Time flies. RAMspeed, a free open source command line utility to measure cache and memory performance of computer systems, changes as well. It has evolved successfully from v1.00 released in November of 2002 as a result of my personal amusement containing about 100 lines of C code to produce one simple benchmark, to the latest versions written in assembly language mostly. There are 3 hardware platforms supported (i386, amd64, alpha) and several most popular UNIX-like operating systems. A quite popular DOS (and Windows) version exists as well. Nowadays, the software offers 18 cache and memory benchmarks for i386 and amd64 machines, though 6 only for alpha ones.
 
So far, RAMspeed has been tested to compile and run with assembly level optimisations on:

• Linux (i386, amd64, alpha)
• FreeBSD (i386, amd64, alpha)
• NetBSD (i386, amd64, alpha)
• Digital UNIX (alpha)

No need to explain here in depth all the benchmarking algorithms implemented in RAMspeed, better look at the documentation supplied and the source code. In general, there are *mark benchmarks such as INTmark, FLOATmark, MMXmark and SSEmark. They operate with linear (sequential) data streams passed through ALU, FPU, MMX and SSE units respectively. They allocate certain memory space and start either writing to or reading from it using continuous blocks sized in power of 2 from 1Kb up to the array boundary. This simple algorithm allows to show how fast are both cache and memory subsystems. There are also *mem benchmarks such as INTmem, FLOATmem, MMXmem and SSEmem. These are supposed to illustrate how fast is actual read\write memory performance. Each of them includes four subtests called Copy, Scale, Add and Triad. They're synthetic simulations, but correlate with many real world applications. You may have seen them already within STREAM and SiSoft Sandra. All *mem benchmarks support the BatchRun mode to enable high-precision memory performance measurement through multiple passes with averages calculated per pass and per run.
 
There are also non-temporal versions of MMX and SSE benchmarks. They have been coded with special instructions to minimise cache pollution on memory reads and to eliminate it completely on memory writes. In addition, they operate with a built in aggressive data prefetching algorithm, though actual behaviour is hardware dependent very much. In a matter of fact, use of non-temporal code allows for significant performance improvements over regular MMX and SSE benchmarks. In some cases, non-temporal MMXmark and SSEmark can deliver almost 100% of theoretical bandwidth while reading.
 
There is also RAMspeed/SMP for multiprocessor machines running UNIX-like operating systems. To be absolutely correct, there are two distinct branches: 2.x.x features support for POSIX threads, and 3.x.x utilises System V shared memory for IPC (Inter-Process Communication) and operates with multiple processes. RAMspeed/SMP v2.x.x is developed no longer due to numerous compatibility and performance issues.
 
Here are several screen-shots to illustrate RAMspeed (FreeBSD) v2.5.0 in action (from 6Kb to 11Kb in size each):
 
Introduction
Additional Information (CPUinfo)
INTmark [writing]
INTmark [reading]
INTmem
 
RAMspeed is more accurate than many other benchmarking tools, more customisable, open source, compact, and gives you much more information to analyse. Some people may say that the lack of some graphical interface is a large drawback, but I rather consider that as an advantage.
 
RAMspeed (UNIX) v2.5.1 (November, 2007) — for uniprocessor machines running UNIX-like operating systems. The source code is available for download (70Kb).
 
RAMspeed (DOS) v2.4.0 (September, 2007) — for DOS (and Windows) systems. Both the source code and a pre-compiled executable are available for download (91Kb).
 
RAMspeed/SMP (UNIX) v3.4.1 (November, 2007) — for multiprocessor machines running UNIX-like operating systems and supporting System V IPC extensions. The source code is available for download (73Kb).
 
All of them are distributed under the terms of The Alasir Licence (TAL), a liberal fairly one. The project is maintained currently by Paul V. Bolotoff, though I keep contributing as much as my free time allows. If you have any development related questions, you better ask him directly.

The colour bars and figures below represent actual performance shown by particular systems: RED stands for INTmem, GREEN — for FLOATmem, BLUE — for MMXmem, VIOLET — for SSEmem. Non-temporal MMXmem and SSEmem are shown in CYAN and MAGENTA respectively. Most machines listed below were running either FreeBSD or NetBSD — because they deliver very good performance, because they run on a wide range of computer hardware, because there is no need to pay for them, and because they are just very good operating systems.
 
Many machines listed below were overclocked properly to some degree. These aren't amateur overclocks close to instability but actual production systems. Click on a system description for a complete performance report. Most of these logfiles are pretty old, so they are subjects for phasing out in favour of some new ones.
 
All clock speeds are real. Memory timings shown are CAS latency (tCAS), RAS-to-CAS delay (tRCD) and row precharge time (tRP). If unknown, question marks are. Other timings are not important as much, though set usually to the most optimal values for a memory type given.