NVIDIA GF100 & Friends speculation

It seems the first iteration of GF104 will really have 336 CUDA Cores, with clocks a little higher than GTX465, if the following image is not shoped. Performance is said to be around GTX465 levels, with good overclocking potential. Power consumption is also said to be about 150W, which is IMO good (not excellent, though) considering that GTX465 consumers around 215W.

 
It seems the first iteration of GF104 will really have 336 CUDA Cores, with clocks a little higher than GTX465, if the following image is not shoped. Performance is said to be around GTX465 levels, with good overclocking potential. Power consumption is also said to be about 150W, which is IMO good (not excellent, though) considering that GTX465 consumers around 215W.
There has to be some sort of mistake there - sure, they could cut the size of multiprocessor to 16 from 32 "cuda cores", but 21 units? That just doesn't make sense at all.
Or then it won't be a full chip, but 22 "cores" sounds strange, too, and 24 sounds a little bit too high
 
There has to be some sort of mistake there - sure, they could cut the size of multiprocessor to 16 from 32 "cuda cores", but 21 units? That just doesn't make sense at all.
Or then it won't be a full chip, but 22 "cores" sounds strange, too, and 24 sounds a little bit too high
What about 48 cores per ShaderModule instead of 32?

2 GPCs, with 4 SMs each for a total of 8, of which one is disabled for yield reasons (at least on the SKU shown in the screenshot).
 
There has to be some sort of mistake there - sure, they could cut the size of multiprocessor to 16 from 32 "cuda cores", but 21 units? That just doesn't make sense at all.
Or then it won't be a full chip, but 22 "cores" sounds strange, too, and 24 sounds a little bit too high

Thats why i said first itineration. GF104 is supposed to be made of 384 "cuda cores", with the higher variant launching a bit later with a 256 bit memory bus and 1GB RAM.
 
Thats why i said first itineration. GF104 is supposed to be made of 384 "cuda cores", with the higher variant launching a bit later with a 256 bit memory bus and 1GB RAM.

edit: just realized I calculated things wrong :D
But ye, that's a possibility, even though it sounds a bit high for middlerange variant
 
edit: just realized I calculated things wrong :D
But ye, that's a possibility, even though it sounds a bit high for middlerange variant

Thats because GF104 probably is not intended to be a middlerange variant? GeForce 8800GT ring a bell? It seems like its GF106 job to go against 5770...
GF104 seems more and more like Fermi done right, by the day.. Probably too little, though, with Southern Islands around the corner..
 
Thats because GF104 probably is not intended to be a middlerange variant? GeForce 8800GT ring a bell? It seems like its GF106 job to go against 5770...
GF104 seems more and more like Fermi done right, by the day.. Probably too little, though, with Southern Islands around the corner..

But it's not enough to compete in high(est) end either, so the "fermi done right" doesn't really fit either :???:
 
But it's not enough to compete in high(est) end either, so the "fermi done right" doesn't really fit either :???:

What I mean with "Fermi done right" is in relation to the better perf/watt. If full GF104 hits GTX470 levels, with lower TDP and temps, that is Fermi done right in my book, as it can be the architecture revision for GF100 follower :D
 
That's kind of more what I was expecting. Even if the performance hasn't changed that much, though, if the power consumption has dropped significantly, I think it could be a much better buy for many.
 
That's kind of more what I was expecting. Even if the performance hasn't changed that much, though, if the power consumption has dropped significantly, I think it could be a much better buy for many.

Did anyone do some kind of perf/shader analysis? How much more performance could 32 "cuda cores", negligencing the 20Mhz increase, give?

32 "cuda cores" are around 6% of 480. With the 20 Mhz increase, probably we will look at around 5-10% more performance? Is it really worth it, if temps and TDP is same? o_O
 
Did anyone do some kind of perf/shader analysis? How much more performance could 32 "cuda cores", negligencing the 20Mhz increase, give?

32 "cuda cores" are around 6% of 480. With the 20 Mhz increase, probably we will look at around 5-10% more performance? Is it really worth it, if temps and TDP is same? o_O

According to that link, there's no increase in memory bandwidth. So (512/480) × (1440/1401) = 1,096, or +9.6%. That would be the theoretical maximum, possibly limited by the memory bandwidth. In practice, I'd expect about +5~6% FPS. In other words, nothing you'd really notice in actual gameplay.
 
There has to be some sort of mistake there - sure, they could cut the size of multiprocessor to 16 from 32 "cuda cores", but 21 units? That just doesn't make sense at all.
Or then it won't be a full chip, but 22 "cores" sounds strange, too, and 24 sounds a little bit too high
You still believe *16 is more real than *48 ?
 
Damien asked a few questions to ATI and NVIDIA's partners: http://www.behardware.com/news/lire/15-06-2010/

With respect to the Radeons, more than 8 months after launch, availability is still problematic.[...] On the other hand and in contrast to what the rumour was telling us, the GeForce GTX 400s are widely available according to the various partners we talked to in Taipei. Sales are however disappointing with partners highlighting the high energy consumption and noise levels, as well as the price/performance ratio as being the major factors keeping customers away.
 
Back
Top