Topic: [XMR] JCE Miner Cryptonight/forks, now with GPU! - page 94. (Read 90814 times)

L1 -> Level 1
L2 -> Level 2
L3 -> Level 3

as the levels get higher, the cache is slower but larger. Maybe AMD is using some clever technique so L1/L2 caches are more shared than they are already, so mentioning it is useless. But absend L1 or L2 cache would have a huge impact on performance.

In wikipedia no info about L1/L2 cache on Epyc. Only L3. Does L1/L2 cache exists on it or not? Just interesting...

P.S. 2Mb L3 cache per thread allready too low... AMD can add more L3 cache on Epyc... Especcially if L1/L2 cache is absent...

When I try the latest version on an AMD Epyc (basically four Ryzen dice on a package) I get the following error:

Thread 30 successfully bound to CPU 30
Allocated 2MB Cached Large Page Scratchpad Buffer for CPU 30 of NUMA node 3 at: 0000021b3c200000
Starting CPU Mining thread 31, affinity: CPU 31
Thread 31 successfully bound to CPU 31
GetNumaProcessorNode failed for cpu 32, error code: 87
Retrying with no NUMA
Allocated 2MB Cached Large Page Scratchpad Buffer at: 0000021b3c400000
Connecting to mining pool support.ipbc.io:17777 ...
Devfee is 1.5%

That's strange because I only defined threads for the CPUs 0 to 31 in the config file.   Why does the miner try to access a CPU (core) 32, which is not present? Apart from that the miner starts mining and has better hashrate than XMR-stak with Bittube: 5200 H/s vs 4900 H/s.  

thanks for the test, it seems my autoconfig is bad on the A10, it allocates too many threads. I'll fix it, thanks.

The multihash (double hash is the 2-case, you can set from 1- to 6- ) is also called low-power in stak IIRC. It's about using, on one CPU core, twice the register and twice the cache to get sometimes twice the speed. The trick is that it let the other cores free, so it consume less power and allows the Turbo to enable, for CPU with turbo.

Technically, it's good, sure when you want to save power, but also when you run out of cores and not of cache. If you have a CPU with 2 cores but 8M cache, normal config would give only 2x2M = 4M cache used.

you may enable double-hash to use 2x2x2M cache = 8M of cache, and get some extra perf.
It works more or less depending on the CPU. It's very efficient on Ryzen, and not at all on Core2.

I looked closer at the A10, and yeah that's a little APU with little cache.
I give you an experimental config that could let you get some extra perf, but not sure, i cannot test, i've no A10.

"cpu_threads_conf" :  
[  
     { "cpu_architecture" : "auto", "affine_to_cpu" : 0, "use_cache" : true },    
     { "cpu_architecture" : "auto", "affine_to_cpu" : 1, "use_cache" : false },
     { "cpu_architecture" : "auto", "affine_to_cpu" : 2, "use_cache" : true },
     { "cpu_architecture" : "auto", "affine_to_cpu" : 3, "use_cache" : false },
]

now finishing 0.29c, the last CPU-only version, with some updates.