http://www.fudzilla.com/content/view/16544/1/
Good news for all, this may be the most anticipated GPU since the G80.
Printable View
http://www.fudzilla.com/content/view/16544/1/
Good news for all, this may be the most anticipated GPU since the G80.
Lets hope that Nvidia bring out some sub £200 parts within the same time frame too.
I heard rumours that it's quite a bit quicker than a Geforce3 Ti200, and it's effectively 40000000 riva tnt's nailed to a board ;)
Well trying to dive that deeply into gpu archeitecture when all we have atm is technical details, is headache induceing. :ill:
Well apart from the normal giveing it more of everything (512 streams, Gddr5 controller, full EEC memory access, higher clock speeds, etc) and reduceing the core architecture to 40nm.
What I can make out from all the technical talk is it's very good on paper, nvidia are using the dx11 progamable thread and their CUDA in a more coherient way.
They've broken the streams up into blocks of clusters of streams (the original G80 used a two block design, but the GT200 was 3 blocks) however it's redesigned to use double scheduling.
In laymans terms, before dx10 cards had a single list of jobs which got fed to each cluster of streams when they where free, this new design breaks the list into two which should in theory make it respond faster esp when doing lots of small jobs.
To do this they also had to greatly increase the L1 cache of each cluster and the L2 shared cache, at the same time at the same time the clusters also have free control over any oth the shared L2 cache, where as before the shared L2 cache was broken into blocks with each cluster only having access to it's own assigned blocks.
And yes that's horribly simplified.
The up shot of all that should mean mean a higher pixel fill bandwidth and far greater flexability to do more things at once and do them more effecently.
It's fmadd numbers are below the 5870 however the newer dpfma is much higher
AFAIK dpfma is what DX11 is based on
fmadd is multiply or add floating point calculation, dpfma is a combined multiply and add floating point calculation.
And yes that's more horrible simplification.
What all this means when the GF100 gpu comes out is that, on paper, it's very flexable and should be very fast for dx11, although not as good in dx10 or dx9 (but still good) and very good for non-graphics stuff and running non-graphics + graphics at the same time.
All told what it look like on paper means little untill we actually see the cards out in the wild.
The most promising point to this news is that nvidia seem to have sorted out the silicon problems they where having and are getting good enough yeilds to announce the release.
So, we may see them available to buy by Feb, March or possibly April then....
Hang on,
Let me get this straight; this is the 100 series? The last one was the 200 series made up with the budget rebranded 9000 which was origionally released in the 8000 series? and before that the 9000 series which was really the 8000s (which should have been the 9000 series but pushed out early) and before that we had the real 8000 series.
Are we counting down? Perhaps a lettering system would be better.
The first card of the Geforce 300 series has been released:
http://forums.hexus.net/hexus-net/17...ml#post1824797
Indeed, 210 rebranded is not a Fermi. I presume that means Nvidia intend re-branding the Fermi into a 400 series before a genuine DX11 card comes out near that price bracket :rolleyes:
My 8800GT is getting a bit old now, are the ATI Linux drivers stable yet? If they are, I could be tempted to abandon the Nvidia ship.
I would check this website out:
http://www.phoronix.com/scan.php?page=home