The memory issue: 66 vs. 100 MHz bus
The first thing to consider is the memory question. Motherboards based on the ALi Aladdin V chipset usually are slightly faster than MVP3 chipset boards when equipped with 512 KB L2 cache. However, equipped with 1 MB of L2 cache, boards with the VIA MVP3 chipset currently represent the top of socket 7 performance. And they have another nice feature: The SDRAM clock can be either synchronous to the system clock or asynchronously, at the same frequency as the AGP bus. As you can imagine this allows more independence from the memory: On the one hand you can still use your tried and tested EDO which certainly was all but cheap still some years ago or 10/12/15 ns SDRAM memory. On the other hand you don’t necessarily have to go for PC-100 memory just to achieve an overall performance gain of approximately 3% which hardly anybody will notice.
The follwing system consists of an AOpen AX59Pro with VIA Apollo MVP3 chipset and 512 KB L2 cache, 64 MB RAM, ATI Xpert@Work 4 MB AGP and Quantum Fireball ST 3.2 GB Ultra-DMA 2 harddisk.
The domain of AMD’s K6-2 certainly is Windows 95 or 98, so I gave up benchmarking in Windows NT. The K6-2 is meant to be a gaming and multimedia CPU, and the only operating system which is fulfilling these demands is Windows 95/98.
You may ask why I didn’t benchmark with a faster 3D card. I could have also used a Voodoo2 card, but the gaming performance becomes so high – particularly in Quake II – that the interesting differences would perish, and since I didn’t want to test 3D accelerators but the system at divergent configurations I gave up doing this.
Reference is this system running at 100 MHz FSB and 100 MHz SDRAM clock. If you still have standard 10 or 12 ns SDRAM DIMMs (PC-66) I would recommend to use them, since you can see similar resluts as with Pentium II systems: What primarily counts is the L2 cache; it should run as fast as possible. The memory performance is less important for business applications. SDRAM clocked at 66 MHz instead of 100 leads to less than 3% performance decrease, in my eyes just peanuts. If you use even slower EDO memory, the loss is around 7%, but still no reason to worry about.
The situation becomes interesting for all Quake II gamers: If you don’t take advantage of the 100 MHz memory option, Quake II runs at 10,0 frames per second instead of 12,0 fps with the higher bus speed, that’s about 16% slower. You even lose 26% performance with EDO memory. Using a 3Dfx add-on card eases this performance deficit, but if you had already thought about getting new RAM as well, this could convince you that maybe it’s time for some new memory.
Incoming is quite memory independent: The loss is narrowly 2% in case of 66 MHz SDRAM and around 6% for EDO RAM. I think you can easily forget about those small differences.
512 KB or 1 MB 2nd Level cache
Motherboards with 1 MB of L2 cache should be preferred because of two important benefits: First the system performance rises a bit, and second the larger cacheable area helps to keep the high system performance you are used to also with larger amounts of main memory.
I also ran the benchmarks you saw above on a FIC VA-503+ and 1 MB L2 cache. The performance loss you will have to expect when using slower clocked memory (66 MHz instead of 100 MHz) is smaller in deed. The business performance measured with Winstone 98 does only decrease by 2% instead of 3% with only 512 KB, Quake II runs about 10% slower instead of former 16% and Incoming becomes 1.2% slower (about 2% at 512 KB). The results with EDO DRAM show roughly the same relation.
This shows very well the advantage of having more cache: A larger L2 cache does not only improve the system performance, but also moderates the negative performance impacts when using older memory or SDRAM which doesn’t run at 100 MHz.
Back to the roots: 66 MHz bus speed
I can imagine that many people are considering to buy a K6 or K6-2 CPU to upgrade their system. Some older motherboards can be switched to 2.2 or 2.1 Volts, sometimes not described in the manual, and they might have a BF2 jumper as well. If then the manufacturer also provides a new BIOS on the website, you have the chance to upgrade the CPU. Many of you may ask whether this makes sense at “only” 66 MHz bus speed, since the K6-2 benefits substantially from the higher bus speed.
Let’s get back to our reference performance with the 512 KB AOpen board:
The difference in Winstone between a K6-2 300 at 66 MHz x4.5 and the maximum performance at 100 MHz x3 is about 1.9 points, this is about 9% less with SDRAM and 13% with EDO DRAM.
Quake II runs about 14% slower on a 66 MHz system than on a system running at 100 MHz FSB. By the way, SDRAM memory is much better for the memory sensitive Quake II: A system equipped with EDO is approximately 16% slower. Again, using a powerful 3D card would compensate the smaller bandwidth of EDO.
The difference between 66 and 100 MHz bus clock with the 3D game Incoming is only about 3%. And you can see the better performance of SDRAM memory once more: In a standard 66 MHz computer using EDO instead of SDRAM has an effect on Incoming by 11%.
Summary Memory Speed
Please take into account that the differences between EDO and SDRAM can vary on other motherboards (and with other chipsets as well) and the results aren’t necessarily the same. These tests are meant to show you the practical differences. It could become even more interesting with motherboards equipped with 2 MB of L2 cache.
You can see that the L2 cache is the most performance relevant factor. As you know that is also the reason for the Pentium II not performing really much better at 100 MHz memory clock instead of 66: The L2 cache already runs at a much higher level and makes the memory question today not as urgent as Intel wants us to believe. The benefits of a higher memory bandwidth can be interesting for very memory intensive applications or the AGP bus, which should be able to gain from the higher bandwidth.
