Thanks for the tip, I go from 248 to 256 kH/s with -i 1 changed to -i 0 on my 570.

My son's 670 went from 142 to 156 with -i 0 instead of -i 1.

Just when I thought that there was nothing left to tweak in the parameter department, hats off to you again,

Michael

I don't know what to do about botnets, but that argument can be used with any coin. As for server farms, at least there is an electrical cost per server that is not vastly different than the one per PC, so I am going to have to expand my notion of fair to include them. After all, the cost per server is at least as much as the cost per PC, so there is no FPGA/ASIC/GPU like cheat.

I get your joke, but I was actually thinking more along the lines of Faircoin, as in something anybody with a computer or tablet could generate.

Michael

It doesn't have to be 2GB per thread, it could be 256 MB per thread. That supports old computers and new CPUs will still win out over GPU threads.

Michael

GPUs will never jump on the bandwagen if your use of scrypt requires 256MB per thread. GPUs have many threads, but they're dead slow (each) compared to CPU threads, especially if the per thread memory requirements are enormous, limiting GPU parallelism.

The thing that cripples GPUs is the lack of significant memory per thread on the video card.

Have a scheme that uses scrypt and requires 2GB per thread (or even as little as 256MB per thread perhaps) and your quad-core CPU with 8 GB of RAM will beat any video card. The GPU can do many threads, but they aren't very fast. Attach large memory requirements to each thread and the GPU (as well as FPGA and even ASIC) is dead in the water.

In a nutshell it's more fair, because people are more likely to have 2-3 computers vs 2-3 video cards per computer. As for server farms and the like, I'm afraid that that would be difficult to get around with any scheme.

At the very least to even the playing field. Everybody has a computer or two in their house and chances are they are not an order of magnitude apart in performance, as is true of video cards. Also, there will be less "cheating" by pros, because the playing field will consist of only one type of player - the CPU.

Also, you can do more interesting stuff like SHA-3-1024 for example as the hash (but still using an scrypt like scheme).

... or you got those FPGA that were rumored to be coming to work  .

Basically something like litecoin, but with the following two characteristics:

1. GPUs should be no better at it than CPUs and hopefully worse.
2. It has to be absolutely FPGA and ASIC proof.

Litecoin had the right idea, but didn't go far enough. The way to really make this happen is to use something like scrypt, but have it access far more memory than is used by litecoin, by tweaking the scrypt params. That way GPUs will not be able to run a gazillion threads, because at least for the foreseeable future video cards will not have enough memory to support them.

Let me know what you guys think...

Where are the specifics? Pool cost is 1%, but what is the minimum withdrawal amount? Is there a fee to withdraw?

LOL, it probably went something like this:

Hotel attendant: Sir, you sure have many suitcases.
Lacan: Oh those... They're just my clothes for the trip.

Ten minutes later, as the hotel lights start dimming...

Lacan: Two outlets? That's all these hotel rooms have is two outlets? OK, where are my octopus splitters? Ahh, there we go... Wait, WTF? The internet is petering out and I don't know why? Could it have anything to do with the lights dimming?

Ten additional minutes pass and Lacan calls the hotel electrician.

Electrician: Sir, why are you splicing the electrical cables connected to the fan?

...

I've seen factory overclocks from 770 MHz to 786 MHz, but 822 MHz does sound high. But, hey if that's safe, maybe I'll give it a go for a while at 1.038 v and see what that nets me.

Michael

Yeah, a 6.2% overvolt is more than I am willing to stomach to push my card. I think a 3.8% overvolt to 1.038 v is about the most I am comfortable with for short gaming bursts (and that nets me the 770MHz -> 900 MHz Core Clock increase I mentioned in my previous post) on the 570.

I've very careful with the 570, because it was one of those uber-expensive Sparkle jobs with the tricked out silent fans and stuff that I paid an extra $50-$75 for. Here are some pics that may bring back memories, I'm sure you've all seen them: http://www.calibrestyle.com.tw/productDetail.asp?id=43

Michael

BTW:

Overclocking my 570 from the stock 770 MHz Core Clock to 900 MHz (a 17% increase in the Core Clock and a corresponding increase in the linked Shader Clock), which my card can do when slightly overvolted (1.0 v -> 1.038 v) raises the rate from about 230 kH/s to 266 kH/s (a 14% increase in hash rate). It does not appear that we are very memory speed bound.

I am leaving my card at stock frequency though, because I want desktop stability and do not want to overheat my card any more than necessary with the torture it is already getting from mining. 670s don't grow on trees you know .

Michael

I'm glad you ptu that edit in, I was thinking you were getting 278 on a single 560se and trumping my 570 

I should mention that I am using -i 1, because there's only one card in this host and I need to use it for other things. In any case, I am ever hopeful that whatever -C tweaks you make going forward to increase your rate won't hamper ours (so perhaps some backwards compatibility on the -C front just in case?).

Yep, exactly what I tried: Same settings for C1 as C2, both gotten from an original autotune with C0. Maybe I should try re-autotuning with C1 to see what happens. Let me do that...

Update:

Retest results for 670 card:

C1:

  Attempt 1. Did an autotune, it crashed the driver.
  Attempt 2. Autotune succeeded and tuned to the usual (i.e., the ones I get from C0) settings of 30x8, so those seem to be good settings regardless of the C switch

C2:
  
  Attempt 1: Surprisingly, no crash this time, also tuned to the usual C0 settings of 30x8 and getting same peak performance as C1 (about 231-232).

So, in the end C2==C1 if it doesn't crash the driver on startup.

I think I am going to revert back to C1 for now as my normal operating mode with a fixed -l 30x8, until you grace us with the next iteration of your wonderful mods.

Michael

P.S. A special thanks in this version for making Ctrl-C / Ctrl-Break work properly. It used to have about an 80% chance of crashing the display driver with prior versions of your app.

P.P.S An amusing unrelated note: I've got an old AMD 5450 also running cgminer 2.11.4 and generating 15 kH/s, but using only 20 (additional) watts to do it according to my kilowatt meter. My gains from this version change alone (i.e., 4/13 -> 4/17) are more than that AMD card is producing .

Here's what I am getting with an NVidia 570 with the 4/17 version compared to 4/13:

C0, a boost of about 3-4 kH (from 218 to 221/222)
C1, a boost of about 13 (from 218 to 231)
C2, crashes the display driver and fails to validate on CPU

An NVidia 670 I also have running went from 141 to 149 with C1 and this version (vs 4/13).

I'll stick with C1 and THANKS!!!  

This version + C1 covers my pool costs and then some... Hats off to you and I just sent you a token 0.1 LTC for all your hard work (and this version, specifically). I hope everyone else using your program donates at least that much to you to reimburse you for your effort.

rPzgwWFBPBGVNFBGWenALxWTduJB6DLnxK