GFLOPS rating of a quad-core ?

[Markfw]

I have googled, and am still unable to find a benchmark program to measure gflops on a PC, specifically on my quad-core boxes. Anybody have a download link ?

 

[The-Noid]

The only ones I know of cost literally hundreds of thousands of dollars from IBM.

 

[RussianSensation]

Since Mflops: 10^6 flops and Gflops: 10^9 flops, can't you take Sisoftware sandra floating point performance (Arithmetic Whetstone I believe) and convert those numbers from Mflops to Gflops by dividing the number by 1,000?

This would imply Q6600 at stock speed is roughly 30-31 Gflops - LINK

 

[Markfw]

Then my Q6600 would have 41 gflops ??? @3.2

 

[Viper GTS]

Surely it can't be that simple. If it is somebody needs to tell this guy that you can buy a brand new, factory warrantied HP for under $700 that is faster than his "supercomputer."

http://browse.geekbench.ca/geekbench2/view/15359

Highest number I see is

Dot Product
multi-threaded vector 7305
7.60 Gflops

Viper GTS

 

[CTho9305]

GFLOPS is largely meaningless as a rating. It depends strongly on what you use to measure it - ignoring SSE, an Athlon or Opteron can do 2 FP ops each cycle, and Core does the same (couldn't find any Core 2 docs). That'd put your CPU at 3.2GHz * 2 FP ops/cycle * 4 cores = 25.6 GFLOPS.

Now, a real program generally won't keep the FPUs busy 100% of the time, so the FLOPS number you'll get could be as low as 0 (for a program that just doesn't perform any floating point calculations) and as high as the theoretical limit (for a program like burnK7).

If you really want a meaningless number, you can use LINPACK (with a FORTRAN compiler like icc), which is what top500.org uses to rank supercomputers. Unless everybody uses the same program to measure FLOPS, you can't compare the numbers at all.

edit: see my other reply below for numbers with SSE.

 

[Markfw]

CTho9305, Thanks for the great info. Since I found out they were 4 X2 3800's (I assume at stock) and I have one of those, I can say that for apps that I care about, my Q6600@3.2 has more power than 4 of the X2 3800's at stock, and for way less than $2500, but somebody else said that was an old article.

 

[PlasmaBomb]

It also matters if you are measuring single precision FLOPS or double precision FLOPS. The article dates from before intel released quad cores (mid 2006 when intel was just about releasing Core2Duo), and states that you could achieve the same performance now for less than 50% of the cost.

 

[Fox5]

Originally posted by: Viper GTS
Surely it can't be that simple. If it is somebody needs to tell this guy that you can buy a brand new, factory warrantied HP for under $700 that is faster than his "supercomputer."

http://browse.geekbench.ca/geekbench2/view/15359

Highest number I see is

Dot Product
multi-threaded vector 7305
7.60 Gflops

Viper GTS

1. That guy only has 4 dual core 3800+ cpus. I could see a single high speed quad core performing close to them.

2. He's measuring double precision (64 bit) while sandra probably measures single (32 bit).

Microwulf only costs $1256 now though, btw. When it was originally designed at $2500, it predated intel's quad cores.

 

[PCTC2]

how about networking 4 Q6600's into a cluster supercomputer? That'd be sub-$3k. anyone got the time and the cash? LOL

 

[StrangerGuy]

What about the 8800GTX? That should stomp any x86 CPU in sheer GFLOPness.

 

[Lord Banshee]

yeah but a 8800GTX can not run normal apps or apps that have already been design for R&D with CPU, it have to be program for a GPU, and it can only do single precision floating point so it is almost worthless lol. The 9800 series is suppose to be double precision, this will be interesting as supercompeting goes.

 

[CTho9305]

I looked into this a bit more, and I think Athlon64 / Opteron are going to be 4*CPU speed (so, 4 GFLOPS per GHz per core), since you can do single-precision (32-bit float) SSE packed adds (2/cycle) and muls (2/cycle) together.