NVidia has published the specifications for the new Tesla M2050, M2070, and M2070Q cards online.  They looks like they’ll do a pretty good job at setting a new bar for GPU computing benchmarks, both being powered by Fermi and boasting many of the same features you get in the newly released Quadro 5000 and Quadro 6000 packages.

One interesting little tidbit about the M2070 and M2070Q.  You look above and may wonder, “Why is there a M2070 and M2070Q?”  The Q is actually the answer to an interesting question a few readers brought up after my Quadro 5000 review. If you’ll remember, there was a noticable (very noticable, some would say ridiculous) difference in the performance between the GeForce cards and even the 3-generations back Quadro.  The Quadro cards excel’ed at the type of OpenGL operations typically undertaken in visualization and CAD packages, which is exactly what they are supposed to do.  However, several large research labs have begun deploying visualization clusters with Tesla’s on board, because they don’t need the video output anyway.  The Tesla offers workstation-level quality, but a slight drop in power consumption and eliminating the unused ports that makes large deployments attractive.  However, how does the Tesla perform in graphics?  It is a Quadro or a GeForce?

I’ve asked NVidia, but haven’t gotten a satisfactory answer.  First it was “Like a GeForce, since they aren’t tuned for Graphics”.  Then it was “Like a Quadro, because they are both Workstation grade”.  Then it was half-and-half, depending on the generation.  The M2070Q solves this definitely, as the Q is for “Quadro”, meaning it contains the same graphics optimizations used in the Quadro.

Now, typical usage in such an environment may make the whole thing unnecessary.  Unlike a real video card that will be connected to a monitor, Tesla’s in such clusters are almost exclusively used in a “readback” buffer mode, where frames are read back into main memory and then transferred to another system (a remote display over the network, a texture buffer on another card, perhaps several of them are combined into a single Ice-T style buffer).  Do the Quadro enhancements help there?  I honestly don’t know.  I suspect they help a little, but the overwhelming majority of your performance hit goes into the readback, which won’t be any different.

Tags