Originally Published: Tuesday, 30 May 2000 Author: Brian Richardson
Published to: enhance_articles_hardware/Hardware Reviews Page: 1/1 - [Printable]

The Great Camino Benchmark: Part Three

Remember what I was doing? Of course not ... the MTV generation has the attention span of an over-caffeniated Kiki The Ferret. My goal was to benchmark an Intel i820 chipset-based motherboard using both SDRAM & RAMBUS memory to compare the two memories under LINUX.

   Page 1 of 1  

Remember what I was doing? Of course not ... the MTV generation has the attention span of an over-caffeniated Kiki The Ferret. My goal was to benchmark an Intel i820 chipset-based motherboard using both SDRAM & RAMBUS memory to compare the two memories under LINUX. The SuperMicro PIII-SCA motherboard offers both RAMBUS & SDRAM DIMM sockets, making this pretty easy. The "baseline" comparison system is an AMI S781 motherboard, using the ubiquitous Intel 440BX chipset.

On both systems, the following components were used:

  • Intel Pentium II 400 MHZ CPU (Deschutes, 4.0 X 100 MHz)
  • DIMM: 64MB PC100 memory
  • RIMM: Samsung 64MB RDRAM @ 300 MHZ clock (PC600)
  • ATI Rage Pro AGP 2X (8 MB) video adapter
  • Seagate 20.4 GB Hard Disk Drive (UDMA/66, 7200 RPM)
  • Toshiba 32X ATAPI CD-ROM
  • Intel EtherExpress PRO/100 Network Adapter

It's also important to note that I used an 80-pin IDE cable on the SuperMicro board, since the system supports UltraDMA 66.

So what's the story? My two benchmarks of choice are BYTEmark v2 and UNIX Bench 4.0.1., both ported to a variety of UNIX flavors. I used the pre-compiled binary that comes with BYTEmark (nbench), but I compiled UNIX Bench under Mandrake 7.0 (it only comes as source).

The results are listed below. The first set of data comes from the SuperMicro motherboard running standard PC100 SDRAM memory (64 MB). The second set of data is from the same system, using RAMBUS memory at a 300MHz clock (a.k.a. PC600 ... RAMBUS uses a double-pumped clock). The third set of data is from the AMI motherboard, based on the Intel 440BX chipset ... it's using the same SDRAM as the first SuperMicro test. The benchmarks were run from a shell (no X-Windows).

Benchmarks for Corel LINUX 1.0s BYTEmark* LINUX/UNIX v2 (10/95)

Memory Index (LINUX)

  • i820 SDRAM: 1.688
  • i820 RDRAM: 1.687
  • 440BX SDRAM: 1.677
Integer Index (LINUX)
  • i820 SDRAM: 1.46
  • i820 RDRAM: 1.46
  • 440BX SDRAM: 1.457
Floating Point Index (LINUX)
  • i820 SDRAM: 2.695
  • i820 RDRAM: 2.702
  • 440BX SDRAM: 2.67
Benchmarks for Mandrake LINUX 7.0 BYTEmark* LINUX/UNIX v2 (10/95) Memory Index (LINUX)
  • i820 SDRAM: 1.689
  • i820 RDRAM: 1.691
  • 440BX SDRAM: 1.679
Integer Index (LINUX)
  • i820 SDRAM: 1.46
  • i820 RDRAM: 1.461
  • 440BX SDRAM: 1.458
Floating Point Index (LINUX)
  • i820 SDRAM: 2.648
  • i820 RDRAM: 2.657
  • 440BX SDRAM: 2.622
UNIX Bench 4.0.1 (index scores) Arithmetic Test (type = double)
  • i820 SDRAM: 33.3
  • i820 RDRAM: 33.3
  • 440BX SDRAM: 33.3
Dhrystone 2
  • i820 SDRAM: 73.4
  • i820 RDRAM: 73.4
  • 440BX SDRAM: 73.4
Excel Throughput
  • i820 SDRAM: 160.8
  • i820 RDRAM: 162
  • 440BX SDRAM: 159
File Copy (1024 buffsize, 2000 maxblocks)
  • i820 SDRAM: 170.6
  • i820 RDRAM: 188
  • 440BX SDRAM: 190.3
File Copy (256 buffsize, 500 maxblocks)
  • i820 SDRAM: 205.5
  • i820 RDRAM: 215
  • 440BX SDRAM: 222.7
File Copy (4096 buffsize, 8000 maxblocks)
  • i820 SDRAM: 6.5
  • i820 RDRAM: 6.5
  • 440BX SDRAM: 5.6
Pipe Throughput
  • i820 SDRAM: 184
  • i820 RDRAM: 186.3
  • 440BX SDRAM: 184.7
Process Creation
  • i820 SDRAM: 255.5
  • i820 RDRAM: 258
  • 440BX SDRAM: 258.6
Shell Scripts (8 concurrent)
  • i820 SDRAM: 125
  • i820 RDRAM: 128.3
  • 440BX SDRAM: 128.3
System Call Overhead
  • i820 SDRAM: 149
  • i820 RDRAM: 149.2
  • 440BX SDRAM: 149
Final Score
  • i820 SDRAM: 97.6
  • i820 RDRAM: 99.6
  • 440BX SDRAM: 98.3

And this means something? Since there aren't any large variances in the scores from one platform to the next, I can assume that RAMBUS on the i820 doesn't give any great bang for the buck. True, it has acres of bandwidth (1.6 GB/s), but the controller doesn't properly utilize it. Using SiSoft Sandra 2000 under Windows 98SE, the memory bandwidth scores were well below 300 MB/s on the RAMBUS configuration.

Why did you do Windows benchmarks? Good question, kids. Since benchmarking for Windows is more common-place, there are a wider set of "established" tools. Those benchmarks (not published here) told the same story as the LINUX benchmarks ... the cost of RAMBUS isn't worth the slight performance gains.

But RAMBUS seems so cool! I am one of many who think that RAMBUS at its current state isn't all that hot. Some development platforms I've seen utilize 80-90% of RAMBUS bandwidth (under Win9x benchmarks), but they are so secret that I cannot even use their names without fear of verbal decapitation from the lawyers of the nameless company. For the time being, system performance will benefit more from faster hard drives and newer CPUs than RAMBUS. For the cost of 128MB of RAMBUS, one can acquire a pretty smoking AMD Athlon or Intel Pentium III CPU.

Didn't you test UDMA/66? I was tempted to install the UDMA/66 kernel patch and take my Seagate Barracuda for a spin, but I ran out of time (damn employers). Since the next release of Mandrake promises UDMA/66 support by default, I'll wait till it comes out ... then I can compare the performance increase a new distribution can make.

What's next, geek-boy ... some Unreal Tournament? Alas, no. The killing fields will have to wait for another day. After a weekend of yardwork and two days of customer meetings, Brian will start in on some VIA chipset motherboards ... we'll see how the Apollo Pro does under my benchmarking scheme.

Brian Richardson awaits the end of another day in the computer industry. Working from the "Silicon Ditch" of Atlanta, Georgia, Brian proudly displays an "RTFM" bumper-sticker on his over-worked mini-van.





   Page 1 of 1