kanecvr wrote on 2016-12-16, 19:31:
Just disable L3 (on motherboard) cache when using a K6-2+/K6-III CPU. Disabling that slow cache usually makes it perform better.
Skalabala wrote on 2016-12-17, 09:37:
kanecvr wrote:Just disable L3 (on motherboard) cache when using a K6-2+/K6-III CPU. Disabling that slow cache usually makes it perform better.
Thanks for the reply bud 😀
But I did try and it losses performance 🙁
Disabling on motherboard cache is a double edged sword. If all you're interested is playing period correct 3D games, be it in software mode or hardware accelerated, disabling on motherboard cache on a K6-2+/K6-3 system will improve performance - BUT - it you're interested in running synthetic benchmarks, you'll find that in lots of cases you get lower scores.
The reason is simple - for the K6-2+/K6-3, on die cache runs at CPU speed. With 256KB of fast on die L2 cache, the slow on motherboard L3 cache that runs at front side bus speed becomes a bottleneck. Disabling it will improve the memory subsystem's latency and memory operations speed.
This behavior can be observed on the intel side of things with the 440BX chipset. While a 450MHz Pentium II's performance in benchmarks is overall higher then a overclocked Mendocino Celeron running at the same speed (Celeron 300 + slotket / Celeron 300A @ 100x4.5), the Celeron is faster in most games, especially first person shooters.
This was all over review sites and gaming forums back in the day. The answer is the same as with the K6-2+/3: The Mendocino Celeron has on die, full speed 128kb of L2 cache, running at 300 (in this case 450MHz) while the Pentium II has 512kb of L2 cache on cartridge, running at half CPU speed, 225Mhz. Whenever the Pentium II is using the L2 cache, it has to wait 1/2 clock cycle before it can talk to the cache, reducing the overall speed of the operation. Large bits of data that only fit into L2 cache will be processed at L2 cache speed, not CPU speed. Smaller bits that can fit into L1 cache will be processed at CPU speed. At least that's how I understand it.
That means that the Mendocino with it's fast on die cache will outperform the pentium 2 in some scenarios. In the same scenarios, disabling the slow L3 cache should improve performance on a K6-3.
The question is - what happens when the data that need to be cached is larger then the amount of on-die cache? This scenario is often encountered is synthetic benchmarks. Well, the Mendocino will take a performance hit because it ran out of L2 and needs to read the rest of the data from much slower system memory. The pentium II on the other hand might not - provided the necessary data fits into it's 512kb of cache. Same thing for the K6-3 with L3 cache vs no L3 cache. If the data fits into 256kb L2, it will perform better with L3 disabled (ex - games). If it does not fit, then the extra 512 to 2MB of additional L3 cache will improve performance (ex. benchmarks and other software) . L3 may be slow compared to on die cache, but it's still faster then system memory.
Again, this is how I understand cache.
