Indeed, large on-board (Motherboard) Cache is actually something I really miss...

That would be a nice buffer between the slow RAM, and the (comparably) lightning fast CPU's.

(I can only speculate that the CPU manufacturers do not want this, because it would tend to 'blend' the effective performance of their different CPU lines, which they so far carefully placed into their respective price segments)

Hm ?

Whenever I disabled the L2 (or L3) Cache of Super7-Motherboards, performance would always drop by as much as 25-40%...

With fast L3 Cache, you can still increase performance quite significantly, and all professional Workstations or Servers for example always sport very large Caches for that reason (not neccessarily located on the CPU itself).

I operate 2 DEC Alpha 21164's, one with 4MB Cache (on Motherboard), while the other has none.
The Result is, that the machine with the 4MB is about 90% (!) faster than the one without it in Memory-intensive Situations...

The old rule is still valid :
Cache is life !

The more, the better, since Memory still remains by far the slowest Link in all modern machines, drastically hampering the CPU...

Since you mentioned "synchronous" Frequencies :
That's true for the (slow) FSB, but even at 200MHz DDR (400MHz effectively), a 2000MHz CPU is doomed to spend more than 4 out of 5 cycles waiting for the Memory to deliver.

In terms of Programming, a single "Cache Miss", or (even worse) a pipeline Stall, forcing it to re-aquire alot of Data directly from RAM, is regarded as the worst performance catastrophy possible for a CPU.

Also, Latencies are still a major factor, since even the best PC400 Modules can barely deliver more than a short burst of Data at its maximum speed. The remaining time, it has to waste time itself to Refresh Cycles, CAS latency, RAS-to-CAS delays etc. etc...
------------
Bottom line is :
For as long as a 2000MHz CPU cannot access RAM clocked at 2000MHz, CL0-0-0, there will always be the need for Cache.

PS.
If your BIOS allows to disable L1 and L2 Cache, try to have some fun disabling both. Then, the CPU is entirely limited to RAM performance...
You will be in for a big surprise, no matter what CPU you test, and no matter how fast your RAM is

That's a really nice eye-opener about actual, true RAM performance, compared to what the CPU could actually do if this limit wouldn't hold it back all the time *g*

I did own a K6-III System, and (as I said) the Performance would drop significantly without the L3 Cache.

Of course, this is only valid for Application Routines that don't fit inside the Cache, but most Applications are really happy to have 512k to 1024k nowaday's, and more Cache doesn't hurt either.

I agree, though, that having it run On-Die is the ideal way of doing it, yet it's the most expensive way, making CPU's more difficult to build...

If there was any Athlon Chipset with L3-Cache, I would immediately buy it !