A PC for make synthesis

I would suggest a dual processor with lots of dual port ram and 64 bit at least. That way when the processor is busy place and routeing, you can be doing the documentation in Word.

So far I think the jury is out as to 32 / 64 bit code .. but I think 32 bit is slightly faster until processors speed up

Simon

It seems you speak about non-x86 MPUs. For AMD64/EM64T MPUs this is not the case. 32 bit code runs in 64 bit mode (OS) as fast as in 32 bit mode and recompiled code with a lot of calculations runs even faster in most cases with a good compiler (thanks to more registers etc.).

Matthias

But you mentioned the two magic words ... Good and Compiler. :-) But I don't know for a fact if its faster or slower.. just initial comments ... the main difference seems to be clock speed.. and the 32 bit CPU's are currently faster ...

at

be

bit

An A64 with a 1M cache. The cache size is more important then clock speed and memory bandwidth. I did some benchmarking of common FPGA development task like simulation and Xilinx builds,

Calculation for synthesis process is more float type or integer type?

The old AMD processors have less performance in float than Intel.

Also, if cache is so important, newer Xeon have lot of cache.

What do you think?

during

Ok, I will buy an Athlon 64.

What model has the best price/performace ratio now?

Thanks Marco

comments

CPU's are

Well, there are already many good compilers for AMD64 compatible architectures like AMDs K8 and Intels Prescott/Nocona:

- PGI Compiler Suite

- Pathscale EKO Suite

- GCC 3.x and 4.x (yes, they are even used for SPEC submissions)

- Intel ICC 8.1

- MS VCC and some more or less exotic compilers (but they are not the bright lights performance-wise). Some of them include fortran compilers.

Look at the SPEC submissions:

In this quarter there are several submitted 64bit results using some of these compilers (look for Opteron SPECint results from HP and Sun or SPECfp from FSC, HP and Sun). Especially the SPECfp results are quite a bit higher than AMDs single submission (using ICC on Windows).

The AMD K8 CPUs (Opteron, Athlon 64 etc.) are also 32 bit CPUs, which have a 64 bit mode. That's the difference to other 64 bit CPUs (MIPS, Alpha, HP PA, Itanium), which are designed as 64 bit CPUs from ground up with some backward compatibility (if they had 32 bit predecessors).

But why does an Opteron with slow registered RAM and only 2.6 GHz and 1 MB L2 cache achieve higher SPECfp results than the fastest Xeon (3.6 GHz) or Pentium 4 (3.8 GHz) and also higher SPECint results (using the ICC compiler)? Because they are different architectures. The K8 has a shorter pipeline (code with many branches doesn't hurt that much) and higher execution bandwidth per clock and the Prescott/Nocona has a longer pipeline with less execution bandwidth per clock, but higher clockspeeds.

So clock speed can't be used to compare them. It's like (since we are in a FPGA group) having 2 FPGAs able to run at 200 MHz but one of them has twice as many gates.

I hope, writing all this gave at least a nice and comfortable update about current technology at least on the x86 MPU market :)

Matthias

But the old AMD processors aren't sold anymore. Since 1999 the Athlon had higher FP performance than Pentium III and although the Netburst architecture managed to get the crown back over time in a lot of applications, the situation changed again. Now the Opteron/Athlon 64 achieves higher performance for most scientific calculations, 3D calculations for games and similar stuff. For video encoders, some audio encoders and some renderers the situation is favouring Netburst architectures in case of single core CPUs.

Regarding the help of big caches the same holds true, which I wrote to Simon Peacock: But why does an Opteron with slow registered RAM and only 2.6 GHz and 1 MB L2 cache achieve higher SPECfp results than the fastest Xeon (3.6 GHz) or Pentium 4 (3.8 GHz) and also higher SPECint results (using the ICC compiler)? Because they are different architectures. The K8 has a shorter pipeline (code with many branches doesn't hurt that much) and higher execution bandwidth per clock and the Prescott/Nocona has a longer pipeline with less execution bandwidth per clock, but higher clockspeeds.

Here is a nice article comparing the Pentium 4 with 1 MB L2 to the new one with 2 MB L2:

Matthias

The best price/performance ratio is the 3400+ or the 3700+ which are 754 pin processors. The downside of those processors is that you can only put

3 DIMMs on them so get 1G DIMMs (you want 2G minimum on a machine for FPGA development). AMD underprices the 754s compared to the 939s because they only have one memory bus. But as you can see from benchmarks the 3400+ runs rings around the 3800+ (939 with 1/2M cache) on NCVerilog. The only places that my 3800+ does better than my 3400+ is where there is a lot of I/O. This is a disk issue not a processor issue. My 3400+ is a laptop and has a slow 4400RPM drive. My 3800+ is a desktop that has a fast 7200RPM SATA drive. A 3400+ with a 7200RPM drive will scream. The 939s are available with a 1M cache also but they are very expensive. However come June there will be dual core 939s available. Dual cores are very useful for a workstation. The Xilinx tools can run multithreaded on Linux machines and even if you aren't running them that way you will likely want to be able to simulate and place and route at the same time.