Page 1 of 1

nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Fri Jan 17, 2014 12:56 pm
by SecretSquirrel
Does anyone have physical access to an nVidia Tesla K20 card that could take some measurements. Apparently an engineering drawing of the card is a rare beast indeed. While I am pursuing formal docs through channels, if anyone has a card and could take measurements for me, I'd greatly appreciate it. I'm looking for the length of the PCB, end to end and the total length of the card from the outside of the bracket to the end of the PCB.

Thanks.
--SS

Re: nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Fri Jan 17, 2014 1:43 pm
by StuG

Re: nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Fri Jan 17, 2014 5:58 pm
by SecretSquirrel
StuG wrote:
http://www.nvidia.com/content/PDF/kepler/Tesla-K20-Active-BD-06499-001-v02.pdf


Yep, we know about that. When you are doing mechanical design work and such, that is an absolutely useless drawing. Something like this is a little more along the lines of what were are after. Though even it is missing a few dimensions.

--SS

Re: nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Fri Jan 17, 2014 9:13 pm
by kumori
I don't have access to a card, but I'm interested in what your designing!

Re: nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Sat Jan 18, 2014 7:04 am
by vargis14
For Basically a professional/server version of the TITAN 5GB of memory seems strange. Since Titan has 6GB of memory and the 780TI has 3GB I though it would be increased in 1.5GB size intervals.
Maybe since enabling ECC will cause some of the memory to be used for the ECC bits, so the user available memory will decrease by 10%, they threw a little extra on.

I could be way off since i just opened my eyes and had only a sip of coffee so feel free to correct me.

Re: nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Sat Jan 18, 2014 12:20 pm
by Bauxite
vargis14 wrote:
For Basically a professional/server version of the TITAN 5GB of memory seems strange. Since Titan has 6GB of memory and the 780TI has 3GB I though it would be increased in 1.5GB size intervals.
Maybe since enabling ECC will cause some of the memory to be used for the ECC bits, so the user available memory will decrease by 10%, they threw a little extra on.

I could be way off since i just opened my eyes and had only a sip of coffee so feel free to correct me.


Desktop ecc adds another chip as that's how their memory bus has been designed. Basically everyone not using ecc has extra dimm traces and logic sitting unused because its been decided that its "ok" to use data integrity as a market segmentation feature as dram density has increased 1000 fold in a decade or so. The whole situation is **** dumb, plain jane desktops still manage tons of data all over the world and they should have gone 9 dram chips standard ages ago. A lot of data buses already have ecc or similar functions running on them because errors do happen period, its nature.

Except for crap like older atoms and more recently some amd apus, every single amd and intel wafer for many years supports ecc when it comes out of the fab. It might be disabled later depending how it is binned. For an example of this stupidity, desktop i3s support ecc as they are sold along with E3 xeons by many server vendors as the bottom tier dual core option, however i5/7 do not. They also make very sure no unlocked cpu supports ecc, though ironically if you don't care about overclocking many E3 xeons can be cheaper than identical speed i5/7s.

A lot of GPUs instead reallocate through their memory controller to accomplish the same algorithm, I believe that titan has 6GB physical.

Tune in next week for my rant on why all hard drives are trash and your file system is a joke even if you don't know it: just remember 2 is one, 1 is none.

Re: nVidia Tesla K20 dimensions -- gerbels, help me out

Posted: Sat Jan 18, 2014 12:38 pm
by SecretSquirrel
kumori wrote:
I don't have access to a card, but I'm interested in what your designing!


We actually produce a FPGA based accelerator card that is very close to the same size as a K20. We were having a discussion about whether we would fit in certain semi-cusom chasis, like a Penguin 2808GT. Since it is pretty expensive to go out and buy a K20, or a server, just to determine whether we fit, we started looking for dimensional drawings. Since we had other chasis to, the easiest thing (or so we thought) would be to get a dimensional drawing of the K20. Best we have been able to find is the above linked PDF from nVidia which is useless for what we need. When you are talking 1 or 2 milimeters being the difference between fitting and not, a generic 10.5 inches without any other info just isn't enough.

--SS