K6-3 Will Bring New Life to ‘Old’ Motherboards

The past few years have seen a dramatic reduction in the prices of motherboards, memory and processors and at the same time have at least doubled system performance. While it would appear that this trend should have reduced the cost of upgrading, the fact is that many of the improvements in processors have forced users to upgrade their motherboards and perhaps memory at the same time. Because of the close connection between these components users are still having to spend several hundred dollars to upgrade their systems to the next performance level even though the individual components continue to drop in price very rapidly. Many users have resigned themselves to the fact that upgrading almost requires purchasing a new motherboard, CPU, and probably new memory.

Apparently, AMD recognized that there may be a number of potential customers who would like to upgrade, but who cannot justify the relatively large expediture to gain perhaps 10% or 20% in performance. These users are waiting for prices to drop even further before making the jump to 100MHz motherboards and PC100 memory. In an attempt to tap this potential market, AMD released their K6-2 processors with 66MHz bus speeds, allow them to jump from perhaps a 233MHz processor to a 333MHz using existing components.

While sales figures for these chips has not been released, the big potential flaw in the concept is that the performance difference between a 233MHz and 333MHz chip both running at 66MHz is barely 12%. There is almost no price difference between the same speed processors for 100MHz and 66MHz bus speeds, and the 100MHz processors can be run at 66MHz in any case, making these new offerings even less attractive.

Enter the Sharptooth (or K6-3), which will have 256K of L2 cache on-chip. This chip will, according to AMD, be fully compatible with standard Socket-7 (and super-7) mobos, assuming they have the proper voltage settings. Essentially, all motherboards in this category incorporate L2 cache, which becomes an L3 cache once the K6-3 is installed. After having some discussions with knowlegable users, reading some comments from AMD spokespersons and doing some testing with currently available products it seems likely that this processor will be the one which allows users to upgrade for a minimal outlay. Most of the ideas and comments in this article were obtained directly from some of these discussions.

In the Intel world, there are no L3 caches. Intel deliberately eliminated motherboard caches from chipsets designed to support the PII, apparently with the idea that the L2 cache on the Slot-X module is all the cache that is necessary for good performance. There is no indication that Intel has re-thought this design decision, even though (as mentioned in our news report “Intel to Abandon Slot 1?”) Intel is abandoning the Slot-X module in favor of a Socket-370 (?) approach that is functionally identical to Socket-7 except for the P6 bus compatibility and MPU (multiprocessor) capability.

In the high-end workstation world, L3 caches are common. DEC (long before its acquisition by Compaq) was routinely shipping Alpha workstations with 8 megabytes of L3 cache, for the simple reason that the 8 meg cache made its workstations run a lot faster. On the other hand, PC OEMs are continually trying to find methods for cutting their costs, which means that some, or all, of them may elect to eliminate the motherboard cache for their K6-3 based systems. With just a little bit of analysis, this idea does not seem so farfetched.

Consider two AMD-based systems which are otherwise identical: System A has a K6-2 on a Super-7 mobo using 1M L2 cache. System B has a K6-3 on a Super-7 mobo with no mobo cache; there is a 256K L2 cache on-chip. System A’s L2 cache is larger and system B’s cache is faster, but on average, the two systems run at the same speed. We know this because of comparisons of Celeron-A/333s with PII/333 systems, which are almost exactly equivalent to our hypothetical A and B AMD systems.

Benchmark results were obtained with the following configuration: AOpen AX59Pro w/1MB, a K6-2 300, 64MB Crucial Technology PC100 SDRAM, W.D. 34300 UDMA HDD, Toshiba XM-5704B CDROM and a Diamond Stealth II S220 with v4.10.01.101 drivers. There was no attempt to optimize BIOS settings, however each test was run with identical settings with the exception of disabling/enabling L2 cache. By changing the jumpers to run the CPU and/or SDRAM at both 100MHz and 66MHz, while also toggling L2 cache on and off we can see the overall effects on the performance. Note that running the SDRAM at 66MHz, while running the system bus at 100MHz means that L2 cache is also running at 100MHz, as the following diagram proves.

The following table illustrates the Winstone 98 Business scores for each configuration:

These numbers show very clearly that with L2 cache enabled and running at 100MHz, the difference between running SDRAM at 66MHz and at 100MHz is 3% or less. With a full speed L2 cache, as the K6-3 will have, this difference would be even smaller. The numbers also show that since L1 cache runs at full CPU speed, there is literally no difference between the 100MHz and 66MHz sytem bus settings when there is no L2 cache (bottom two cells). The addition of L2 cache on a system running both CPU and SDRAM at the same speed results in approximately a 25% improvement, whether at 100MHz or at 66MHz, while simply increasing the SDRAM speed with no L2 cache provides almost a 15% performance gain.

What all of this illustrates is that the performance gains are realized by speeding up the heirarchy level that is satisfying most of the ‘hits’ for CPU demand accesses. If those hits are satisfied by L2, then the SDRAM speed itself is of little consequence. It would therefore stand to reason that even a K6-3 system with no L3 cache and 66MHz SDRAM speed would still match or outperform a K6, K6-2 or even a Pentium II with the same speed processor on a 100MHz system bus due to the full speed L2 cache. A 1MB L3 cache running at 100MHz would provide even more of an improvement than simply running the SDRAM at 100MHz, because the access is approximately 3 times faster. This could mean an additional 5% to 10% or more when all components are running at 100MHz.

Though all of this information is not new or revolutionary, what has not yet been discussed is the performance implications for those with motherboards that do not support a 100MHz bus speed. Owners of such motherboards as the FIC PA-2007, AOpen AP5T/AX5T, M Tech R581A, Abit TX5 and other boards with 2.2v support will still be able to gain a substantial advantage with the K6-3 without even requiring new memory. As can be seen with the above table, simply increasing the L2 cache speed from 66MHz to 100MHz gives a 7% performance boost, so increasing it to full CPU speed will gain even more – perhaps as much as 15% to 20%. With a K6-3 running at 5.5x66MHz (366MHz) with 66MHz SDRAM or even EDO, users should actually get performance that compares very favorably to a current K6-2 or PII 400MHz without having to purchase a new motherboard or memory. For the first time in quite awhile, a manufacturer has apparently provided an upgrade solution that really *will* save money