For socket 7 systems you should select your key components (motherboard, CPU, RAM) according to this order:
- External clock speed (system clock speed)
The higher, the better. But be careful with 112 MHz bus speed; I can only recommend this setting for people who know what they are doing. - CPU clock
Again, the faster, the better. - Memory clock speed
As you have seen, 100 MHz of course is faster. It depends on you whether this is worth buying new memory or not.
(Over)clocking – 266 MHz
There are some questions which lead me to do some additional benchmarks: Is the performance difference between a K6-2 and a Pentium II at higher clock speeds still the same as with 266 or 300 MHz? How clock-sensitive is the 3Dnow! Instruction set? Does the K6-2 become faster than a Pentium II at even higher clock speeds or will the faster L2 cache of the Pentium II be more advantageous?
The 266 MHz version of AMD’s K6-2 cannot benefit from the 100 MHz bus speed; that’s the main reason for the quite low benchmark results. The deficit compared to a Pentium II 266 is about 11%. Clocking it with 250 MHz and 100 MHz bus speed is the alternative – and the performance results do not differ so much now: The K6 is only behind by 5%.
This chart shows incoming to be memory clock sensitive. Clocked at 266 MHz, the K6-2 is about 4% slower than a Pentium II 266. At 100 MHz bus speed and 250 MHz, the K6-2 is even faster by 5%.
Quake II is also bus clock sensitive: The 66 MHz setting is 15% slower than the Pentium II, even with the special Quake II 3DNow! drivers. The defecit shriks to little 2% with 250 MHz and 100 MHz FSB again.
(Over)clocking – 300 MHz
The fastest CPU settings according to AMD’s specifications allow 100 MHz bus speed, which reduces the windows performance gap to the Pentium II by 4%.
The performance difference with Incoming is almost nothing (the K6-2 is about 1% slower)
Thanks to the 3DNow! drivers the difference with Quake II is about only 4%.
(Over)clocking – 350 MHz
The new K6-333 CPU has still some performance margins; this means overclocking it to 350 MHz should be successful with almost all 333 MHz CPUs, maybe with a 300 MHz type, as well. You should be familiar with the rules of overclocking from Tom’s guides, thus I think it’s not necessary to mention the risks and insecurities again. All others should wait for the official 350 MHz version which is to come still in fall 1998.
The Winstone differences between the two rivals were 5% at 250 MHz (the second official setting for the 266 MHz version), 4% at 300 MHz. The question was: Will the difference decrease to 3% or less? The answer may disappoint: The difference has grown to 7% due to the even faster L2 cache of the Pentium II 350 (175 MHz) and the 100 MHz memory.
Also incoming profits of the new features of all new 100 MHz Pentium II CPUs. The K6-2 at exactly the same clock speeds is 4% slower. Though the Pentium II couldn’t outperform the K6-2 at 300 MHz, he will do this at higher clock speeds.
Quake II runs slower by 11% at 800×600 at 350 MHz. The Pentium II is clearly faster here.
(Over)clocking – 112 MHz bus speed
Some motherboards (e.g. FIC VA-503+) allow 112 MHz bus frequency. Please consider that PCI and AGP cards as well as all on board components are being overclocked with this setting. As we’ve seen above, the L2 speed is the key to high system performance. To ensure the highest possible performance, I took the FIC motherboard with 1 MB L2 cache again. Now let’s see what we can expect:
You can see that the K6-2 clocked at 336 MHz is only 4% slower than a Pentium II 350 MHz and about 3.5% faster than a Pentium II 300 MHz. Both L2 cache and SDRAM are clocked with 112 MHz and help to increase the windows performance.
The K6-2 is only slower by 2% compared to the Pentium II 350 and approximately 4% faster than a Pentium II 300 MHz. This difference comes mainly from the high bus speed of the K6-2 during these tests.
Everybody knows that Quake II just wants pure performance, no matter how. The fast bus and memory speed helps the K6-2 to be only 2.5% slower than a Pentium II 350 and 15% (!) faster than a 300 MHz Pentium II. Considering that a K6-2 333 MHz clocked at 336 MHz and with AMD’s 3DNow! support provides Quake II performance at the level of a Pentium II 350 MHz I think that’s the ultimate Quake II gaming machine for people who cannot afford buying a Pentium II.
(Over)clocking – Summary Clock Speed
The benchmarks show the K6-2‘s qualities to compete with the Pentium II at medium clock speeds (266 to 333 MHz), but it also makes pretty clear that the Pentium II will always take advantage of its higher clocked L2 cache. That’s why it’s almost impossible that a K6-2 400 or 450 will be able to achieve an overall performance comparable to the Pentium II, only games could profit from the 3DNow! instructions. AMD knows about this and also plans to equip future CPUs with an integrated 2nd level cache. The move to 100 MHz bus speed was important for socket 7 and of course enables much higher performance. But we should not forget that those performance gains mostly come from the L2 cache which benefits tremendously from higher clock speeds.
It’s not the main memory which should be clocked as fast as possible; it’s the L2 cache which should run at a clever relation to the CPU clock speed. Using the external clock for the L2 cache will never enable highest performance (socket 7), and clocking it at full CPU clock is fast, but very expensive (see Pentium Pro or Pentium II Xeon). It seems as if half CPU clock seems to be the best compromise between high performance and costs; Intel also plans further CPUs based on this concept (Katmai, Coppermine, both Slot 1).
The K6-2 still has great chances to become very popular. DirectX6 is about to be released, and many game software developers have agreed to support 3DNow! Intel is also working on a new instruction set which will possibly be called “MMX2”. If AMD succeeds in establishing large amounts of CPUs and convincing enough software developers to support 3DNow!, the way for winning customers from Intel may be paved. Let’s hope that sun of 3DNow! doesn’t set before its maximum blossom.
