mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware

Reply
 
Thread Tools
Old 2021-03-05, 20:36   #166
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

3×5×373 Posts
Default

prime95 on Windows 10 gave mostly 2 cores/worker optimal, until reaching large exponents (~300M). But latency can be long compared to assignment expiration. See second attachment of https://www.mersenneforum.org/showpo...4&postcount=11 for prime95/win10 exhaustive benchmark, for comparison to any ubuntu/mprime results to follow.

Last fiddled with by kriesel on 2021-03-05 at 20:36
kriesel is online now   Reply With Quote
Old 2021-03-05, 21:27   #167
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,657 Posts
Default

Quote:
Originally Posted by kriesel View Post
prime95 on Windows 10 gave mostly 2 cores/worker optimal, until reaching large exponents (~300M). But latency can be long compared to assignment expiration. See second attachment of https://www.mersenneforum.org/showpo...4&postcount=11 for prime95/win10 exhaustive benchmark, for comparison to any ubuntu/mprime results to follow.
Thanks, though as usual I needed reading glasses. :)

My own rule of thumb for manycore-system runs is to go with whichever combo has the fewest #workers and yields at least 95% of the maximum total throughput, to keep completion times reasonable.
ewmayer is offline   Reply With Quote
Old 2021-05-01, 21:43   #168
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,657 Posts
Default

Of possible interest to our UK readers, a cheap refurb 68C/272T 7250 KNL.

Priced some RAM modules, going to hold off for now, but do occasional price-checks, posting links here for reference.

From the SuperMicro mobo manual:
Quote:
The K1SPE has six (6) DIMM slots that can support 384 GB of Registered RDIMM up to 64 GB size and DDR4 2400 MHz speed, 1.20V SDRAM. See Chapter 5 for details.
(BIOS applicable settings related to memory: Memory Configurations -> Memory Map -> Channel Interleaving)
Recommended RAM modules: DDR4 1.2V-2400 is the fastest supported RAM frequency:
Code:
DDR4 1.2V-2133:
GB-Per-DIMM	Recommended Modules
 8:
	Samsung M393A1G40DB0-CPB	$50 per dimm
	Hynix HMA41GR7AFR8N		$44 on Amzn
16:
	Hynix HMA42GR7AFR4N-TF		$56 on Amzn
32:
	Samsung M393A4K40BB0-CPB	$139 on Amzn
	Hynix HMA84GR7MFR4N-TF		$139 on Amzn

DDR4 1.2V-2400:
GB-Per-DIMM	Recommended Modules
 8:
	Samsung M393A1G40EB1-CRC	$50 per dimm
	Hynix HMA41GR7AFR4N-UH		$40 on Amzn
16:
	Samsung M393A2K40BB1-CRC	$73 on Amzn
	Samsung M393A2K40CB1-CRC	$80 on Amzn
	Hynix HMA42GR7AFR4N-UH		$65 on Amzn
	Hynix HMA82GR7AFR8N-UH		$85 on Amzn, but as low as $58 used [MicroMark had best price]
32:
	Samsung M393A4K40BB1-CRC	$157 on Amzn
	Samsung M393A4K40CB1-CRC	$172 on Amzn; this was the 1 module where
					the Sponsored-Ad top-placer in my DDG searches,
					Nemix.com, had the best price, $150
	Hynix HMA84GR7AFR4N-UH		$157 on Amzn
64:
	Samsung M386A8K40BM1-CRC4Q	$297 on Amzn
	Micron MTA72ASS8G72LZ-2G3B2	$298 on eBay, $350 on Amzn

Last fiddled with by ewmayer on 2021-05-22 at 00:36
ewmayer is offline   Reply With Quote
Old 2021-05-01, 23:34   #169
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

22×37×67 Posts
Default

Quote:
Originally Posted by ewmayer View Post
Of possible interest to our UK readers, a cheap refurb 68C/272T 7250 KNL:
That sold days ago.

Last fiddled with by ewmayer on 2021-05-02 at 03:06 Reason: Still sets a price level...
Uncwilly is offline   Reply With Quote
Old 2021-05-07, 10:53   #170
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

3·5·373 Posts
Default

Quote:
Originally Posted by ewmayer View Post
Of possible interest to our UK readers, a cheap refurb 68C/272T 7250 KNL.
...
Recommended RAM modules: focus on DDR4 1.2V-2400, the fastest supported RAM frequency:
That 7250 was the chip package only, requiring a compatible motherboard etc., so a possible upgrade path for a 7210 system. Such processors are quite commonly available on eBay etc.

It might be better to go with 2133 RAM. From https://www.mersenneforum.org/showpo...72&postcount=8 (bold emphasis mine):
Code:
I found the following in the specification update:
"KNL25.Operating With DDR4-2400 Memory May Cause Unpredictable System Behavior
Problem:Operating the processor with DDR4 memory configured to operate at 2400 MT/s may cause unpredictable system behavior.
Implication:When the erratum occurs, the system will exhibit unpredictable system behavior.
Workaround:It is possible for BIOS to contain a workaround for this erratum.
Status:No Fix"
Latest BIOS I found was 2.0a and does not address the preceding issue.

