NVIDIA GT200 Rumours & Speculation Thread

[Ailuros]

Like the "65nm R600" supposedly hitting +900mhz.
And then RV670 supposedly clocking +900mhz.
BTW- Roughly the same GPU but different rumors with quite a bit of time between them.

I wasn't aware about a 65nm/R600 rumour of the past; anyhow many RV670 cores are obviously capable of clocking =/>900MHz and it shows from the overclocks many users can yield. It doesn't surprise me one bit, since the chips on the X2 are clocked by AMD at 825MHz.

Anyhow back on GF-Next: while I could easily accept that a G92b could reach something like =/>750/1875, I wouldn't be so sure that GF-Next with a much higher core complexity would even come close to those kind of frequencies. If I take the ALU domain: TMU/ROP domain ratio on G92 it's at 2.5x times the frequency; if you take that as granted you'd get for 2.4GHz on the ALU domain, 960MHz for the TMU/ROP domain. That sounds even too much IMHLO for a 55nm/G92.

 

[Jawed]

Also this:

5.1.2.2 Constant Memory ... The cost scales linearly with the number of different addresses read by all threads. We recommend having all threads of the entire warp read the same address as opposed to all threads within each of its halves only, as future devices will require it for full speed read.

It does sound like the warp size will stay constant at 32 "threads" but shared memory banks will double to 32, and the 2x8 ALU + 2 interp ALU per cluster configuration might change to something like a 2x16 + 2 interp setup.

I think they could double operand fetching to 32-wide while keeping the ALUs the same. This widening would simply save on the logic that supports 16-wide reads at twice the rate.

In other words I don't think this provides any evidence that the ALU configuration per multiprocessor has changed.

That would keep the cluster count for a 256ALU GPU at 8 (if the interconnect is a crossbar which I thought it was, then scaling up to double the number of ports might be an issue, no?).

I'm cautious, too, about the size of the crossbar between the clusters and the ROP/MC partitions. Keeping it at 8 clusters seems reasonable. Particularly if there's 8 partitions for a 512-bit bus.

The control logic would have to be beefed up to deal with the need to issue 2x the number of instructions per cycle (and probably also track more warps).

If you double the width of a multiprocessor that doesn't necessarily double the number of instructions being issued.

Indeed NVidia could halve the instruction issue rate per multiprocessor by going to 4-clocks per instruction - G80...G92 are 2-clocks per instruction, but only for vertex shaders (and GS?).

Alternatively a 16-lane multiprocessor would require 2-clock instructions, retaining the baseline instruction issue rate of G80.

Texture power (per-clock per-cluster) would remain where it is now unless the TMUs are also widened...

I agree, for what it's worth.

Jawed

 

[aaronspink]

Ugh, can you imagine always having to explain the difference between Nvidia and AMD "processors" if that were to happen? I think referring to the number of ALUs is a far lesser evil! :smile:

I have no problem with them referring to ALUs! Just don't call ALUs processors!

Aaron Spink
speaking for myself inc.

 

[aaronspink]

http://translate.google.com/transla...d-5051-1-1.html&langpair=zh|en&hl=en&ie=UTF-8

Another GT200 rumour. 512-bit MC and OVER 200SP so maybe these rumours of 240SP are correct. I believe that GT200 SPs is stronger than G80s/G92s when we compare it 1:1.

I'll believe 240 SP when they actually show the 1600 mm^2 die. Likely its 20 or 24 SPs just like G80/92 is 16 SPs. Once again ALUs != SPs!.

 

[trinibwoy]

I'll believe 240 SP when they actually show the 1600 mm^2 die. Likely its 20 or 24 SPs just like G80/92 is 16 SPs. Once again ALUs != SPs!.

You weren't kidding huh?

 

[aaronspink]

You weren't kidding huh?

nope! if we are going to the future where we are going to be using GPUs for anything except just bog standard graphics it IS important and does make a difference. Plus it pisses me off when people co-op already defined and standardized terms for marketing purposes.

Also it does make a difference in evaluations architectures because as you increase the width of the SIMD the effective efficiency of the SIMD units across the workloads decreases. OTOH, most of the workloads have very little thread interdependence and therefore get very good parallel instruction scaling.

It would be interesting if it was actually possible to get efficiency info for both the G80/92 architecture and the R6xx architecture.

Also its easier to scale up an architecture already designed for a large number of processors with a reasonable interprocessor interconnect than it is to keep scaling instruction schedulers, compilers, and SIMD width. This is both from a physical design standpoint and a logic design standpoint.

Aaron Spink
speaking for myself inc.

 

[Jawed]

I guess we could just co-opt SP to mean Single Precision (MAD)

Jawed

 

[Jawed]

It would be interesting if it was actually possible to get efficiency info for both the G80/92 architecture and the R6xx architecture.

If you look at the GPGPU work being done you can see people are still struggling to understand the parameters that determine performance/efficiency for something apparently simple:

http://forums.nvidia.com/index.php?showtopic=47689&st=40

it's taken over a year with the combined efforts of NVidia and some clever people to get SGEMM working "well", ~200GFLOPs, on 8800GTX. There's quite a few useful insights in that thread.

Jawed

 

[2900guy]

is it really that hard for nvidia to released something that will be to the 8800 GTX what the 8800 GTX was to the 7900 GTX? sinlge chip 2x+ faster than 8800 GTX.

 

[Chris123234]

is it really that hard for nvidia to released something that will be to the 8800 GTX what the 8800 GTX was to the 7900 GTX? sinlge chip 2x+ faster than 8800 GTX.

I think the problem is that AMD has been unable to do that.

