Some specs revealed in the upcoming GT212

Joined
Apr 1, 2008
Messages
2,783
gt212.png


http://en.expreview.com/2009/01/05/specs-of-gt212-detailed.html

Since ELSA revealed that GT212 would be another flagship card of NVIDIA, GT212 has been a focused topic. Being the top GPU this year, GT212 is coming to take the place of GT200(b) (such as GTX260/280) Series.

Our source indicates that the number of stream processers increases from 240 (GT200) to 384, and the number of TMUs also increases to 96. Such improvements are fairly significant, kind of like 128 SPs of G92 to 240 SPs of GT200. As the second-generation rendering structure, GT200 features 10 TPCs, and we’re wondering if GT212 would follow this structure. If so, GT212 will feature 16 TPCs, and there’re only 6 TMUs in each TPC, while there’re 8 TMUs in GT200. GT212 is very likely to adopt the third-generation rendering structure - each TPC contains four SMs and 8 TMUs, and there’re 12 TPCs, so there’re 12*4*8=384 stream processors in total.

Generally speaking, the stream processors of GT212 increase by 60%, but TMU just increases by 20%.

The memory interface of GT212 will reduce from 512bit to 256bit, which is quite similar to AMD RV770. GT212 will aslo use GDDR5 to make up the loss of memory interface. The memory clock frequency of GTX280 is 1107MHz. If the memory bandwidth of GT212 has to achieve that of GT200, its data rate must be above 4.5GHz.The new products are usually superior to previous ones in specification, so the data rate of GDDR5 is supposed to be around 5GHz. NVIDIA and AMD both use GDDR5 on their middle-end lineup, which makes us to worry about the productivity of GDDR5.

As to processing technology, GT212 undoubtedly features 40nm processing. Because of the sharp increase of stream processors, the number of transistors reaches 1.8 billion, while that of GT200 is 1.4 billion. Thanks to the new processing, the core die area is controlled within 300mm^2, and the power consumption can be also reduced greatly.

GT212 is scheduled to be launched in Q2 this year. As the coming flagship product from NVIDIA, it’s really worthy of our expectation.

So they are going to use GDDR5 and shrink the memory bus back to 256 bit. I'm more interested in the GT300 though.
 
Hmm...looks like the green team stole the red team's playbook.
 
I'm kinda disappointed they had to use a 256 bit memory bus at the expense of GDDR5. I like the specs except the 256 bit bus really bothers me.
 
Why, if and that is if they use gddr5 then why would they need to bother with the 512bit bus? No company actually wants to use that width of bus with all that it entails, faster memory with a smaller bus is the easier way to go.
 
I'm kinda disappointed they had to use a 256 bit memory bus at the expense of GDDR5. I like the specs except the 256 bit bus really bothers me.

Nothing to worry about, as long as it runs fast :D.
 
what would be the problem in using 256-bit bus?
it's only so good for bandwidth, and with GDDR5 there's plenty bandwidth to make up for the cut in the bus width.
 
It's cheaper/more cost effective to use 256bit than 512bit and with GDDR5 you'll get the same throughput.
 
Okay, re read what you wrote and i think i get what you were saying, didn't read right at first.
 
crappy 256 bit bus! :mad:

I'll wait till GT300 thanks for a nice 448 or 512 bit bus with GDDR5
 
With GDDR5 speeds I don't know that we'll need it... but with native resolutions of 2560x1600 it may prove useful.

Buying 55nm and waiting to see what happens with 40.
 
They're also dropping a load of GT1XX series cards, including a 1.5 gigabyte GT130. Which has me interested, if I can put that memory through good use with CUDA. Have it work on encoding and decoding or act as a super fast framebuffer for editing or something, anything that might ease the workload.
 
crappy 256 bit bus! :mad:

I'll wait till GT300 thanks for a nice 448 or 512 bit bus with GDDR5

Clearly if someone makes a 256bit part with their latest high end it's telling you that 256bit + DDR5 is enough bandwidth. Do you really think they're arbitrarily designing it as 256bit at the expense of performance? Clearly memory bandwidth isn't a bottleneck if they're designing it this way from the ground up.

