Opteron 244 (1.8 GHz) benched with Prime95 22.12

[nomadicus]

Interesting to see this benchmark article. The thing that caught my eye was the 4 processor Opteron 9GB/s memory throughput.
Anyone care to speculate what this could mean to a future prime95?
-- edit -- and the hyper transport busses?

[Dresdenboy]

Originally Posted by nomadicus

Interesting to see this benchmark article. The thing that caught my eye was the 4 processor Opteron 9GB/s memory throughput.
Anyone care to speculate what this could mean to a future prime95?
-- edit -- and the hyper transport busses?

It's a nice number but it's not of use for Prime95 if the calculations won't go faster inside the CPU. But if there are some free resources (say many machine cycles where the FPU is waiting for work) one could optimize the client to spread the data in a way that it could put a part into local mem and others into mem of other CPUs. But that's far from being optimal for this hardware configuration since every CPU could work on it's own set. If other CPUs want data from one CPU then local bandwidth would decrease significantly. Better let them each have their own stuff to calculate :)

One interesting thing for Prime95 could be a workstation mainboard with an own northbridge which is connected to the CPU via HT. This could increase max. bandwith of single-channel Athlon 64s or even increase bandwidth of a single Opteron with dual-channel RAM. But do we already need already more than 6.4 GB/s for Prime95 on a single CPU? Currently it's not much faster than Athlons which have max. half the bus bandwidth (and much higher latency). I'm still investigating if the Opteron is already calculating as fast as it could.

Regards,
DDB

[ewmayer]

Originally Posted by Prime95

Benchmarks from a 1.8GHz Opteron:

AMD Opteron(tm)
CPU speed: 1764.54 MHz
CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2
L1 cache size: 64 KB
L2 cache size: 1024 KB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
L1 TLBS: 32
L2 TLBS: 512
Prime95 version 22.12, RdtscTiming=1
Best time for 384K FFT length: 26.539 ms.
Best time for 448K FFT length: 31.959 ms.
Best time for 512K FFT length: 35.640 ms.
Best time for 640K FFT length: 47.547 ms.
Best time for 768K FFT length: 58.636 ms.
Best time for 896K FFT length: 71.287 ms.
Best time for 1024K FFT length: 79.153 ms.
Best time for 1280K FFT length: 100.040 ms.
Best time for 1536K FFT length: 119.910 ms.
Best time for 1792K FFT length: 147.761 ms.

Tom Cage just sent me the first results for a gcc 3.4 64-bit build of Mlucas on an AMD64. On his 2.2GHz machine Mlucas (with zero assembly code aside from the odd prefetch macro) gets about the same performance as Prime95 does on the above 1.8 GHz CPU. That is surprisingly close.

http://hogranch.com/mayer/README.html

[optim]

Does MLucas run faster than Mprime/Prime95 on AMD AthlonXP? If yes, can I use MLucas to win the $100,000 prize though GIMPS, just as if I was using Mprime/Prime95?

BTW: My AthlonXP 2GHz has about the same performance as this Opteron 1.8GHz, meaning that A64 integrated memory controller and other benefits make up for 200MHz of frenquency.

[ewmayer]

Originally Posted by optim

Does MLucas run faster than Mprime/Prime95 on AMD AthlonXP?

No. Except for a few prefetch-related assembly-code macros, Mlucas is generic C, and lacks any specific optimizations for x86-style CPUs, whether they be Intel or AMD. That's why I expressed surprise that it was even close to Prime95-style performance on any such architecture.

[E_tron]

I'm surprised that the celeron D is doing well with Prime95. The extra 128k of L2 helps a lot.

http://www.mersenneforum.org/showthread.php?t=2862