Summary: In heated competition with Intel's Pentium II line of CPUs, AMD has made some dramatic improvements for its K6 line. We compare the recently released K6-III against the older K6-2 and show why cache can really matter.
In competition with Intel's Pentium II class chips, the AMD K6-III CPU(.JPG, 11K) includes a whopping 256 kilobytes of cache RAM. What does this cache do? To put it simply, it's used to bring out all the hidden performance of the CPU and take it closer to the theoretical maximum of performance. It does that by stashing some "often used" data, or by buffering write operations from the CPU, thus compensating for the much slower main memory and motherboard's bus. It allows a fast CPU to do more work, instead of just wait faster.
In this review we intend to find out what kind of performance difference can be seen, compared with AMD's earlier chip: The K6 and the K6-II.
The Test System
My test system's configuration is as follows.
For comparison: AMD K6-2 300
EPoX MVP3E-M Super 7 motherboard with 1MB cache.
1 x 128MB SDRAM DIMM, populated with NEC A80 (8ns rated) chips.
Symbios SYM 22801 Dual Channel UWSCSI card.
IBM DDRS-39130W (9ES), 9GB 7200PRM UWSCSI HDD.
ELSA GLoria Synergy 8MB AGP
Installed on this system is OS/2 Warp 4.0 (Traditional Chinese version) with FixPak 5. Additional applications loaded and running at the time of the benchmark are a typical representation of what you might expect on a normal computer and include the SysBar/2 Task Monitor, Pipe Monitor and clock. No other applications or utilities were running at the time.
Installing the CPU
When installing a CPU you should always apply some heatsink compound onto both the heatsink and the CPU. Contrary to popular belief and some hardware review sites, the heatsink compound is to be applied THINLY, and not generously onto the mating surface of the CPU. These are the steps that I recommend:
The Imagined Problems
There were a LOT of rumors flying around hardware newsgroups about the AMD K6-III sucking your motherboard's power regulator dry due to the high current demand. I'm happy to say that there's nothing to worry about. There were claims that the AMD K6-III 400 12.4 amp thirst would be too much. Maybe so for some older motherboards with linear regulators, but it's highly unlikely that a good, modern motherboard will have any trouble supplying the kind of current that's needed.
The AMD K6-2 400 uses 11.25 amps and the 450 uses 12.5 amps at their maximum current consumption level. In comparison, the K6-III 400's 12.4 amp requirement seems to be a small increase, especially considering that K6-III have over 20 million transistors while K6-2 had just over 8 million.
The Real Problems
The only problem I experienced during the whole testing phase was the occurrence of mysterious crashes at 100MHz x 4 (a motherboard bus speed of 100MHz, with a clock multiplier of 4, driving the chip at 400mhz). The crashes still occurred even if I reduced the multiplier to 3.5 (for a 350MHz CPU speed). Upon contacting the AMD Tech Manager who had supplied me with the CPU, it was explained that there are occasional problems with 100MHz bus speeds and CXT core K6 CPUs on older motherboards. The recommendation was to try a different bus speed and, surprise-surprise, the thing worked as stable as a rock at 83MHz bus speed and a 5x multiplier. I'm not sure if this applies to production K6-III CPUs, but I don't think there should be much problem with most newer motherboards.
This is unfortunate, since it almost always pays off to increase the bus speed as high as possible, even if the ultimate speed of the CPU is lower than with other configurations that use higher multipliers. This is because the faster the motherboard/bus speed, the faster the chip can communicate with all other parts of the system - including memory, video and I/O. So for example, a computer with a CPU running at 330MHz on a motherboard running at 66mhz (with a 5x multiplier) will be out-performed by a computer with a CPU running at 300MHz but with a motherboard clocked to 100MHz (and a 3x multiplier). Even though the first one has a chip going at 330MHZ, it's laboring under a 66MHz bottleneck and isn't talking to the memory and video as fast as the second computer can.
Without a good benchmark that attempts to simulate real-world usage I settled on SysBench Version 0.9.4b for OS/2.
First I would like to point out that file I/O tests have shown some strangely inconsistent figures, (like faster I/O with K6-III at 373.5MHz than 415MHz) so I have removed that section from this review.
Graphics performance shows an average improvement of only 2%, this may be due to poor drivers from ELSA than anything else.
CPU performance is much more interesting, it shows an 7.4% increase in performance compared to K6-2 at the same clock speed and external bus frequencies. Even the floating point performance went up a little bit, by about 1.5%.
Dive performance shows the difference in AGP bus speed, I believe that EPoX engineer played safe and preset AGP speed to 2/3 divider if the CPU bus speed is set to 83MHz. As you can see, the slowest speed K6-2 won because of the advantage in raw AGP speed. It also showed interesting performance between the K6-2 and K6-III clocked at 373.5MHz, an increase of 5.6%. Even with the external L3 cache turned off, the K6-III beat K6-2 at the same frequency with a performance difference of 1.7%.
Memory performance is where K6-III starts to glow. At 373.5MHz, the K6-III surpassed the K6-2 by a whopping 28.7%! Even without the help of external L3 cache, it still hold an advantage of over 23.3%.
The detail of the memory benchmark below shows the reason why it is recommended that the K6-III system should still be equipped with a cache.
With external L3 cache enabled, you can see that the performance comes down in a ladder fashion. The first level, which shows over 1.2GB/s is the level of performance when CPU can simply access the L1 cache. All K6-2/K6-III CPU have similar result there until it hits 40KB, when the 32KB of L1 cache's data section is overfilled. As you can see, the K6-2's memory performance sharply drops starting at the 40KB mark. With the K6-III, it manages to slow the drop of performance a little longer, with several times the memory performance that the K6-2 have. At 320KB, the performances of K6-III drops again, this performance drop in K6-III is due to the fact that the on-die 256KB L2 cache is no longer large enough, thus the performance drops again. After that, both CPUs faced another drop at 1200KB level, where even the motherboard's 1MB cache isn't enough to be of help.
Figures given are SysBench totals for each category. Higher numbers are better.
Well, I might be a little biased, but I'm honest to goodness surprised that I can see such dramatic improvement on memory performance compared to the K6-2. It's no wonder that AMD is happily demonstrating the K6-III 400 beating Pentium II 450 systems wherever they can.
For a lot of Super 7 power users who need more than 256MB of main memory, there's always been a problem about not being able to find a 2MB cache equipped motherboard. There's only one that I know of that comes with 2MB cache, and even that has a few drawbacks compared to other popular Super 7 motherboards. Without the 2MB cache, most, if not all Super 7 motherboards will not be able to cache memory over 256MB. Due to OS/2's design, the memory is being used from top down, and because of that the non-cached memory will be used first, dramatically slowing down the system's performance. Cry no more! The AMD K6-III is here to the rescue: it allows you to use memory out of motherboard's cache boundary with an almost un-noticeable decrease in performance. With the recent price cut, I'm happy to say that I strongly recommend the AMD K6-III to those who would like to upgrade or build a new system.
I'd like to thank Country Technical Manager, Mr. Dasga Lin of AMD Taiwan for lending me K6-III CPU for the review. Also, look for my review of the AMD K7 in the coming months.
|Copyright © 1999 - Falcon Networking||ISSN 1203-5696||June 1, 1999|