Last fiddled with by kriesel on 2021-05-07 at 11:06
kriesel is online now   Reply With Quote
Old 2021-05-07, 10:57   #171
paulunderwood
 
paulunderwood's Avatar
 
Sep 2002
Database er0rr

380210 Posts
Default

Quote:
Originally Posted by kriesel View Post
... so a possible upgrade path for a 7210 system.
Scary number of pads on the "chip" -- would have to be carefully accurate putting into the socket.

Last fiddled with by paulunderwood on 2021-05-07 at 10:59
paulunderwood is offline   Reply With Quote
Old 2021-05-07, 15:03   #172
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

3·5·373 Posts
Default

From https://www.supermicro.com/manuals/s.../MNL-1891.pdf:
"These units are sold as complete systems with the motherboard, processor, and
heat sink assembly factory installed. These components are not user serviceable,
and the user should not attempt to remove or modify these components other than
adjustments specifically outlined in this manual. Making other adjustments to these
components risks damaging the system and voiding the system warranty."
There's a liquid cooling loop block mounted atop the package by two alignment pins,
4 Torx fasteners, and distinct tightening and loosening orders given on the block label.
Like torquing down a multicylinder engine head.
I've decided not to mess with mine.
Attached Files
File Type: zip xeon-phi-7250-cpu-cooler GEDC0007.zip (2.83 MB, 35 views)

Last fiddled with by kriesel on 2021-05-07 at 15:04
kriesel is online now   Reply With Quote
Old 2021-05-07, 20:37   #173
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

1165710 Posts
Default

Quote:
Originally Posted by kriesel View Post
It might be better to go with 2133 RAM.
I updated the current-prices table in my post above to add the officially supported 2133 DIMMs - as with 2400, the current sweet spot is at 16GB, one could buy 6 of those and thus max out the memory bandwidth to the resulting 96GB for around $360.
ewmayer is offline   Reply With Quote
Old 2021-05-25, 22:40   #174
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,657 Posts
Default

Ordered 192GB = 6 x 32GB-per-dimm of RAM for my KNL last night - the 2400MHz actually proved slightly cheaper (I walked both 2133 and 2400 through to checkout in separate-site shopping baskets), as the seller didn't charge me CA sales tax, saving around $80. Guess I'll find out if the "Unpredictable System Behavior" issue Kriesel mentioned is rare or not.

I ended up going with 192GB as opposed to the slightly-cheaper-per-GB 96GB option, because I need to do shakedown testing of the Mlucas v20 p-1 code, and 192GB should allow me to run 17 simultaneous p-1 jobs, each on 4 distinct cores, with a maxalloc setting of 12GB per job, which will yield 60-70% of stage 2 primes paired. (The min-RAM option, which needs ~1.5GB RAM per p-1 instance for current-wavefront exponents, yields ~25% primes-paired.) Or, assuming a random mix of progress among the 17 jobs and each spending roughly the same amount of time in stages 1 and 2, I might be able to use maxalloc of around 20GB per job, which would cut an additional 10% off the stage 2 runtime. We shall see.
ewmayer is offline   Reply With Quote
Old 2021-06-01, 15:59   #175
diep
 
diep's Avatar
 
Sep 2006
The Netherlands

13538 Posts
Default

hello i see kriesel benchmarked woltmann code on it. How fast does it work for say much smaller prime testing if we'd try to use it for riesel testing? Which is what is it some millions of digits max.

How fast is it versus LLR with latest woltman inside on a Hashwell or broadwell?

On ebay i see the 2699ES getting offered. Can be put in a 2 socket system. Will be 2.1Ghz clocked so that's:

2 cpu's * 2.1Ghz * 22 cores * 32 flops (hashwell/broadwell) = 2956.8 gflops = 3 Tflops and of course each cpu has its own memory channels to it.

Roughly 1500 euro to asemble system for so i counted for motherboard+cpu's+ram

A 7290 system times 2 cpu's would be more expensive than this and intel will not like to support it anymore and i do not know about development tools for these cpu's. Usually the intel c++ compiler and such is paid when i toy at home as i have a company (which is not for profit busy in prime numbers - commerciel for robots/drones and especially 3d printers - as all this here is my hobby obviously).

So what matters is how fast Woltman is with his library at the knights landing.

Any benchmarks done there for non-mersenne primes?
diep is offline   Reply With Quote
Old 2021-06-01, 20:58   #176
paulunderwood
 
paulunderwood's Avatar
 
Sep 2002
Database er0rr

2·1,901 Posts
Default

Quote:
Originally Posted by diep View Post

Any benchmarks done there for non-mersenne primes?
3*2^6000000-1 with 64 instances (optimal) is 4.15 ms/iter. About 7 hours per test. (Xeon Phi 7210) YMMV.
paulunderwood is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
AMD vs Intel dtripp Software 3 2013-02-19 20:20
Intel NUC nucleon Hardware 2 2012-05-10 23:53
Intel RNG API? R.D. Silverman Programming 19 2011-09-17 01:43
AMD or Intel mack Information & Answers 7 2009-09-13 01:48
Intel Mac? penguain NFSNET Discussion 0 2006-06-12 01:31

All times are UTC. The time now is 11:19.


Sat Sep 18 11:19:03 UTC 2021 up 57 days, 5:48, 0 users, load averages: 0.83, 1.18, 1.29

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.