SPARC64 IXfx coming soon

[jhu]

to a super computer far away from you.

I used to have a Sun IPC. I miss those days.

 

[Joseph F]

What are the performance figures compared to the 10-core Xeons? (Aren't there a few inter-uarch benchmarks?)

 

[Soulkeeper]

"performance levels of 236.5 gigaflops and performance per watt of over 2 gigaflops."
not bad, what's the i7 do ? around 70 gigaflops ?
"85GB/s" must be quad channel ddr3

 

[IntelUser2000]

"performance levels of 236.5 gigaflops and performance per watt of over 2 gigaflops."
not bad, what's the i7 do ? around 70 gigaflops ?
"85GB/s" must be quad channel ddr3

1.848GHz, 16 cores, 236.5GFlops. Let's do the math.

236.5/16/1.848=Looks like it can achieve 8 DP flops per cycle.

Comparison:
-10 core Westmere EX: 2.4GHz x 4 DP Flops/cycle x 10 cores = 96GFlops
-4 core 2600K: 3.4GHz x 8 DP Flops/cycle x 4 cores = 108.8GFlops
-8 core Romley Xeon: 3GHz x 8 DP Flops/cycle x 8 cores = 192GFlops
-16 core Interlagos: 2.6GHz x 8 DP Flops/module/cycle x 8 modules = 166.4GFlops
-8 core Power 7: 4.14GHz x 8 DP Flops/cycle x 8 cores = 264.96GFlops

 

[Joseph F]

Even though they may be able to do a similar number of "operations" per second, aren't the operations themselves so different that different architectures can't be compared like that?

I was thinking along the lines of linpack or perhaps, some kind of a cross-architecture video encoder. (not that I know of one)

 

[Cerb]

They can be compared, sure. They all have scalar and vector FPU, and can issue at least one double operation every cycle. FP benchmarks would likely need to be tuned for each CPU, though. Fujitsu's paper numbers are as good as anyone else's paper numbers, and are likely 100% accurate for a small loop fitting in cache. I'm sure each computer could have SPECfp run on it, giving a moderately fair comparison for application performance.

But...

They are mainly selling their management software, compilation software, proprietary network filesystem, interconnects, and a history of solid RAS from Fujitsu SPARC computers. The FP performance of each CPU will be eclipsed by commodity devices in short order. Total system performance and reliability...that's more fuzzy.

 

[IntelUser2000]

I was thinking along the lines of linpack or perhaps, some kind of a cross-architecture video encoder. (not that I know of one)

Yea, Linpack can get pretty close(90%) to theoretical. It fully utilizes the FPU meaning its not really bottlenecked by anything else.

I think the chips I mentioned are roughly comparable for FLOPs counting. It's not like Niagara, Cell, or any other chips that blur the line between general purpose CPU and specialty chips.

Still if you want to put millions of those cores together to get into top500 rankings, there's no substitute for peak FP power.

 

[lol123]

I think the benchmark you're looking for is SPECfp.