Also, you do realize that ATI's 256 bit with GDDR5 gives the same bandwidth as the 512bit GT200 with GDDR3, right? Why do you want to bolster parts of the graphics card which will have no impact on performance by going to 512bit with GDDR5?

The type of memory is as much of a factor in bandwidth as the bus width. You can't judge cards off of such a simplistic factor as bus width... it's like trying to judge cards purely based on clocks today, when the architectures are so different that it's impossible to do a direct comparison purely based on clocks.
 
crappy 256 bit bus! :mad:

I'll wait till GT300 thanks for a nice 448 or 512 bit bus with GDDR5
oh good lord what a silly thing to say. 256-bit with fast 4400mhz gddr5 or 512-bit with 2200ghz gddr3 would net you the same thing. gddr5 speeds are going to get even faster and putting that with a 512-bit bus would be wasteful at this point.
 
Hmmm, I wonder what the next generation of shaders will have over the 280 gen? I know the shaders were re-written from 8-9 series to GT2xx series, maybe theyll do it again?

All I can say is CUDA could kick some ass in a 3 way SLI set up with these cards :)
 
There is a serious advantage to going to the 256 Bit bus and that is a higher shader clock speeds. We might see a 700mhz stock speeds which would make it an insanly fast card, we'll just have to wait and see though, should have black market cards hitting the chinese sites with benchies in 3 months.
 
There is a serious advantage to going to the 256 Bit bus and that is a higher shader clock speeds. We might see a 700mhz stock speeds which would make it an insanly fast card, we'll just have to wait and see though, should have black market cards hitting the chinese sites with benchies in 3 months.
why would a 256bit bus have much effect on overclocking?
 
It wouldn't.

Hmmm, I wonder what the next generation of shaders will have over the 280 gen? I know the shaders were re-written from 8-9 series to GT2xx series, maybe theyll do it again?

Shaders on G8x, G9x and GT2xx are all the same. But yeah, we should expect another big change with GT300.
 
why would a 256bit bus have much effect on overclocking?

Because with a smaller bus width you can run tighter timings to reduce the latency. With the wider bus you have to relax the timings to keep the latency in check. Or something like that. I've been out of the GPU architecture way too long. I remember it being cited as one of the main reasons the G200s weren't reaching the clock speeds they were hoping for, which in turn has been the real downfall of the GT200.
 
Because with a smaller bus width you can run tighter timings to reduce the latency. With the wider bus you have to relax the timings to keep the latency in check. Or something like that. I've been out of the GPU architecture way too long. I remember it being cited as one of the main reasons the G200s weren't reaching the clock speeds they were hoping for, which in turn has been the real downfall of the GT200.
I thought the size of the chip was what was holding back the cards from overclocking. so could the chip itself be smaller with a 256-bit bus? oh and IMO they overclock pretty good for such a huge 65nm chip. also Im a little confused about what timings you are talking about and how that would affect overclocking the core/shaders.
 
I thought the shaders on the GT series were re-designed? I read something like 40% more efficient, or was that all FUD?
 
I thought the shaders on the GT series were re-designed? I read something like 40% more efficient, or was that all FUD?
I think it was 50% and that turned out to be BS I believe. that was some crap that people made up after reading leaked info about the card. the card was about 50% faster but that was to be expected based on its specs.
 
I'll probably suspect this card will only be slightly faster than the 280/285 GTX only, but who knows, while its being limited by that stupid 256 bit memory bus, its like 9800 GTX all over again. Pass.

I'll wait for the GT300 which hopefully has a 512 bit and GDDR5.
 
I'll probably suspect this card will only be slightly faster than the 280/285 GTX only, but who knows, while its being limited by thatu an stupid 256 bit memory bus, its like 9800 GTX all over again. Pass.

I'll wait for the GT300 which hopefully has a 512 bit and GDDR5.

Please read what the majority of posters have already said. This card will not be limited by memory bandwidth with a 256 bit memory bus on DDR5... Just because the numbers are larger doesn't mean in the given situation they will always be better...

