I admit to being a geek to some extent but I was surprised to find myself so happy after reading a set of slides yesterday. In my opinion we could not ask for a better endorsement for Niagara2.
Here is the Summary Page of Dave Patterson's presentation:
Berkeley View:
* Must help manycore revolution succeed
> 13 Computational Building Blocks vs. Old Benchmarks
> Autotuning to find best code vs. Traditional Compiler
* 1st example of autotuned building block:
> Sparse Matrix Sam Williams et al, “Optimization of Sparse Matrix-Vector
Multiplication on Emerging Multicore Platforms,” SC07, Nov. 2007.
* Niagara 2 Bad at 20th Century Metrics:
> Lowest Peak
> Performance (1/7) and Slowest Clock Rate (1/2)
* Niagara 2 Great at Berkeley View Metrics:
> Easiest to Parallel Program/Autotune,
> Best Power Efficiency (3X-4.5X MFLOPS/CPU Watt), &
> Highest Actual Performance (1.5-2X Faster)
For more information: View Wiki Page