Tesla K10

Can’t wait to get reliable pricing information on this card. Not the normal direction for the Tesla product line but if the price is sensible then it could be a decent amount of SP performance and video memory for the money. In the UK a pair of GTX 680 4GB cards will cost you about £1000 so I’m guessing that a K10 will be what £2000? £2500?

Usually Nvidia priced their Tesla about 5x their GeForce equivalent. :angry: The GeForce equivalent of K10 is GTX690. So it is likely it might sell for $5,000.

However, at this price, I suppose the sales will be very poor. If their sales department is smart, they should price them no more than $2,500 :no:

Would I be right in assuming that the ECC on the memory would be configurable like on Fermi and that if enabled it would reduce the memory capacity and bandwidth relative to the quoted figures?

The GK110 whitepaper says that for the K20 they have made some changes to reduce the performance penalty for ECC, so I assume that K10 will also have some penalty. The question to ask NVIDIA is whether that performance loss is comparable to Fermi, or if GK104 will also have some ECC improvements like GK110.

Curious - how do you reduce the performance loss besides just having fewer bits of parity?

I got a quote for the Tesla K10 too. Around £2100 (ex VAT) versus £750 for a GTX 690 (which only has half the memory). Doesn’t sound too bad put like that does it?

I have been watching the talks in GTC and have been very impressed with the evolution.
However, there is one thing i am not clear.

The K10 is based on the GK104 chip.
The K20 is based on the GK110 chip.

The GK 110 supports the Hyper-Q and the dynamic parallelism features.
Does the k10 support these features too?

Cheers,
Than

Hyper-Q and dynamic parallelism are GK110 features not found in GK104. Please see the following whitepaper for the new features of GK110:

http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

For those who don’t feel like reading the full whitepaper, there is a quick summary here:

http://www.nvidia.com/content/PDF/kepler/NV_DS_Tesla_KCompute_Arch_May_2012_LR.pdf

Given that GTX 690 's DP performance is 1/24 of its SP performance, its SP performance is 5600GFLOPS and DP is 234GFLOPS. Both numbers beat the 3x priced K10. If I don’t need ECC and I want to play games, maybe I should get 690 for $999??? :no:

Yes. This strategy has worked well for me for 5 years now. You just have to be willing to deal with occasional card death and keep an eye out for overheating and memory errors (run your calculations twice, for example). When you can take responsibility for the hardware and the software (and don’t need the Telsa-specific features), GeForce can be very productive. In a computing service environment, like a large multi-user cluster, the uncertainty of consumer hardware might not be acceptable and you need hardware that makes stronger reliability guarantees. But if it is just you and your desktop, a little bit of care will save you a lot of money.

Did you buy K10 eventually? Is it just an unclocked 680 with ECC, more RAM and being more “stable”? Does it give us surprises??

A very bad thing is NVIDIA may have a contract with some big company,such as dell. They refused to sale their 16PCIe box with GTX card. :thumbsdown:

I would imagine it is Dell being risk averse (much like their enterprise customers). They have to support the configurations they sell, and most of their customers are not going to screw around with GTX cards in high-end computer systems, so they aren’t going to test such configurations.

Smaller computer assemblers are usually more flexible and would probably be more than happy to use GTX cards. (If you are trying to build a computer with eight GTX 690 cards, then you really should be comfortable building that from scratch and dealing with the problems involved. :) )