More fur starting to fly.
Suggestions that Intel may have made a large contribution to ZD and Winstone 99 may have been part of the deal.
anandtech.net
posted 12-01-98 04:56 PM ET (US)
Please do it correct this time!
You do a good job all the time, but this one time you messed up. I think it happens to all of us.
Thank's Anand.
Thoughts on the K6-2 400Mhz & Winstone 99
Those of you who have been visiting the Brotherhood website for a while know that I don't really like to criticize other
people's articles. However, unfortunately, sometimes it's not possible to avoid that kind of thing. Many of you have been
e-mailing us regarding AnandTech's recently published K6-2 400Mhz benchmarks, using Winstone 99. They were intrigued
with the fact that Anand said that without the proper optimization, the K6-2 400Mhz was slower than the 350Mhz unit, and
with the scores presented by him using Winstone 99. Let me tell you that I have a great respect for Anand and for the
work that he has been doing, and I know that he's not lying about the scores that he got. However, what I do question is
Winstone 99 itself. First of all, we had a K6-2 400Mhz unit running here for a few days, without the Write Allocation
optimization, and let me assure you that by no means is the K6-2 400Mhz slower than the K6-2 350Mhz even without the
proper Write Allocation optimization feature. Although the redesigned Write Allocation delivers quite a performance boost,
the difference between Write Allocation enabled and disabled in the older K6 & K6-2 CPUs isn't that much that would
make a 400Mhz unit run slower than the 350Mhz version. Believe me, I've been handling K6's for a long time and I know
what I'm talking about, and I did compare the new K6-2 400Mhz with the older one. The difference made by the old Write
Allocation is so small in terms of performance that it was even inside what we consider to be an error margin, that's why I
said we couldn't come to any conclusions about the new core tweaks (at least not without the proper optimization).
Second, we all know that Intel's Pentium II and Celeron (128Kb) CPUs have always been slightly faster in business
performance than its AMD and Cyrix equivalents. The scores in Winstone 98 seemed to confirm that. However, strangely
enough, even before those scores were posted, I received the information that Ziff-Davis was receiving a large investment
from Intel, and that Winstone 99 was being made to Intel's specifications. As this was just a rumor, without any evidences,
I kind of ignored it. But when the Intel CPUs started showing nearly 10% of performance increase over the AMD ones in
Winstone 99, it certainly smelled funny. I can tell you that I am a heavy user of business apps and multitasking and I have
used both Intel and AMD CPUs and I have never noticed any significant difference between the two of them in regular
Windows tasks. Considering that the CPU is not the only responsible for the system's performance, this 10% difference
should be quite noticeable, at least through other benchmarks.
Anand posted that this difference in Winstone 99 was due to the fact that Winstone 99 makes a high use of multitasking
by loading several heavy apps at the same time, while Winstone 98 only loads a single app at a time, and in a
multitasking environment CPUs that have a faster L2 cache like the Pentium II (which runs its L2 cache at half the CPU
speed) and the 128Kb Celeron (which runs its L2 at full CPU speed) would run much faster than a CPU that runs its L2
cache at the external clock speed (which is much slower than the two above). I haven't even ordered my copy of the
Winstone 99 CD-ROM yet, so I don't really know if Winstone 99 actually loads several apps at the same time. But
Anand's comments are wrong. First of all, it makes no difference for the CPU if it's running one application or several in
terms of caching, because the caching system would work the same way. The difference in performance between a slower
and a faster L2 cache when running two apps at the same time or two apps in multitasking would be exactly the same,
because the caching system is the same, unless these apps were so small that their cached instructions would fit entirely
in the L1 cache when being ran alone, and would make use of the L2 when being used together. But we know that in
real-world that doesn't happen very often. The point with several apps being executed at the same time is that it raises the
area covered by the Principle of Locality, in other words, it makes a better use of the cache. If a certain app only repeats
about 160Kb of its code, then it'll effectively use only 160Kb of the cache. But if we have two of these apps, they would
both use 320Kb of the cache at the same time. As we can clearly notice by this pattern, the difference when it comes to
multitasking is not in the speed of the L2 cache, but rather in the amount of cache of the CPU. The speed of the L2 cache,
in real-world (and I'm talking about the Windows environment as we know it, it could be different in other platforms that
have much smaller apps and very light OSes), isn't really the bottom-line in multitasking, the size of the L2 cache is. And
in that case, the difference between the K6-2 and the 128Kb Celeron should even decrease a little bit because of the
bigger size of the K6-2's L2 cache. In other words, those weird Winstone 99 scores aren't due to the Intel CPUs having a
faster L2 cache. In fact, I don't even believe that Winstone 99's scores reflect the real-world performance of those CPUs at
all. Therefore, until I'm able to test (or to receive additional information on) Winstone 99, I'd recommend you guys to avoid
using it for comparisons. This is not the first time that we've seen benchmarking programs producing absurd scores.
Benchmarks can be designed to show almost anything, if you throw in some specific assembly coding you can make a
486 CPU apparently outperform a Pentium one... it all depends on how the tests are made. So, I'd recommend all
unexperienced users to stick with the benchmarks that we recommend. There are probably benchmark programs out there
that are much better than the ones that we recommend, however, it's very risky to trust a sole benchmark and make your
choice when purchasing a CPU based on it.
[This message has been edited by James (edited 12-01-98).] |