This card should at least out perform the 295 and can you imagine a GX2 version of this card? *drools just thinking about*
 
Please read what the majority of posters have already said. This card will not be limited by memory bandwidth with a 256 bit memory bus on DDR5... Just because the numbers are larger doesn't mean in the given situation they will always be better...

This card should at least out perform the 295 and can you imagine a GX2 version of this card? *drools just thinking about*
I dont see it beating the GTX295 unless it has insane clocks and even then thats doubtful. it should be around 50% better or so then the GTX285 though.
 
Apparently people read the memory bus without truly understanding what it does, and call it off. First of all, these are speculations. Second of all, did anyone forget the 4870 is GDDR5 256-bit? What are ya complaining about if the performance is the same?
 
Because with a smaller bus width you can run tighter timings to reduce the latency. With the wider bus you have to relax the timings to keep the latency in check. Or something like that. I've been out of the GPU architecture way too long. I remember it being cited as one of the main reasons the G200s weren't reaching the clock speeds they were hoping for, which in turn has been the real downfall of the GT200.

If that is true (and I have no idea if it is), then it would only affect the memory clocks. The shaders aren't tied to the 512bit bus. I think what you heard was just some people making crap up.

Shaders on G8x, G9x and GT2xx are all the same. But yeah, we should expect another big change with GT300.

No they aren't. The GT2xx made some changes to the shaders. One of which was doubling the register size per SP, which resulted in a ~10% gain in performance (according to nvidia)
 
I think this card will beat GTX295. With 40nm and 300mm^2 it will probably clock around 800-1000Mhz clock core.
 
I think this card will beat GTX295. With 40nm and 300mm^2 it will probably clock around 800-1000Mhz clock core.
you can probably forget about 1000mhz being realistic but who knows. all we can do is guess at this point but I will go on record as saying this card will not beat the gtx295. :p
 
I'm tired, I may be on crack, but I remember reading something about smaller bus sizes interacting somehow to cause slower shader clocks. Can't find it now, I'm probably on crack. :(

As for the idea of 1 Ghz clock someone posted, you really are doing crack if you think you're going to see a 50% increase in shadder clock speeds from flagship to flagship.
 
No they aren't. The GT2xx made some changes to the shaders. One of which was doubling the register size per SP, which resulted in a ~10% gain in performance (according to nvidia)

The register file and shaders are not the same thing. The larger register file simply allowed the ALU's to perform closer to peak in longer more complex shaders as it allowed the chip to have a sufficient number of threads in flight to properly hide latency. If you were to compare G80 and GT200 on a simple shader you would see no benefit to the larger register file.

Another improvement was more effective scheduling of the MUL on the SFU unit. On G8x,G9x MULs that could potentially run on the SFU were still scheduled primarily on the main ALU.

But the shaders themselves did not change - they're still the same old MAD + SFU/MUL that we've come to know and love/hate. Same with RV770 by the way, it's the same shaders there that we were introduced to with R600. It's the stuff feeding them that's been changed up a bit.
 
The register file and shaders are not the same thing.

Of course they aren't. Shaders HAVE registers, just like CPUs have registers.

It's the stuff feeding them that's been changed up a bit.

The registers have nothing to do with the feeding of the SP. I would argue that the SP changed due to the doubling the size of the registers, just like I would argue that a CPU changed if you doubled the number of CPU registers. The processing aspect of the SP may not have changed, but the registers are definitely a part of the SP, and they changed, so therefore the SP changed.
 
The registers have nothing to do with the feeding of the SP.

Really? So where does the GPU store state for inactive or stalled threads? The register file is not part of a shader. That's like saying L1 cache is part of the floating point pipeline in a CPU.

The processing aspect of the SP may not have changed, but the registers are definitely a part of the SP, and they changed, so therefore the SP changed.

Take a look at R600.
 
Shaders HAVE registers, just like CPUs have registers.

This is where you're not being consistent. Shaders on a GPU are analagous to execution units on a CPU. You can't equate a shader to a CPU. A CPU has execution units and registers/cache and a GPU has shaders and registers/cache. If you want to refer to shaders as the all encompassing processing unit then you would be referring to one of the clusters. In which case GT200 would only have 30 "shaders" based on that definition.
 
Back
Top