importance of CPU cache size

[James Heinrich]

I'm sure there's a thread about this somewhere, but I couldn't find a nice concise source of information: What is the exact effect on (Prime95) performance for various L1 cache sizes?

For example, take 4 somewhat similar CPUs, same clock speed:
* AM2 Sempron 3200+ 1800MHz, 128kB
* AM2 Sempron 3400+ 1800MHz, 256kB
* AM2 Athlon64 3000+ 1800MHz, 512kB
* 939 Opteron 165 1800MHz, 1024kB
(on the Intel side you'd have similar examples, from 256kB to 2MB)

What is the performance hit/gain at each level of cache, for both trial factoring and LL? Obviously more is better, but by how much and is there a threshold of great benefit?

[Citrix]

, How does L2 size affect prime95?

[James Heinrich]

Oops, yes, I obviously meant L2 cache in my above question, although if someone wants to comment on the L1 cache sizes as well, please do!

[dsouza123]

For AMD CPUs the L1 and L2 are exclusive (non duplicating)
so the sizes are added together. The L2 is larger but slower.

The L1 is split between code and data.

You would have to test each of the four CPUs
to find it out, unless someone has already done so,
there is a benchmark thread and also a page
with timing results.

http://www.mersenneforum.org/showthread.php?t=59

The type and speed of main memory will also have some effect on timings.

The data for TF should fit in the cache maybe the calculating code too.
The data for FFT doesn't fit, don't know about the code.

[Prime95]

Check out http://mersenne.org/bench.htm

I know the Intel chips suffer greatly at 128K L2 cache. I think the 512K and up chips all perform about the same.

[James Heinrich]

I made http://mersenne-aries.sili.net/bench.php yesterday and compiled all the results from the perpetual benchmark thread, but unfortunately there's a shortage of benchmarks with TF results (only the most recent ones), so I don't really have enough data to compare.

I'll be trading my 1024kB Opteron in for a 128kB Sempron on Sept 1 (yes, I know, it's sad), so I'll be able to do some direct comparisons for that at least, so within 2 weeks we'll know what bestcase-worstcase looks like (for AMD at least), although filling in the gaps with 256kB and 512kB would be nice too.

[Prime95]

TF is likely immune to L2 cache size changes. The small sieve it uses was designed to fit in an 4KB or 8KB (I don't remember which) L1 cache.

[James Heinrich]

Originally Posted by Prime95

TF is likely immune to L2 cache size changes. The small sieve it uses was designed to fit in an 4KB or 8KB (I don't remember which) L1 cache.

Confirmed -- TF is immune to cache size.

Comparing my Sempron2800+ (128kB) with my (former) Opteron165 (1024kB) at 1800MHz (stock for the Opteron, overclocked by 12.5% for the Sempron), TF times are essentially identical for the both of them (slight edge to the Sempron, but within 1%). FFT times are reasonably close, but it gets significantly worse at larger sizes:

Prime95 64-bit version 24.14

         1024kB    128kB
       +----------------
58-bit |   4.24     4.22
59-bit |   4.26     4.22
60-bit |   4.52     4.49
61-bit |   4.67     4.63
62-bit |   5.53     5.49
63-bit |   5.93     5.92
64-bit |   7.71     7.62
65-bit |   9.13     9.05
66-bit |   9.09     8.97
67-bit |   9.10     8.95
       | 
 512K  |  26.64    26.22    - 2%
 640K  |  34.57    37.04    + 7% 
 768K  |  42.05    44.85    + 7% 
 896K  |  50.34    53.79    + 7% 
1024K  |  55.63    59.90    + 8% 
1280K  |  70.69    76.09    + 8% 
1536K  |  87.25    93.09    + 7% 
1792K  | 105.13   112.43    + 7% 
2048K  | 117.76   125.59    + 7% 
2560K  | 154.63   168.12    + 9% 
3072K  | 189.65   212.47    +12%
3584K  | 228.13   265.69    +16%
4096K  | 254.53   313.01    +23%

See full benchmark details on my benchmark page