Like the T in T-buffer? (My one and only guess. Only wasted enough time to type this message. :smile
NVIDIA GT200 Rumours & Speculation Thread
- Thread starter Arun
- Start date
- Status
Like the T in T-buffer? (My one and only guess. Only wasted enough time to type this message. :smile
Another hommage to senior Tarolli? LOL
Sound_Card
Regular
You would be wrong then. Every review(think back to 2006) refered to the shaders in G80 as "stream processors" because they were told by Nvidia to call them that. Errr... so just no... plain no....
http://www.elitebastards.com/cms/in...sk=view&id=228&Itemid=27&limit=1&limitstart=2
Twice as fast??? That's hardly reasonable, besides, amd/ati are the ones who need to get on the ball, they still couldnt put out a single gpu solution to best the G80 after all this time, needed 2 GPUs to do it, and at the price of higher power draw and inconsistent performance.
When ATI stops dicking around with their current seemingly inefficient architecture nvidia will surely retaliate
Interesting link, but I don't see any reference where they count it as 256 SP's. (IOW: please reread what I wrote. Carefully.)
aaronspink
Veteran
As long as ATI/AMD is behind in performance, Nvidia don't care to release something faster. "That is how the game of War - works"
aaron, the problem isn't the word 'processor'; the problem is the word 'stream'. A stream is, by definition, coherent (especially in the context of computing); on dictionary.com, one of the definitions is 'a continuous flow or succession of anything'.
The implication that a stream processor would simply be processing a continuous flow of data, rather than require control logic per piece of data, isn't a new concept. The term has been employed in Academia projects substantially before NVIDIA released the G80 AFAIK - look at the examples at http://en.wikipedia.org/wiki/Stream_processing if you don't believe me.
I agree it's bad terminology because of the common understanding of the word processor as refering to something with dedicated control logic for instruction handling, but that doesn't mean it's fundamentally incorrect or that NVIDIA came up with the idea of calling it that way.
The implication that a stream processor would simply be processing a continuous flow of data, rather than require control logic per piece of data, isn't a new concept. The term has been employed in Academia projects substantially before NVIDIA released the G80 AFAIK - look at the examples at http://en.wikipedia.org/wiki/Stream_processing if you don't believe me.
I agree it's bad terminology because of the common understanding of the word processor as refering to something with dedicated control logic for instruction handling, but that doesn't mean it's fundamentally incorrect or that NVIDIA came up with the idea of calling it that way.
aaronspink
Veteran
I know all about Dally et al and stream processors, he'd given plenty of talks about it at DEC SRC and DEC WRL and several coworkers had been previous students of his. And by their definition G80/92 would be 16 stream processors.
Um, YES IT IS in the number context they are using. They are referring to a single ALU as a stream processor. This is BS of the first order. Do some research (it is somewhat arguable whether they should be able to call their design a 'stream processor', period, but we'll save that debate for a later time...).
At best you could describe the G80/G92 design as being a chip consisting of 16 stream processors.
And as I'll always point out, 'Stream' is an adjective to identify a sub-type of processor. The word processor is authoritative in the definiton and therefore any specific definition of a 'stream processor' much fit within the definition of processor. By definition, 'stream processor' implies an instruction pointer per instance.
If you really really really want to push it, we can get into the various reasons why G80 ISN'T a stream processor or at the very least contains at most 1 stream processor contains 16 execution clusers.
Aaron Spink
speaking for myself inc.
If you really really really want to push it, we can get into the various reasons why G80 ISN'T a stream processor or at the very least contains at most 1 stream processor contains 16 execution clusers.
Regardless of whether or not Arun wants to push it, I'd be interested - is this essentially because everything starts with a (single) stream of geometry data from which other streams (vertices/fragments) are derived? I.e. would a "16 stream processor" die have to be able simultaneously process 16 completely independent geometry streams?
[And if so - I didn't realize Imagine/Merrimac had that capability. I always thought of them as a single stream processor with multiple VLIW execution clusters.]
aaronspink
Veteran
One of the staples of a stream processor architecture is that the computational clusters pass the stream between them in a serial method in order to leverage time domain parallelism and locality.