 

[Rangers]

I think the problem is that AMD has been unable to do that.

So, Intel wont release Penyrn cus AMD has nothing good also?

Or, why did Nvidia release G70 long before R520 (doing immense damage to AMD btw)?

Totally bonkers, Nvidia is out to destroy AMD any way they can. If they had something faster, it'd be out the door.

I mean really, what would it even hurt? If R&D continued apace..

 

[Mart]

I think the problem is that AMD has been unable to do that.

Let's not forget that the GPU market isn't really an expanding market. Having the upper hand doesn't mean you keep a steady stream of money, it has to dry up sometime when the market is saturated with the current generation product, even if it's vastly superiour to the competition.
There are a lot of people with an 8800GTX or Ultra in their machines that are desperate for a new graphics card to double their performance. I would certainly replace mine as soon as I could, if the new card is > 1.5 times faster. Thus far, all we got is a lousy shrink with higher clocks. I'm not going to spend money on that, it's only marginally faster than what I've got. I think a lot of people are thinking along the same lines.

 

[Jawed]

Yeah, anyone who's still in denial about the lateness of GT200 should look at the official information posted here:

http://forums.nvidia.com/index.php?showtopic=36286

Does CUDA support double precision arithmetic?

CUDA supports the C "double" data type. However on G8x series GPUs, these types will get demoted to 32-bit floats.

You should be careful to specify float constants in single precision (i.e. "1.0f") so that they will not accidentally be compiled to double precision on future hardware.

NVIDIA GPUs supporting double precision in hardware will become available in late 2007.

So, forget the "NVidia doesn't need to compete" baloney.

Jawed

 

[Razor1]

Yeah, anyone who's still in denial about the lateness of GT200 should look at the official information posted here:

http://forums.nvidia.com/index.php?showtopic=36286

So, forget the "NVidia doesn't need to compete" baloney.

Jawed

That doesn't tell us anything, the gt200, wasn't even talked about at the time of that post.

I agree with mart about saturation of the market, but ......

How much is the cost of ramping up a new chip? What is the offset of not ramping a new chip vs.? What is the potential market for the new chip vs.? What is the possibility of cost savings from a newer process vs older process?

Lots of things to look at when there is no competition.

 

[Sound_Card]

Yeah, anyone who's still in denial about the lateness of GT200 should look at the official information posted here:

http://forums.nvidia.com/index.php?showtopic=36286

So, forget the "NVidia doesn't need to compete" baloney.

Jawed

Jawed, I believe that dates to 2007.

And the products that came out in late 2007 for Nvidia I believe was the Tesla's, and G92's which I believe support double precision to some degree or fully. Errr, but I'm not sure.

 

[Rufus]

is it really that hard for nvidia to released something that will be to the 8800 GTX what the 8800 GTX was to the 7900 GTX? sinlge chip 2x+ faster than 8800 GTX.

Umm, yes?

Four years and 400 million dollars in the making, NVIDIA G80 represents for the company their first brand new architecture with arguably no strong ties to anything they've ever built before.

 

[Arun]

Umm, yes?

Pfff, $400M... NV makes that in net profit every quarter nowadays! Why can't they spend that money on R&D instead of adding to shareholder value? Oh wait... (and I figure 2900guy probably meant in terms of performance, not architecture)

 

[Freak'n Big Panda]

That doesn't tell us anything, the gt200, wasn't even talked about at the time of that post.

I agree with mart about saturation of the market, but ......

How much is the cost of ramping up a new chip? What is the offset of not ramping a new chip vs.? What is the potential market for the new chip vs.? What is the possibility of cost savings from a newer process vs older process?

Lots of things to look at when there is no competition.

NV would most definitely release GT200 if it was ready. Not having competition would not alter that. What it would affect is the price of the chip which would probably be @ 600+ if there was nothing from ATI to compete with it. Hell they could price it at 900+ if they didn't want to cannibalize 9800GX2/9800GTX sales and just rake in huge profits on each board sold. Then when ATI did come out with something to compete with they could drop the price accordingly. I don't see any reason for them not to release it as soon as possible.

Anyways to address your questions directly:

How much is the cost of ramping up a new chip?

It wouldn't go down over time, and they'd have to do it eventually regardless.

What is the potential market for the new chip

small but not totally insignificant. There's also something gained when you absolutely dominate the competition. Halo effect and all.

What is the possibility of cost savings from a newer process vs older process?

With an undisputed performance leadership NV would have the freedom to price GT200 at whatever point they wanted too.

 

[Kowan]

Let's not forget that the GPU market isn't really an expanding market. Having the upper hand doesn't mean you keep a steady stream of money, it has to dry up sometime when the market is saturated with the current generation product, even if it's vastly superiour to the competition.
There are a lot of people with an 8800GTX or Ultra in their machines that are desperate for a new graphics card to double their performance. I would certainly replace mine as soon as I could, if the new card is > 1.5 times faster. Thus far, all we got is a lousy shrink with higher clocks. I'm not going to spend money on that, it's only marginally faster than what I've got. I think a lot of people are thinking along the same lines.

Agreed. I'm using SLI'd GTXs and waiting to upgrade.

 

[Jawed]

That doesn't tell us anything, the gt200, wasn't even talked about at the time of that post.

How is "GT200" relevant?

I agree with mart about saturation of the market, but ......

How much is the cost of ramping up a new chip? What is the offset of not ramping a new chip vs.? What is the potential market for the new chip vs.? What is the possibility of cost savings from a newer process vs older process?

Lots of things to look at when there is no competition.

LOL a load of excuses like this come up each time when NVidia is not executing as well as its fans like to pretend it always does.

Jawed