That website is confused, Gigabyte GT220 OC is not GDDR3. It is one G of DDR3.Not that I blame them: Two chips came to market … in three different memory configurations … each available in two sizes … from a plethora of vendors … :-D
I did get my hands on a Palit GT220 Sonic Edition 512MB card with GDDR3 memory. It seems that most GDDR3 models are limited to 512MB RAM due to the higher component price.
Device to Device Bandwidth
.
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 24997.1
Yummy.
In my app benchmark doing parallel 4x4 matrix inversions this achieves 74 GFlops/s whereas a GT220 with DDR2 memory peaked out at 59 GFlops/s.
So if your application is memory bound, you’re definitely going to see some speedup from the GDDR3.
Now the unfortunate thing is that this particular Palit model does not run at all on my OpenSuse 11.1 - the driver just hung. Fortunately I was able to run the primary display on an older 8500GT instead - and then CUDA works on the 220GT as well (even without kernel time limits).
That really does look like a nice development part - competent enough to do real work and it should be pretty straightforward to extrapolate single precision results straight to a GT200 based production system.
As an aside, my local vendor of choice might have let the cat of the bag on the GT240 (which is supposed to be a 96 core compute 1.2 part with either GDDR3 or GDDR5 ram on a 128 pin bus). Their product list already includes an ASUS GT240 card half showing (no specs or prices), so it might well be released very soon.
Try another driver? Although not mentioned anywhere, even drivers as old as 190.32 supports the card. Which one do you have installed now?
Indeed! It has the same bandwidth to main memory as my E1200 has to level 2 cache. I can tap it for 50Gflops of realtime audio processing with very low latencies, kernels running once every 0.3 ms. There is just no Core2 part in existence that could do that, especially not below €80 (including an extra G of DDR3 ram)
All is not rosy though. It doesn’t do video (as in television) very well. Actually it is ghastly, which is a bit surprising for a card shipped with an HDMI output.
Tried the 190.42 as well as the 195 beta driver (CUDA 3.0).
Turns out in was the vga=0x31a setting passed to the linux kernel by GRUB that changes the text console to a high resolution mode. This triggered the crash on starting X.
Probably some weird interaction with the graphics card’s BIOS.
Christian
Ah, OK!
Tips: I use vga=0 to get at the most low-level 4 bit 640x480 mode so that I can run and watch debug messages, set variables while also editing the source without tilting the running RT-application. This on the console(s) with RT-kernels having turnaround times below 0.3 ms. The SVGA modes with higher resolution and more colors would in this scenario break as soon as you just scroll off the edge.
[As I understand, you have a second GPU for display, so may not personally have any use of such info.]
Tips #2: ‘setfont lat1-12’ gives you 30% more room on the somewhat cramped console
Tips #3: ‘modprobe vga16fb’ (as root) makes the setting above stick even if you take a detour around X … The switch to X still kills any RT-application with that kind of time constraints though.
UPDATE: The above problem was later solved by Nvidia with Linux driver version 195.22 including switch back and forth to X (within “reasonable” limits.)
You can forget about vga16fb for now.
Is anyone willing to “take a bullet for the team”, buy a GT240 card and post the deviceQuery? I think the single-slot Zotac GT 240 cards are particularly interesting as upfit GPUs for some 1U compute nodes:
Not requiring supplemental power is also a plus. The only downside with the current Zotac models is having to choose between 1GB GDDR3 or 512MB GDDR5.
The bandwidth gap between GDDR3 and GDRR5 is pretty significant. I would go for a 1GB GDDR5 model from another vendor.
Who has 1GB GDDR5 single-slot? The only real reason I see to get these for CUDA is to fit them in places a GT200-series GPU won’t go. If you have the room, go with a GTX 260 for ~$40 more and get compute level 1.3. We have a number of (well-funded) department heads who would love to be able to upgrade their existing cluster compute nodes for CUDA.
Palit are the only company with a release day 1Gb GDDR5 part, and it is a dual slot card…