Time per gigatype RAM with NN engines to exhaustion

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Dann Corbit
Posts: 10187
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Time per gigatype RAM with NN engines to exhaustion

Post by Dann Corbit » Wed Dec 04, 2019 10:52 pm

With an RTX 2080, how fast does it consume RAM?
I have found that my 1080TI eats 32 GB in an hour, and the analysis slows to a crawl (and the machine starts to get balky).
If you have 2x RTX2080, what is the maximum time you can analyze before RAM is exhausted with 64 GB and with 128 GB?
Has anyone looked into this?
I think that there is a way to limit maximum RAM, but if you do that, how is the analysis impacted?
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

dragontamer5788
Posts: 160
Joined: Thu Jun 06, 2019 6:05 pm
Full name: Percival Tiglao

Re: Time per gigatype RAM with NN engines to exhaustion

Post by dragontamer5788 » Wed Dec 04, 2019 10:55 pm

Dann Corbit wrote:
Wed Dec 04, 2019 10:52 pm
With an RTX 2080, how fast does it consume RAM?
I have found that my 1080TI eats 32 GB in an hour, and the analysis slows to a crawl (and the machine starts to get balky).
If you have 2x RTX2080, what is the maximum time you can analyze before RAM is exhausted with 64 GB and with 128 GB?
Has anyone looked into this?
I think that there is a way to limit maximum RAM, but if you do that, how is the analysis impacted?
Isn't this just a simple "nodes per second" x "bytes per node" calculation?

I don't know what the NPS of a RTX 2080 or 2x 2080 is. But that's probably the only calculation you'd need to run.

zullil
Posts: 5680
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: Time per gigatype RAM with NN engines to exhaustion

Post by zullil » Wed Dec 04, 2019 11:36 pm

Dann Corbit wrote:
Wed Dec 04, 2019 10:52 pm
With an RTX 2080, how fast does it consume RAM?
I have found that my 1080TI eats 32 GB in an hour, and the analysis slows to a crawl (and the machine starts to get balky).
If you have 2x RTX2080, what is the maximum time you can analyze before RAM is exhausted with 64 GB and with 128 GB?
Has anyone looked into this?
I think that there is a way to limit maximum RAM, but if you do that, how is the analysis impacted?
With that hardware, knowing the type of analyses you do, you should plan on a minimum of 512 GB of RAM.

dragontamer5788
Posts: 160
Joined: Thu Jun 06, 2019 6:05 pm
Full name: Percival Tiglao

Re: Time per gigatype RAM with NN engines to exhaustion

Post by dragontamer5788 » Wed Dec 04, 2019 11:50 pm

zullil wrote:
Wed Dec 04, 2019 11:36 pm
Dann Corbit wrote:
Wed Dec 04, 2019 10:52 pm
With an RTX 2080, how fast does it consume RAM?
I have found that my 1080TI eats 32 GB in an hour, and the analysis slows to a crawl (and the machine starts to get balky).
If you have 2x RTX2080, what is the maximum time you can analyze before RAM is exhausted with 64 GB and with 128 GB?
Has anyone looked into this?
I think that there is a way to limit maximum RAM, but if you do that, how is the analysis impacted?
With that hardware, knowing the type of analyses you do, you should plan on a minimum of 512 GB of RAM.
Is RAM necessary? Shouldn't the search tree be effectively cached on a NVMe Optane drive or NVMe Flash drive?

The strict memory requirements of MCTS are probably worth investigating. I'd assume that most "visits" down an MCTS, even with virtual loss, would result in the "same section of the tree" being visited over and over again. So 128GB of RAM, backed by 1024GB of NVMe Flash (operating on an mmap'd section of code or something) probably would work out fine.

At a minimum, the root of the MCTS tree would be visted over-and-over again (and therefore be cached in RAM). But most of the nodes in a 512GBs MCTS tree won't be revisited.

Stephen Ham
Posts: 2428
Joined: Wed Mar 08, 2006 8:40 pm
Location: Eden Prairie, Minnesota
Full name: Stephen Ham

Re: Time per gigatype RAM with NN engines to exhaustion

Post by Stephen Ham » Wed Dec 04, 2019 11:58 pm

Dann Corbit wrote:
Wed Dec 04, 2019 10:52 pm
With an RTX 2080, how fast does it consume RAM?
I have found that my 1080TI eats 32 GB in an hour, and the analysis slows to a crawl (and the machine starts to get balky).
If you have 2x RTX2080, what is the maximum time you can analyze before RAM is exhausted with 64 GB and with 128 GB?
Has anyone looked into this?
I think that there is a way to limit maximum RAM, but if you do that, how is the analysis impacted?
Hi Dann,

A NN engine on my RTX 280 Ti eats 32GB in about 30 minutes. After an hour, most of my RAM is consumed. The PC becomes sluggish and my mouse sometimes then disappears. Switching to a new position doesn't always free up the consumed RAM. So, I have to exit my NN engine and then return to get sufficient RAM back

Another mystery is that GPU usage sometimes drops. Normally it stays around 98-99%, but over time it can drop to near 0%.

All the best,
-Steve-

corres
Posts: 1758
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Time per gigatype RAM with NN engines to exhaustion

Post by corres » Thu Dec 05, 2019 12:00 am

NN engines do not need fast working memory/page file.
So if you set up a big (50-100 GB or more GB) page file on a HDD it may be appropriate.
Obviously if you use SSD instead of HDD it will be better.

Dann Corbit
Posts: 10187
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: Time per gigatype RAM with NN engines to exhaustion

Post by Dann Corbit » Thu Dec 05, 2019 12:11 am

I already have one, but I might buy another one of these:
https://www.newegg.com/gigabyte-aorus-2 ... 6820009013
I saw that the motherboard can handle it.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

corres
Posts: 1758
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Time per gigatype RAM with NN engines to exhaustion

Post by corres » Thu Dec 05, 2019 7:17 am

Dann Corbit wrote:
Thu Dec 05, 2019 12:11 am
I already have one, but I might buy another one of these:
https://www.newegg.com/gigabyte-aorus-2 ... 6820009013
I saw that the motherboard can handle it.
If you would buy a common SSD and connect to a SATA driver it would be good too.

Dann Corbit
Posts: 10187
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: Time per gigatype RAM with NN engines to exhaustion

Post by Dann Corbit » Thu Dec 05, 2019 7:23 am

I bought a 2 TB PCIE4 SSD for my swap space. It does 5 gb/sec read and 4.4 gb/sec write.
With extended guarantee, I paid 500
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

schack
Posts: 140
Joined: Thu May 27, 2010 1:32 am
Contact:

Re: Time per gigatype RAM with NN engines to exhaustion

Post by schack » Fri Dec 06, 2019 5:01 am

I was under the impression that NN engines didn't need nearly as much traditional hash space as AB engines. Is that not correct?

Post Reply