Time per gigatype RAM with NN engines to exhaustion

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

jjoshua2
Posts: 99
Joined: Sat Mar 10, 2018 6:16 am

Re: Time per gigatype RAM with NN engines to exhaustion

Post by jjoshua2 »

If you view the SSD as a consumable then you don't mind if it wears out in 4-6 years, as it will be worthless anyway even if it still works. I just looked up a sata 840 evo 250gb was $200 in 2013. Now we can get a good 1tb nvme SSD for under $100. In 5 years can probably get a 8tb SSD for $200 thats much faster again that can really utilize pcie4/5. The main thing to avoid is having it break the month after warranty expires ;)
ernest
Posts: 2041
Joined: Wed Mar 08, 2006 8:30 pm

Re: Time per gigatype RAM with NN engines to exhaustion

Post by ernest »

Dann Corbit wrote: Sat Dec 07, 2019 7:15 pm
Most ssd devices have about the same endurance as a regular mechanical disk.
Very curious statement, to say the least ! Basis ?

From what I have read, ssd failure after too many writes is intrinsic to ssd...
corres
Posts: 3657
Joined: Wed Nov 18, 2015 11:41 am
Location: hungary

Re: Time per gigatype RAM with NN engines to exhaustion

Post by corres »

ernest wrote: Sun Dec 08, 2019 4:32 am
Dann Corbit wrote: Sat Dec 07, 2019 7:15 pm Most ssd devices have about the same endurance as a regular mechanical disk.
Very curious statement, to say the least ! Basis ?
From what I have read, ssd failure after too many writes is intrinsic to ssd...
Every SSD have a self-repair tool. This tool marks the bad cells and the control of SSD can not use them. The consequence is the effective capacity of SSD will decrease in time more faster than an HDD..
The endurance of SSD depends on the type of controlling system and structure of memory cells. The cheaper SSD has smaller endurance but its price compensates this issue. Undoubtedly an SSD wants more checks than an HDD.
dragontamer5788
Posts: 201
Joined: Thu Jun 06, 2019 8:05 pm
Full name: Percival Tiglao

Re: Time per gigatype RAM with NN engines to exhaustion

Post by dragontamer5788 »

jp wrote: Sat Dec 07, 2019 12:10 pm
dragontamer5788 wrote: Fri Dec 06, 2019 10:26 pm This sort of traversal would minimize the hits to the hard drive, while still providing a MCTS traversal of some kind. SSD Flash storage might be fast enough to provide the ~50,000 positions/second that LeelaZero can provide, but hard drives are just way too slow
But I read that SSDs will be harmed by that sort of activity. Right?
Reads don't harm SSDs, only writes.

A typical modern, TLC SSD cell wears out after ~1000 writes or so (depending on technology), but there's very good load-balancing algorithms. With proper load-balancing, the write endurance grows to ~500TBs of writes (for 1TB drives). The bigger your SSD drive, the more load-balancing and write-endurance you get. (2TB SSDs should have over 1000 TBs of rated write endurance).

A hard drive writing 4k sectors at 500-operations/second is roughly 2MB/s write speed. Random-IOPS are the worst for hard drives... it would take 7-years for a hard-drive to random-IO write 500TB of data. Of course, hard drives could get there through sequential-write (~200MBs write speed), but sequential writes are harder to program in practice.

SSDs have 100,000 operations/second, so yes a TLC SSD could reach 500TBs of writes in just 14-days. Hmmm... I guess keeping an eye on S.M.A.R.T data and seeing how many total writes go to the SSD will be an important metric to keep track of. I "feel" like MCTS would be mostly reads, but... it'd be worthwhile to investigate the number of writes of these schemes.