NVIDIA GF100 & Friends speculation

1GiB @ 192-Bit? 4*1024MBit + 2*2048MBit chips? Is this possible?

Also 28% more performance with only 14% more TMU/ALU-peformance seems a bit strange, since GTS 450 did not looked so bandwidth-starved
From my experience it IS bandwidth starved , especially with AA on of course , and that is compared to GTS 250 , some PhysX titles perform better on GTS 250 too .

Any idea whether the increased ROPs count would actually benefit the card ?
 
From my experience it IS bandwidth starved , especially with AA on of course , and that is compared to GTS 250 , some PhysX titles perform better on GTS 250 too .

Any idea whether the increased ROPs count would actually benefit the card ?
No way imho. 16 or 24 ROPs doesn't really matter since the chip can only output 8 pixels per clock anyway. Now depth/stencil tests don't suffer from that, but even with only 16 ROPs there's plenty of throughput there.
I consider it unlikely asymmetric memory configurations (4 1gbit and 2 2gbit chips) are supported too, hence it's probably either 192bit/1.5GB or 128bit/1GB.
Now if that's 128bit the 28% quoted performance increase actually fits perfectly: 13% more memory clock + 15% more shader clock = 28% more performance. Wouldn't surprise me the least if nvidia calculated like that :).
Let's see though a 192bit bus certainly might help a bit, while I don't think GTS450 is tremendously bandwidth starved, for a 50% bandwidth increase getting another 10% overall performance increase (which is all that's needed for about the quoted 28% increase) might be quite possible.
 
But the shop listings and performance leaks points to the 1gb/128bit solution, 10-15% faster than 450.
http://vr-zone.com/articles/geforce-gtx-550-ti-spotted-in-eu-e-tail/11343.html http://vr-zone.com/articles/nvidia-geforce-gtx-550-ti-benchmark-leak-/11420.html

I Guess these are the pictures removed from ht4u: http://vgacentral.wordpress.com/2011/03/02/some-gtx-550-ti-slides-from-cebit/ (btw, 25% faster in F1-2010 would mean 70% faster than 450, or a massive driver improvement: http://www.computerbase.de/artikel/...-geforce-gtx-560-ti-sli/16/#abschnitt_f1_2010 )

But I don't see how you would make 1gb/192bit with standard sized chips.
 
Last edited by a moderator:
So this will be competing with the 6850/70. That's bad news for AMD. This is a cut down GF114 right? that was already 36mm^2, it will literally be cut in half to get to the 192SP count we've seen so a die size of 180mm^2 isn't out of the question. Barts has a die size of 230mm^2 so Nvidia will be able to sell for a lower price for similar performance.

The 400 series may have been a bust for Nv, but the 500 series has been absolutely stellar. Fermi redeemed...
 
Last edited by a moderator:
Nah, even IF it's 192bit it should be slower than 6850. It's not a GF114, but a 106/116, just like the 106 is already selling in 192 bit configurations for notebooks (450M, 555M iirc) - with normal expected 1.5gb..
GF106 is reported ~240mm^2 btw, so not exactly a die size advantage..
 
The GF106 and GF116 has a 192 bit memorybus, only 64 bit is disabled by the GTX 550 Ti and GTS 450. ;)
The funny part is, a full 192 bit/24 ROP/768 MB SKU would probably be a few percent faster.

I guess Nvidia sticks to 128 bit/1024 MB because it's more marketing-friendly (1GB > 768MB).
 
The GF106 and GF116 has a 192 bit memorybus, only 64 bit is disabled by the GTX 550 Ti and GTS 450. ;)



The funny part is, a full 192 bit/24 ROP/768 MB SKU would probably be a few percent faster.

I guess Nvidia sticks to 128 bit/1024 MB because it's more marketing-friendly (1GB > 768MB).

768MB and 1.5GB versions of the GTX 550 would be even more marketing friendly. Sadly some people, would think they just bought a card equal to the GTX 580 at 1/4 of the price, if they saw 1.5GB on the box!

As for the one 64bit channel being disabled, Donanimhaber posted some slides the other day, that have been taken down now at the request of Nvidia, one of which showing the 550 having a +70% memory bandwidth over the 450. How could this possible if one 64bit was disabled? It wouldn't!

