I remember reading someone's thoughts in much of the same vein back in 2006 about GDDR4... and we all know how that turned out.
Nvidia GT200b rumours and speculation thread
- Thread starter nicolasb
- Start date
I remember reading someone's thoughts in much of the same vein back in 2006 about GDDR4... and we all know how that turned out.
I remember reading someone's thoughts in much of the same vein back in 2006 about GDDR4... and we all know how that turned out.
GDDR4 was expected to scale much further than it did. Technical limitations turned out an underwhelming product which resulted in its low adoption rates.
Cookie Monster
Newcomer
They got sued by one and only Rambus. Yes its that time of the month again.
I remember reading someone's thoughts in much of the same vein back in 2006 about GDDR4... and we all know how that turned out.
I don't think you can really say that differences between GDDR3 & 4 are anything like the same as between GDDR3/4 and GDDR5.
GDDR5 provides much, much more bandwidth than either of these two earlier technologies. NV will simply have to use it if they are to remain competitive in the longer term.
Or they postponed the entire idea until a smaller manufacturing process would had been available, because it became quite some time ago bleedingly obvious that the result would had been way too large for anything 65 or 55nm.
AMD probably postponed RV870 for 40nm for the same reason.
Well, i don't know how it was really called but that's the chip that should came instead of GT200.
Jawed
Legend
Which would imply that GT200 wasn't late, per se. So as well as the performance of this "G100" being strangled by bandwidth, NVidia was also desperate to include double-precision. The two reasons in combination giving rise to GT200.
Maybe NVidia then rushed to produce GT200. Perhaps it was originally a GDDR5 design aimed at releasing "now", but it was pulled forwards and sacrificed GDDR5. Being forced to implement extra MCs and ROPs, NVidia had to cut cluster count.
Jawed
Maybe NVidia then rushed to produce GT200. Perhaps it was originally a GDDR5 design aimed at releasing "now", but it was pulled forwards and sacrificed GDDR5. Being forced to implement extra MCs and ROPs, NVidia had to cut cluster count.
Jawed
There is another rumour =)
According to it this 384 SPs chip was in developement concurrently with GT200 but while GT200s target was Quadro/Tesla/CUDA market, this chip was intended for the gaming 3D market. But then at some point they've decided that GT200 is better anyway -- for the reasons i've posted earlier. That fits quite nicely with what we have now from any point of view.
As a result of that i'm expecting that GT212 (or 216? still can't remember which is which =)) will probably be this 384 SPs chip shrinked to 40nm with additions in the form of FP64 support and GDDR5 memory controllers (256- or 384-bit wide). But that's pretty much what everyone expects right now i suppose.
Well, they always was. But that doesn't mean that they always will be =)
Sure, but what has changed this gen to cause this paradigm shift?
FP64 precision.
In games it's not only unnecessary at the moment, but it's also relatively expensive, transistor-wise.
But in dedicated GPGPU applications it is very much a basic feature, that is, if you want your chip architecture to be taken seriously by the scientific community.
So, it makes sense to develop a chip with it for GPU computing, and another chip without it for high-end gaming.
Most GPGPU software apps would still work in consumer FP32-only hardware, but the speed penalty would further entice the sales and paid post-sales customer support of higher end (and much more lucrative, due to higher average selling price) equipment.
I think it would make sense to develop a GT2xx or GT3xx without FP64 units in order to spend those transistors improving performance in 3D games and new image quality features instead.
Yes, you are right on that.
No, AFAIK clearspeed is the only one making HPC specific chips. The market (professional/HPC) is too small to be sustained by specially tailored chips. ASPs on Quadro line are already v high and Tesla is a growing, but nascent market and yet they don't make special chips for them.
They already have plans for that. If you go through the CUDA docs, you will find a gap in compute capability (nv's number given to each card so that their capabilities can be versioned and documented). All 8xxx and 9xxx cards have compute capability 1.0 or 1.1. GTX280 and 260 have compute capability 1.3 and all DP enabled devices have compute capability 1.3 or more. I reckon that the gap at 1.2 will be filled by future mid-range/lowend derivatives of gt200.
You may see this as an investment they've tried to make with GT200 being so GP/CUDA-oriented. So yes it's small but pretty much everybody expects it to grow significantly in the nearest future.
And FP64 isn't the only reason for making a separate chip for serious CUDA calculations -- another one as you've said is significantly higher ASPs of that market which theoretically allow you to make much more complex chips and boards.
You may see this as an x86 mainstream vs IA64/POWER supercomputers situation. I think it's inevitable that GPUs will eventually came to similar market separation. Wether NV is ahead of that time with GT200 is another interesting question =)
Yet RV770 offers significantly higher DPFP performance...
Too bad there's no Top 500 supercomputer based on it yet...
the point being is that this is not necessarily something that you need to add significant area for, detracting the gaming performance.
