting at here, as per-ROP bandwidth would increase with the proposed transition to 256-bit/GDDR5 so efficiency should increase as well, since NV's ROPs are rarely running @ 100% utilization (ever?).

I don't know if a 20-30% clock boost is reasonable with such a massive chip undergoing such a minor shrink... I can see a potential GTX 290 with 650/1500 clocks on 55nm but beyond that would have to be up to the individual AIB partners...

Also, a GDDR5 GT200b SKU would not have to provide less bandwidth than GTX 280 does already, NV would just have to use 2.2-2.5GHz chips instead of down-clocking 2GHz chips as ATi does.

So yes, per-ROP bandwidth would increase (assuming NV doesn't double ROP count per partition or create "super ROPs" that achieve twice the throughput of their current ROPs).

 

Getting rid of 256-bits of memory interface should make for a better than minor shrink - and getting rid of half the ROPs would make it even better. Also, we're talking 65-40nm here, which is a big shrink. That's Oushi's question (45nm is supposedly really 40nm).

Jawed

 

Good point. I did not account for these facts previously. I'm not sure you can toss out half the ROPs and still maintain Z fillrate or AA sample perf. though. Even with the efficiency improvements GDDR5 and a new memory controller would bring, it's going to take a hefty dose of increased core clock to account for a halving of ROPs.

Someone needs to run some bandwidth efficiency tests on GT200 (i.e. overclocked and underclocked).

My mistake. I thought your clock speed increase was in reference to GT200b for some reason.

 

i am very sad for nvidia for making this huge brilliant chip and still ATI can catch up
with it so easly , can someone explain to me what is holding GT280 back or it's just ATI
had made a superior chip than GT280 ??