Article: Tesla 10-Series Analysis [Part 1]

Okay, I think this will be easier to understand with real chips. Note that the DP numbers per SM for these chips are mostly speculation based on my understanding of NV's strategy:

GT200: 1DP/SM.
GT206: 0DP/SM.
iGT206: 0DP/SM.
iGT209: 0DP/SM.
GT212: 2DP/SM.
GT214: 1DP/SM.
GT216: 1DP/SM
GT218: 1DP/SM.
iGT21x: <=1DP/SM.

 

I think in the end they will tie their SP multipliers together for DP ... DP in the G200 was just a quick hack (which was unfortunate on a chip which ended up way late anyway).

 

Too expensive? A $650 flaghship 576 mm^2 flagship isn't too expensive? Bleeding marketshare isn't too expensive? They can afford a seperate design for the low end where margins are low, but not Tesla where they can charge a premuim... please.

And if they include 2 DP/SM they can throw graphics perf/mm^2 out the window.

So the customer uses different ratios of FP64:FP32 chips based upon their need.

 

There's no evidence of "bleeding marketshare" yet. Hold your horses...
Also, how many low-end Tesla's or Quadro FX'es, or Quadro Plex'es do you know ?

The only low-end Quadro's are from the NVS line, which isn't really focused on 3D performance, but big-business driver reliability, 2D visualization, etc.

 

It is inevitable given the current situation. Sure, the average consumer isn't very aware of these things and takes a while to come around... but it will happen.

I fail to see the point here.

 

Nvidia survived R300 and R4xx (where even their high-volume, low-margin Geforce FX parts were substantially weaker than nowadays), so this high-end stuff, coupled with Tesla sales, etc, will probably off-load the cost of it.
It's not like GT200 was ever meant to last in its present high-end 65nm form for too long, or that a G92b is more expensive to make than a RV770...

The point being that it makes sense to have a separate low-end chip for graphics (where volume counts, and every penny saved on chip production means a lot to the bottom line when you multiply it by the millions of units produced at TSMC and UMC), but it doesn't make sense to have a high-end chip for graphics, and another substantially different for FP32/FP64 GPGPU (both segments are low-volume niches, despite being very profitable).

 

I have to disagree... It is much better to have 2 superior products than 1 inferior product IMHO. The worst thing you could do to a low-volume market is make it even smaller by losing market share.

G92b doesn't have any DP.

 

dedicated DP

I don't like the dedicated double precision decision. Being a CPU oriented guy I would have liked to see IEEE754 compliant single and double precision hardwares.

It is just a matter of time until we get these standard compliant vector math units.

 

You're conveniently ignoring volume.

Just make a business case for it: estimate NRE, production cost and volume. Don't forget to include the opportunity cost of a whole team not being able to work on the next gen product.

Should be interesting.

 

As opposed to margins? Selling chips at a loss is always nice...

Presently, nvidia is fortunate to be able to leverage CUDA, but that will not always be the case.

 

Margins based on manufacturing cost may be high but if the Tesla line is allocated the full cost for a chip those margins won't look so hot. Right now Tesla gets a free ride off the Geforce line.

 

Who says they're selling the low end stuff at a loss?

Why not?

 

But if the GTX isn't making any (or very little) $$$ then where does the free ride come in?

If they keep the $650 pricepoint that will hurt their sales, if they lower the pricepoint that will hurt their margins. Based on performance the GTX280 really shouldn't be priced any higher than $499. Nvidia can't really get away with charging whatever they want anymore as AMD is competitive again. However, in the GPGPU market they still don't have much (any?) real competition, but that is going to be changing soon.

The FP64 units in the GTX 280/260 are OK for the present market (no competition), but it looks like a disastrous plan for the future to me.

 

I certainly didn't...

Larabee won't be using CUDA and AMD hasn't exactly warmed to it... where is Cg now?

 

Cg is just a shading language for DirectX, OpenGL and the PS3 (it works on ATI GPU's too, i think), very similar to Microsoft's HLSL.
It just so happens that some GPGPU software could be coded with it, running within DirectX/OpenGL.

CUDA is an API built from the ground up for GPGPU software applications, and it will run simultaneously -but not on top of- with the DirectX or OpenGL API's (CUDA's newest compiler can even target x86's SSE2/3 Floating Point units, so it will run on "Larrabee" with minor modifications).
Didn't you read the article that originated this very thread ?

 

The GTX 280 is just one chip. We're talking about the general concept of taping out multiple dies for different segments. It makes no sense to do so for a low volume segment like GPGPU.

If the consumer GTX doesn't make money how would the much lower volume Tesla line do any better? You seem to be completely ignoring the amortization of R&D and other fixed costs.

 

I know what it is, which is why I referred to it....

I wasn't talking about Cg's GPGPU functionality, but the fact that it was never supported by ATI (as a shading language).

Also, I didn't say Larabee couldn't run CUDA, but that certainly isn't the developement environment Intel will be pushing.

Versus ignoring the fact that nobody will buy an inferior and by default irrelevant product at a much worse price/performance ratio.

While we are at it, anyone have actual numbers for the volume of Tesla products sold and projected growth of the GPGPU market?

 

Uh you've lost me now. How would taping out a separate chip help with that?

 

Trini, we tried...

 

LOL, ok... I'll check back in ~18 months.