For example in the canonical Stanford model assuming 4 clusters, cluster A is the start of the stream processing and fetches the initial ops and data from memory, applies function F1 and passes the results to cluster B which fetches any additional data it needs from memory applies function F2 and passes the results to cluster C. Cluster C fetches any additional data it needs from memory applies function F3 and passes the results to cluster D. Cluster D fetches any additional data it needs from memory applies function F4 and writes the results to memory or IO device.
The actual operational capabilities of the clusters is variable but generally they are capable of each executing an independent instruction stream in order to accomplish their function. In the case of something like video, one cluster would do input stream demux, another would do motion estimation, another would do audio decode, another idct, etc.
And yes imagine/merrimac are single stream processors. Which would be the category that the G80/G92 most closely resembles except the G80/G92 lack the unit to unit (or cluster to cluster) streaming that is most critical to the performance of a stream processor because of the ability to bypass memory for results. From a purely functional perspective G80/G92 more closely resembles something like a special purpose multi-core processor than a pure stream processor.
Effectively, G80/G92 is as much a stream processor as a Core2 Duo and Core2 Quad is a stream processor, they are certainly capable of executing a program written using a stream processing programming model and run through appropriate compilers but that is really just because they are capable of emulating the functionality using memory and are essentially turing machines. Cell more closely resembles a stream processor due to the SPE to SPE DMA capabilities than either G80/G92 or Core2.
Aaron Spink
speaking for myself inc.
Each SP has its very own instruction pointer, among other things.aaronspink said:
Except G80 does have cluster-to-cluster streaming. It's not because it's not exposed in CUDA that it's not there. You also can't access the rasterizer in CUDA, but that doesn't mean there isn't one.
aaronspink
Veteran
DerekBaker
Newcomer
Is this the launch of GT200?
http://www.theinquirer.net/gb/inquirer/news/2008/04/02/nvidia-throws-party
http://www.theinquirer.net/gb/inquirer/news/2008/04/02/nvidia-throws-party
TimothyFarrar
Regular
Kyle thinks G92b is on the way. He also proposes that the 9900 series is a contingency plan in case RV770 brings the pain. All this sounds to me like whatever they're cooking up for a next generation part is either borked, broken or incredibly late.
http://www.hardforum.com/showpost.php?p=1032310319&postcount=1
http://www.hardforum.com/showpost.php?p=1032310319&postcount=1
DegustatoR
Legend
9900 cards should use G92b, it's really simple to figure out logically if you have any idea about the size of G1xx chip family.
The plan as i see it is to exchange all the G92s with G92bs in the current GF9 line-up plus maybe to introduce a faster clocked G92b based card to counter RV770 -- which should use the GF9900 moniker.
So it's not really a contingency plan against RV770, but more of a side effect of 'shrinking' all of G92 cards for higher profits.
I believe that G100 is a much faster and costier chip which will compete with RV770 X2 (if there can be such a thing as competition between AFR solution and 'normal' hi-end videocard). Wether it's late or not is really hard to say until Q3 and some final specs showing up -- they might allow us to figure out if this G100 is the same top-end chip which was supposed to be released in the end of 2007 or not. If not then it will be hard to say anything about it's schedule.
The plan as i see it is to exchange all the G92s with G92bs in the current GF9 line-up plus maybe to introduce a faster clocked G92b based card to counter RV770 -- which should use the GF9900 moniker.
So it's not really a contingency plan against RV770, but more of a side effect of 'shrinking' all of G92 cards for higher profits.
I believe that G100 is a much faster and costier chip which will compete with RV770 X2 (if there can be such a thing as competition between AFR solution and 'normal' hi-end videocard). Wether it's late or not is really hard to say until Q3 and some final specs showing up -- they might allow us to figure out if this G100 is the same top-end chip which was supposed to be released in the end of 2007 or not. If not then it will be hard to say anything about it's schedule.
- Status
