ION gpu spec wrt. CUDA

hello,

are there any more detailed informations about the ion<->cuda relationship?
compute compatibility, multicores etc

thanks in advance

moik

Ion is basically our brand name for the GeForce 9400 mGPU combined with Atom CPUs. The specs for the 9400 are in the programming guide - 2 multiprocessors (16 processors), compute capability 1.1.

We have one of the Acer Revo systems here, and it’s a nice little box!

grr we need 1.3 :whistling:

thanks for your fast answer.

we are looking for an embedded cuda solution for image preprocessing

and our software definitly needs 1.3, because atomics on shared mem are so useful.

is there anything in scope?

greetings,

moik

I believe you only need 1.2 then? (From memory 1.3 only has double-precision added).

There are some note-book GPUs out now that support 1.2 I believe.

Moik,

What size embedded are you referring to? Power/weight restrictions? I’m finishing up an embedded CUDA-capable system right now for a project I’m working on that might be of interest.

Simon, do you know if the Pico-ITX ION is coming out? It seems a 9400GM is pretty skimpy, but hopefully still capable. And if it’s anything like the price of the Mini-ITX ION than it’s well worth losing more cores. Hell, you could even buy 2 and multi-task. 17cm x 17cm Mini-ITX is much better than a desktop strapped to the back of a 20 pound robot but the PICO could actually fit in my robot.

-Mack

well we don’t have any concrete size or power specs in mind, at this stage just small and low powered. ;-)
But a tesla c1060 is definitly oversized.
our project is going into the “smart cam” direction.
and we dont want to go the fpga way.
it would be great if nvidia will offer solutions there.
the tegra plattform is promising, but i dont know about any cuda capabilities.

for smart cams we dont need the full massive horsepower of a telsa (clockrates, memory),
parallism is much more important even if fpgas won’t be beatable there (parallism/power consumption).

cudas big advantage are fast implementation cycles compared to fpgas.

@riedijk: sure? compute compatibility 1.2 also allows atomics on shared mem?
(update: just too lazy to rtfm: programming guide 2.2 s.112: … compute compatibility 1.2)

thank you for your answers,

moik

thats the point!

in my dreams there would be a box like the picoITX box you posted with a firewire 1394 bus and DMA zero copying for direct transfer to GPU memory and cc1.3 .

well ok, and some other things B)

I won’t take credit for the idea; Joe Stam who co-wrote the CUDA stereo example suggested I look at a company called Liantec. So, at this point the best thing available is a Mini-ITX from Liantec that carries something they created called the Tiny-Bus which lets you seat a 8600M GT. I have the ITX-6965 which is 17cm x 17cm (standard Mini-ITX) and a 2.2 GHz core 2, they now have the Intel Penryn chips which are better on power. Mine is about 60W CPU+GPU vs the newer set with 2.2GHz drops to around 50W total heat dissipated by the heat sinks. The 8600M is the only thing that will fit, unfortunately. You can buy the cards from mxm-upgrade.com. The guy who runs the store said he was having bios issues with the 9600M a few months ago so I don’t know if he’s figured it out. Even then, the 9600M doesn’t seem to be that much of an improvement so you’re not hurting too much there by not getting the 9-series card.

I’m hoping that the Pico-ITX ION might actually make its way to the commercial stage but it looked from blogs that it was just a demo version to show what they could do for netbooks. The Pico has a 9400M GS which only has 2 multiprocesors where as the 8600M GT has 4.

I’m using UNC’s GPU-SIFT running Cg at around 7 1/2 hz on 640x480. However, there library is definitely meant for processing folders with saved images, not live feeds from a camera so I have to do a few extra memory copies. I’m on a Mac so I haven’t gotten their CUDA version to work yet. Once I finish my thesis I’m planning on starting from scratch with their CUDA code as an example, or maybe SURF rather than SIFT because it’s faster along with some firewire code I’ve been writing on top of libdc1394.

I can say your best bet is Liantec, or praying for the Pico-ITX ION

I’ll add that to get firewire I’m using a FirePRO low-profile card from Point Grey and a PCIe extension cable to allow the 1394 card do mount flat. I think the total height of the box we’re building is something like 7-8 cm, 10cm worst case. I have a mech-E undergrad doing the design for me so I’m not sure off the top of my head.

The Pico-ITX board shown on that page is just our reference design, we don’t sell this. I’m not aware of any of our board partners shipping a Pico-ITX design so far, but I’ll keep you posted.

Your robot sounds cool, you should post some pictures!

You SHOULD sell it. I’ve been amped for a commercial version of your reference design since you announced ION. So far, nothing but Mini-ITX. It has been disappointing to say the least. The Pico-ITX was damn :heart: sexy :heart: and I know a lot of other HTPC’ers who feel the same way about it. PLEASE WAKE YOUR PARTNERS UP, OR SELL US SOME OF THE REFERENCE DESIGN!

You can buy an Acer Revo which is based on this platform. Its size makes me think it’s using pico-ITX as well.

Not even close. It uses a custom motherboard that looks to be a little larger than mini-ITX. The register has some images of it

here

Quote from that link: The Revo is built around a Pico-ITX motherboard and has a tiny lozenge-shaped case that measures 195 x 210 x 37mm

Does that look like a Pico-Itx board to you? It has got SO-DIMMs socketed on top side of the main board, and that complete fills the inside of the case, which itself is a fraction larger than the 170mmx170mm mini-ITX board size

is a video ok, too?

http://www.youtube.com/watch?v=JkjmwOZN7bg

this car is equipped wtih massive computer power and a GPU.

it is research and this robot is an experimental car.

but we are trying to narrow the gap between our experimental robot and consumer cars.

This directly leads us to “smart cams”.

In my opinion, it would be a great deal if nvidia provides its excellent GPU interface CUDA

on more downsized, embedded devices (without discussing how “embedded” is exaclty defined here ).

There is a trend towards mobile devices, as still mentioned TEGRA seems promising. But i couldn’t find

any information on supporting CUDA.

@simon: is there any official statement of nvidia which covers this topic? It would be interesting for me, especially when discussing with people from automotive industry.

bests,

moik