mrau wrote:

hi there, […]
Show full quote

hi there,

so i have an fx 6200 as main rig and it is kinda suboptimal;
i noticed the l3 cache is very slow, even main memory has more throughput;(but it cannot get used, since the l3 cache limits the data flow...)
does anyone know if there are any tools to disable the l3 cache on this cpu? id like to try and see if the main memory accessed directly would not perform better;

thanks

mrau

I don't know of any tools to disable the L3 cache, but what I know is that even if it somehow is slower than the RAM, it has a much lower latency, so it should not be hindering performance in any way.

i know the latency is better than with ram, i dont remember the exact values, but l3 cache here has only about 10gb/s throughput, this is mediocre, i cannot believe this wouldnt slow down everything, my very very common kingston cl9 value ram is a lot faster even;

What program are you using to measure the cache speeds? I have an Intel Core i5, where the L3 cache is on a seperate die inside the CPU package, and I get over 44GB/s in copy operations. The FX series should have the L3 cache in the same die as the CPU itself, and therefore they should be much faster. The software I am using to test is Aida 64 BTW.

I believe it is impossible to disable L3 cache on FX CPU. It wouldn't help anyway... FX 6200 is slow overall, not just the cache. You can try to overclock the CPU and cache as well - 2.4 GHz should be possible on L3, maybe even little more.

btw - Core i5 always have on-die L3 cache. Clock depends on the i5 gen - SB/IB has full speed (same clock as cores), Haswell and later runs cache a little slower. The only separate-die cache is L4 on Broadwell and some BGA variants of Haswell and Skylake.

One little thing - there is no fx 6200. There is either 6100 or 6300.

Actually, FX 6200 indeed exists - https://en.wikipedia.org/wiki/List_of_AMD_FX_ … .2832_nm_SOI.29 Not as common as 6100 or 6300, but still a real CPU.

http://www.cpu-world.com/CPUs/Bulldozer/AMD-F … %20FX-6200.html

i was using aida too, to read the cache speeds (when i was on windows);

i know intels caches are faster, a _lot_ faster, thats why they can make proper use of fast memory (i guess);
if all data goes through l3, then l3 speed limit, limit the entire dataflow - doesnt that sound logical?

edit: i lost the screenshots i made back then, but i see on google that others get similar results, which is even as low as 10gb/s on l3 cache..
also the newer generation vishera seems to have improved cache reaching some 35gb/s
latencies are still similar, half of what is declared for the main memory

Oh, I guess it does exist. My bad.

havli wrote:

btw - Core i5 always have on-die L3 cache. Clock depends on the i5 gen - SB/IB has full speed (same clock as cores), Haswell and later runs cache a little slower. The only separate-die cache is L4 on Broadwell and some BGA variants of Haswell and Skylake.

Yeah sorry, I confused the separate memory controller die to the L3 cache in Clarkdale i5 processors.

AMD L3 cache speed varies with NB frequency, to increase by more than a single step tends to need a touch of CPU-NB voltage.

My 965BE is currently running with DDR3 1600 and the NB at 2400 (3x the "half clock" of the RAM), needed two clicks of CPU-NB voltage to stabilize it

Another reason why I had stayed with Phenom 2 before switching over to an i5 back in the day and I was pissed with that shitty performance the FX series gave. I don't know how FX procs behave when overclocking other than they are limited on the memory controller and they use a lot of power but that is the option you have beside upgrading. Switching to Intel things are just easier all the way around even the first gen i5/i7s overclock with ease and still perform ok for modern games.

If you ever get into the market for something that has some grunt to it looking into doing a low cost Xeon 2011 build as the procs can be had on eBay at a low cost even current gen 3 and gen 4 xeons. Gen 3 and 4 have the advantage as X99 boards are upwards of $100 cheaper than X79 so you can buy a decent amount of DDR4 with the savings. Rocking a 12 core that only cost $105 plus they are not power hogs unlike some AMD procs 😉

You can get 12 core Broadwell Xeons for little as $90 us from China but just be sure that in the listing it says "X99 compatible".

The catch is these 12core xeons are clocked at less than 2 GHz, which makes them quite slow for regular desktop.

the l3 even on Bulldozer is faster than your ram, it's a positive for performance.

havli wrote:

The catch is these 12core xeons are clocked at less than 2 GHz, which makes them quite slow for regular desktop.

Not really unless you run something that isn't threaded well and in that case it wouldn't run very well on an AMD proc either, in threaded workloads like FTK, adobe, and others so those cores get the job done. Plus a lot of these xeons are pretty efficient these days.

Running forensics software even on a modern quad core is Painful like running vista with a lot of malware on a netburst celeron /vomit.