New LLT program in the pipe..

[kriesel]

LL (with state of the art for LL error checking during the computation in the software application, which GP_LLT does not have) requires extremely reliable hardware, and that is frequently not sufficient to produce accurate results in long tests.
(Fleet average, historically, for LL error rate is ~2% for wavefront tests, ~21% for exponents at ~3x wavefront, projected ~91% for exponents at ~10x wavefront, & projected ~99% for 30x wavefront (gigadigit EFF prize territory), unless interim residues are manually compared frequently between simultaneous runs on the same exponent. Backup save files would need to be saved for retreat to the point before an error was detected by comparing runs in progress. See 4th attachment of https://www.mersenneforum.org/showpo...6&postcount=14)

PRP (with state of the art for PRP error checking during the computation in the software application, which prime95, mlucas, and gpuowl DO have) does not require particularly reliable hardware, and usually does produce accurate results in long primality tests, even on unreliable hardware. It's common to run prp/gec on hardware too unreliable for other use in GIMPS.
On extremely unreliable hardware, in PRP/GEC, errors detected may result in little or no progress on iteration count, and that's a good thing, because such unreliable hardware if unchecked would certainly produce wrong results running either unguarded PRP or LL. I've seen that work, in prime95, on an old AMD Athlon system I picked up at a thrift store for $15; its reliability deteriorated to the point it made no further progress in PRP/GEC because errors were so frequent.

Most of the fleet does not contain ECC ram, so relying on ECC for accuracy of results is a design error.
ECC does not prevent the erroneous effect of software bugs or occasionally reaching values in the fft terms that may result in roundoff error or a cosmic ray hit somewhere other than in ram or hardware error due to overclocking/overheating/age either. ECC will ensure that an erroneously computed value is stored and retrieved from ram without additional modification by memory error. That's all.

[Uncwilly]

Originally Posted by kriesel

Insisting on staying with LL for production testing will prevent acceptance of the program by end users.

Originally Posted by kriesel

That's all.

Further application of thongs and switches is unlikely to levitate the equine. It is discouraged as much as FTC-LL's.

[kriesel]

Thongs?!

I took the GP_LLT program for a multiday test. Observations, reactions and results here.

[Uncwilly]

Originally Posted by kriesel

Thongs?!

#1 https://en.wiktionary.org/wiki/thong#Noun

[mrh]

Originally Posted by Uncwilly

#1 https://en.wiktionary.org/wiki/thong#Noun

Back in my day, it always meant footwear, and then I don't know what happened...