Topic: [XMR] JCE Miner Cryptonight/forks, now with GPU! - page 11. (Read 90791 times)

I got reports that on Opteron (old non-aes AMD) i also give +30% perf, so maybe the Nehalem is a special case of non-AES cpu which runs bad on my code

@migo:
Looks like my miner uses 8 threads and stak uses 16, and stak is probably right to do such.
Can you copy-paste the beginning of my log, where it detects the CPUs ?

Otherwise, try this manual config, which is very like the one of stak:

@migo:
Looks like my miner uses 8 threads and stak uses 16, and stak is probably right to do such.
Can you copy-paste the beginning of my log, where it detects the CPUs ?

Otherwise, try this manual config, which is very like the one of stak:

@Unclwish:
try those:



But on a small AES intel CPU, i'm not that faster, when i am. Except on Bittube. It's possible i don't even compensate for the 1.5% fees, assuming you compiled xmrig with no fees.

CN_Turtle Testnet – Now that PoW algo [1] has been announced for block 1,200,000 lets start testing it all out. The project needs some people to run some infrastructure (seed nodes, pool, explorer). We’ll also need some testers to send a few hundred hashes at it and check the miners are all OK as they get updated. In the next day or two we’ll be able to compile the testnet bits [2] and even just run straight from containers [3] then we’ll aim for an upgrade height for around midnight Saturday UTC specifically block 1,060,600. So anyone that wants to help out with infrastructure bits please ping me SoreGums#8071 in discord [4] to get added to #dev_testnet, everyone has access to view this channel if you’re curious as well. – soregums

Surprising.
Just benched, between last JCE and last xmrstak, official release, not recompiled myself.
Using autoconfig to mine Monero v8 in both cases, 4 threads on my Core2Quad Xeon oc@ 2.666G



So that's even more than +30%.

What CPU and algo did you test on?

Sure, online is the 0.33k CPU Linux

Masari: yeah, yet another fork
And again with no reference dedicated miner and no test pool.
Otherwise it looks like just Masari with the v2 additions, so if that's it, it will be easy to add.

I've still found no pool for test on Stellite v8 or Turtle v2.

Sure, online is the 0.33k CPU Linux

Masari: yeah, yet another fork
And again with no reference dedicated miner and no test pool.
Otherwise it looks like just Masari with the v2 additions, so if that's it, it will be easy to add.

I've still found no pool for test on Stellite v8 or Turtle v2.

Hi!

About the Athlon-FX, i'm about to release a CPU version with the backport of the Bulldozer fix and optim, as detected by Unclwish.
Compared to the 33j, it should give you a few extra hashes, on par with the older version 33h

Best CPU algo:
Some maths. The Cryptonight algo is made of three main parts, and a few negligible other:
* Explode
* Scratchpad
* Implode

The 1st and 3rd are simple AES rounds on predictable memory addresses. A non-AES cpu will make it slow, unless using JCE where it would be 35% faster, but still quite slow, and an AES Cpu like the Athlon-FX will make it almost instantly. An AES CPU is like an AES cypher ASIC.

GPU do these steps quite slow, because an EAS round on GPU involves dozens of computations and LDS memory accesses.

The step 2 is slow in both cases.
The CPU are overall very fast, but they compute something like 4, 8, 12 CN hashes at the same time, on a common CPU (ultra-pimped Threadrippers/Xeons of course can do 64 or more, JCE supports up to 256).
GPU do like 500 to 4000 at the same time, so they are overall more efficient.

The trick is to choose for CPU the biggest steps 1 and 3 and the smallest step 2.
The best algo for this are:
* Cryptonight MKT (slightly smaller step 2 than Monero)
* Cryptonight Fast/Masari (better, the step 2 is half of Monero)
* Uplexa (still better, the step 2 is quarter of Monero)

Also, while the normal cache allocation is 2M, Heavy-like algo (Ryo, Haven, Tube...) use 4M and Uplexa use 1M. GPU have such a large memory to work with that the difference is tight, in both case they use all their memory and computing power.
But modern CPUs tend to run out of cache (while old CPUs like Core2 had a lot of cache but lacked cores). Use a twice smaller cache allocation allows to use twice more cores, and often double the performance.

This is why mining Monero on CPU often makes it run at 50% or 66%, while mining on GPU always makes it run at 100%.
GPUs dont' run out of cache because they simply don't really use it (*), they rely on their fast memory. They are designed for this. A CPU has a much slower memory but a very efficient cache.

My advice: technically, mine uPlexa. Just i don't know if its market value will get good, but the algo is fine for CPU.

(*) A typical real-life GPU operation is blitting a 64M texture in a game. In such case, having 256K or 512K cache or even 16M cache is useless. This is why GPUs are designed to work with async uncached fast memory, while CPU are cached and sync.

edit:
Here's the 0.33k CPU Windows (linux comes next)

Very minor revision, with a fix for Bulldozer family on x64 for v8. No other change.

Fork status:
AscendingNight: done but hidden, and likely to be dropped since the devs no longer reply on Github and the only 2 pools are offline. Sad, it was a good CPU algo, like uPlexa.
Stellite v8: done but hidden, i want a real pool to validate my devs, and haven't found any.
Turtle v2: ditto
Hycon: not started yet and i don't like this coin, looks like a normal CN algo with pedantic netcode changes. I may add it but not sure. I don't see a reason to mine it neither on GPU or CPU.