As the release date of Tesla C870/D870 approaching (next month),
dose anyone have any idea on how much will they cost? we have
been waiting for too long.
Why do you want the C870? Is it just the memory? You do realize it’s just an extremely overpriced version of a regular graphics card, right? I think someone mentioned it will cost $2,000-$3,000.
The comment that a C870 is just an 8800GTX is not correct.
C870 certainly has more memory (1.5GB per card), plus a number of compute-specific validations and warranties. The pricing will be below the rumoured ranges too.
Paulius
We’ll see about the price. “Compute-specific validations and warranties” sounds like a load of bs. Maybe you mean that you’ve turned down mem frequencies to reduce bit errors. Probably just a feature for the paranoid, but I can do that too for free.
Hey Alex,
In terms of validations, these are actual compute/memory tests that pass on every Tesla product, not at all a clocking difference. In terms of warranties, these are Tesla specific and cover things such as longer replacement time-frames and availability. So for some folks, this won’t make any difference and that’s cool. Telsa is being using in some places where these benefits (I suppose “paranoia”) are demanded.
-dh
In a cluster environment, its not paranoia, its a simple requirement. Think cluster systems where the typical job is 128 nodes (the kind of system Tesla is targeting). Now, imagine that the average node crashed once every 30 days. So the average job will only last about 6 hours before crashing… which is worthless to those who typically run jobs for 96 hours.
I have by no means done the tests myself, but my hunch is that good geforces aren’t going to be much less reliable than the Teslas. There’s still no ecc, and the capacitors aren’t going to get much better. I’m pretty sure the “validation” just amounts to nvidia doing a final test and saying something like, “we’re sure it works.” I think those words are very soothing to the paranoid, but my question is “well… were we expecting it not to?” I think all the high-quality board manufactureres’ cards work very well, and obviously everybody performs validation (and if not “compute” validation, then I guess the plain one that simply runs computations).
As for extended availability… yeah, that’s something that is of special concern to enterprises (who often buy weird overpriced gear that the world forgets about in two years. caughtesla). But who here thinks there’s not going to be a ready supply of 8800’s five years from now?
But it’s ok, this is definately not as bas as trying to sell Quadros to engineers on the basis of “precision.”
I saw the popular NVIDIA GFLOPS graph comparing GPUs and CPUs recently updated to include the C870 and its dot was roughly 10% above the 8800GTX. 514 GFLOPS if I recall correctly. Is the chip the same as the 8800GTX just clocked faster or something else?