Did anyone watch madmax code for the access-pattern his code reads / writes Data from/to disk? Somewhere i read the original plotter would read / write in 26kb-blocks
How do you feel about shortening the lifespans/killing the nvme’s at least ramdisk is going to last for the forseeable.
That would be a good move. Have the option for those with fewer disks and more processor power/ram to plot sequentially and those with more disks and less processor power to plot in parallel.
Having said that wasn’t one of the aims of Chia to no be CPU intensive, to discourage the GPU, ASICs etc…
1 Like
tmpdir_stagger_phase_major: 1
tmpdir_stagger_phase_minor: 5
# Optional: default is 1
tmpdir_stagger_phase_limit: 1
# Don't run more than this many jobs at a time on a single temp dir.
tmpdir_max_jobs: 4
# Don't run more than this many jobs at a time in total.
global_max_jobs: 16
# Don't run any jobs (across all temp dirs) more often than this, in minutes.
global_stagger_m: 34
# How often the daemon wakes to consider starting a new plot job, in seconds.
polling_time_s: 5
the rest is on default, 3 or 4 threads per job, and 4 1Tb 980 pro plotting drives.
2 Likes
On an i9-9900k, 3200 GHz DDR4, Samsung Evo Plus 2tb, 86 minutes.
1 Like
So finally Chia is getting greener.
2 Likes
i9-10900 2.80ghz 56 minutes
mention by Gene
then Bram Cohen:
“We can’t merge in the madmax plotter under the current license of GPLv3. If it’s released Apache 2 we’d be very interested in merging it as the default, with the caveat that we need to try a bunch of different platforms/hardware/circumstances and see what works best on each.”
3 Likes
Interesting, it looks like Gen 4 M2 makes quite a difference over the processor speed/threads.
With the traditional plotter, I was definitely losing speed on P3 with the CPU setup, but not as much with MadMax.
Ryzen 5 3600 3.6 GHz (6C/12T)
64 MB 3400 GHz DDR4
Firecuda 2TB Gen 4 (single stick for both tmp drives)
53 Minutes
1 Like
Gen 4 would be much better. My rigs are all Gen 3, now too late to buy Gen 4 motherboard, etc. I’m nearly done plotting.
1 Like
I’m getting 29 min plots with 5950x if I have the node/farmer running, my tmp drive is 3 2tb SSD in raid0. About 25 min without the node running. Seems like a lot of penalty!
If you win a block in the meantime, not so much a penalty, you never know. Why would you waste such a machine as farmer?
Every machine that i have is plotting (5950, 5900, 5800, 3600), all the others are harvesting and this one is farming too. The amount of plots per day I’ll lose if i move the farmer to another machine will be about the same, but honestly, it seems like a lot. It’s not using CPU to warrant this few minute time increase. I’m going to try again without the node to see if i actually go down to 25 min or if it could be something else.
1 Like
Now with Madmax in mind, the best NVMe to buy is a 500+ GB with a good reputation, a motherboard with 8 RAM plots, and some 32GB DDR4 sticks. (The budget friendly build.)
2 Likes
Well, not quite. You see, SSDs are essentially a raid 0 of NAND chips. The fastest SSDs are also the biggest of them.
4 Likes
well yes in theory, but not in real-life, especially for consumer models. If you look at the reviews from Toms for example, you will see that sustained speeds for 1 or 2 TB models are basically the same for many models. Just too many other factors at play.
2 Likes
If not speed, overprovisioning which is directly tied to expectancy of TBW is tied to the size. 500TB drives have literally 1/4 the TBW of 2TB drives.
I have been getting around 2000 sec plot times on a 5800X using 2 Gen4 2TB nvmes.
My dual 2680v1 rig is doing about 2600 second per plot. In order to make that though I needed to raid 2 nvme’s for temp2.
I have a feeling the 5800X is pretty close to optimum for cost / performance ratio on new kit.
And I’m thinking the dual xeon I have is probably very near the mark for used gear.
