Yeah. To contextualize that a little...

The i7-740QM was introduced almost 3 years ago, June 2010, at release time costing $378 (if you bought in lots of 1000).

I didn't find historical pricing data for the 445M video card, but here's a relevant data point: I bought my GeForce 470 at almost exactly that time, for $420 retail (so the bulk price would be around $400, I think?). Measured in raw FLOPS, the GF470 is over three times as fast as the 445M (the video card tested).

So dollar for dollar, TFA undervalues the GPU by about a factor of three?

Sources: http://en.wikipedia.org/wiki/Comparison_of_Intel_processors http://en.wikipedia.org/wiki/Nehalem_%28microarchitecture%29 http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_p...

I'm working on running the tests on 2x Xeon E5-2630 and a Tesla K20c. Will post the results here in a couple hours.

You think the average programmer (user) has a Tesla computing device on his computer ?

The 445M is a mobile GPU and quite low end at that. In fact the Teslas are consistently slower than the GTX cards for single precision (until their k20 line that is). So something like a 480 would be ok. 580 would be ideal.

You can buy an hour with two Tesla GPUs for $2.10 on EC2.

Are you allowed to install software on an On-Demand Cluster GPU Instances or you must use what is provided by Amazon ?

cuda comes pre installed with gpu instances. but you can install sotware as well.

A 445M is an sm_21 Fermi device with 3 SMs (144 cores). This is a low-end GPU.