With one 64bit channel over the 450, you are instantly at +50% mem bandwidth. Add some clock frequency to that and there you go.

Things are simple me thinks. Just take this graph, enable the disabled parts, increase clocks, presto, the GTX 550!

arch.jpg
 
Things are simple me thinks. Just take this graph, enable the disabled parts, increase clocks, presto, the GTX 550!
Sure yes, but this doesn't mix with 1GB (unless it uses an asymmetric memory configuration).
And if it's just 128bit, the chip perfomance / area isn't really too good at (supposedly the same size as GF106) 230mm² (but well having disabled parts wouldn't be helping there of course). That would be just barely faster than HD5770 (except for HAWX 2) with the announced clocks - no way to touch a HD6850.
Even if it's 192bit, adding ROPs and bandwidth won't do THAT much - yes should be enough to really leave a HD5770 behind, but still a good deal slower than HD6850.

Actually some leaked numbers are out. These confirm 128bit and barely faster than HD5770, at a much higher price (for now). That said, I guess if it won't sell it can easily sell for cheaper, however the GTS->GTX name change seems totally unjustified to me for a product with just 15% higher performance (in fact, the relative difference to the next higher product, the GTX 560, has significantly grown compared to GTS 450 -> GTX 460).
Note that it looks like AMD will release HD6790 in the near future, which is presumably (my guess - other possibility would be crippled Barts) a slightly higher clocked Juniper XP.
 
Last edited by a moderator:
bandwith is good, who plays without AA?
what a stupid card now.

well probably the naming is to differentiate from the cut-down version. It expect it to be 144SP, ddr3 on 192bit, and might well be called GTS 540.
for some reason there can only be one GTS in a nvidia line up.

indeed try to sell a GTS 550 1GB when there's a cheaper GTS 540 1.5GB ; or imagine wasting the cut down card (which is the missing card filling a gap) by naming it GT 540.
nvdia knows what they're doing, those letter games are a crucial part in the perception of the product.
 
I think with the asynchronous memory configuration, this card will be more dependant on using the right set of drivers than anything before it. So, quite likely prematurely leaked numbers are even more worthless than on any other card so far.

That said, I don't see the unique selling point in this card either.
 
I think with the asynchronous memory configuration, this card will be more dependant on using the right set of drivers than anything before it. So, quite likely prematurely leaked numbers are even more worthless than on any other card so far.
No I think you got that wrong, the leaked numbers seem to indicate there's no asymmetric memory configuration, and it's just a slightly pimped GTS 450.
That said, I don't see the unique selling point in this card either.
Well it's not a bad card per se, while the GTS 450 couldn't beat the HD 5770 this probably can (if just barely). It just needs to be priced accordingly. I'm not sure how it will compare to the HD6790 but if that's really Juniper things won't change much (there is likely not much clocking headroom available over HD5770).

Blazkowicz said:
bandwith is good, who plays without AA?
Sure yes but these nvidia cards are not that dependent on memory bandwidth. Just look at GTX 460 1GB/768MB difference.
 
PCGH says Nvidia confirmed on CeBIT, that GF116 is able to do the combination of 192-Bit and 1GiB memory: http://www.pcgameshardware.de/aid,8...Cebit-weitere-Details-Update/Grafikkarte/News
Nvidia hat - ohne bislang Details zu nennen - die Kombination von 1,0 GiByte Videospeicher und 192 Bit Speicherbus bestätigt. Man habe mit den neuen Chips flexiblere Möglichkeiten, Speicherausbau und ROP-Partitionen zu kombineren und so auch mit drei 64-Bit-Partitionen vier bzw. acht GDDR-Chips anzusteuern
 
The last 256MB will have to share their bandwidth when addressed?

Prefer the first 512MB, then 50% of the extra chips?
 
There is completely nothing impossible about this, combination of different density mem's chips could easily solve problem
Pretty sure that would needlessly complicate the memory controller, which would make for a significant engineering challenge to get it all to work without a performance hit. I sincerely doubt that nVidia would do that.
 
Back
Top