Nvidia snaps out snappier Tesla GPU coprocessors

GPU chipmaker Nvidia knows that it has to do more to grow its Tesla biz than slap some passive heat sinks on a fanless GPU card and talk up its CUDA parallel-programming tools. It has to keep delivering price/performance improvements, as well. And that's exactly what it's doing with the new Tesla M2090 GPU coprocessor. Back …

[Heff]

...one other thing about the 590

as well as slurping power under full load, they run _hot_, they run _loud_ and they aren't rated for extended use under full load : ATi and nVidia have realised with desktop videocards that cards only really need to hit peak rated performance in bursts, and thats what they're built to do. Tesla units are built and designed around running under load for long periods of time and providing double-precision, something which you cant do with the 590, and nvidias mini-supercomputer department will doubtless pick up more work once crunch-happy modellers have melted a few stock 590s running hot 24/7.

[Destroy All Monsters]

WHERE exactly...

...do all the heatpipes on the M2090 go???

[Sorry that handle is already taken.]

Re: WHERE exactly...

By the looks of it, right in to the copper slug covering the GPU.

[lunatik96]

What is that dang thang?

This card is one major reason nVidia is in trouble. Look at that radiator on the board. Heat and process technology is holding the green goblin from advancing past SpidyAMD in the graphics market. Since nV can't compete on the CPU front, they have turned to Tesla products (all over-hyped and more power hungry). However, nV has done more to advance lower cost HPC performance. Lest the ghost of Phys-X and Havok appear in the middle of the night, homage paid, notorious mention. CUDA and OpenCL are the future. However, AMD is lurking in the corner with Fusion.

[David Hicks]

As someone that's been toying with OpenCL...

..I'm impressed at the performance figures.

Now if they would just put some effort into fixing the damned OpenCL compiler and runtime everyone could be happy. The total lack of feedback from the compiler on about half of all compilation errors is annoying. The fact that OpenCL kernels often compile and run and do nothing when they have blatant programming errors also doesn't help.

Still, forefront of technology and all that....