Help to choose components for a Ryzen rig

[CRGreathouse]

Originally Posted by LaurV

On the SSD vs HDD debate, we would never buy a cheap HDD (in our life we succeeded to destroy at least 14 HDDs), but we also see no utility on buying a SSD.

A 480 GB SSD is only $100, I don't see why anyone would bother with spinning rust except for cold storage of files (photos, videos, etc.).

[danaj]

Re SSD vs. HDD, thanks for the analysis.

For this purpose, if you can get a 120ish GB SSD cheap, or have it as a hand-me-down from another computer or person who got a bigger drive, go for it. But otherwise an HDD would be just fine. There's no particular need in this application for one.

I love SSDs for Windows machines, especially gaming ones. Unlike my Linux machines, all our Windows boxes constantly putz with the disk. Especially if it is in a bedroom, this will drive you crazy. At least you get a bunch of audible indication that it is even busier when you start a program -- it is diligently flogging the disk like a BDSM mistress. Use an SSD and get joyous silence, not to mention notably better performance. But this isn't a Windows machine.

For graphics, I'm not familiar with AMD, but for these kind of boxes I use the built-in graphics or get a cheap card (e.g. $10-$25 US, passively cooled, low idle power). I don't like having it completely headless even if it spends most of its time with a monitor turned off or even taken somewhere else. At least I know if I need to I can attach a monitor and open xterms as needed.

For memory, 8GB should work though I would do 16GB just because I like to run some other things occasionally that take more memory. Memory bandwidth seems to be important when running lots of threads, so higher speed seems like a good idea.

Using a Kill-a-watt unit years ago it's surprising how little power is used even when the CPU is going full speed. My overclocked 3930K used a lot more than most of my machines, at over 200W. I like Seasonic PSUs for the quiet, the build quality, and the efficiency, but they do charge for it. I think 350W is adequate. I agree looking at high efficiency e.g. 80+ Gold (or more) is worthwhile. Unless you have a need for a space heater I suppose.

[danaj]

Way back on one of the threads about the prime gap finding program, I noted that it was a bit lazy in that it does a parallel search of a given block that could be unbalanced, leading to some threads finishing earlier than others but sitting idle waiting for all threads in that block to finish. In practice it doesn't normally seem like a big deal, but I also run large blocks (days to finish one). A different design could get around the issue.

smallgaptest with merits-20180404.txt.

i7-6700K @ 4.2GHz

time threads
362.5s 1
181.5s 2 (2x)
121.1s 3 (2.99x)
92.7s 4 (3.91x)
89.3s 5 (4.06x)
87.1s 6 (4.16x)
81.2s 7 (4.46x)
76.1s 8 (4.76x)

i7-4770K @ 4.3GHz

time threads
415.8s 1
212.6s 2 (1.96x)
142.5s 3 (2.92x)
107.0s 4 (3.89x)
119.8s 5 (3.47x) [yes, this is slower]
99.6s 6 (4.17x)
95.8s 7 (4.34x)
90.4s 8 (4.60x)

Looks like scaling through 4 threads works pretty well. 8 threads is faster but not much gain for the effort from hyperthreading. It's still faster in absolute terms.

With 8 threads I also ran with an empty merits file, so the test prints out gaps it found (that are not close to being records). It seems to run just slightly faster (75.6s and 90.0s).

[robert44444uk]

Originally Posted by danaj

Way back on one of the threads about the prime gap finding program, I noted that it was a bit lazy in that it does a parallel search of a given block that could be unbalanced, leading to some threads finishing earlier than others but sitting idle waiting for all threads in that block to finish. In practice it doesn't normally seem like a big deal, but I also run large blocks (days to finish one). A different design could get around the issue.

smallgaptest with merits-20180404.txt.

i7-6700K @ 4.2GHz

time threads
362.5s 1
181.5s 2 (2x)
121.1s 3 (2.99x)
92.7s 4 (3.91x)
89.3s 5 (4.06x)
87.1s 6 (4.16x)
81.2s 7 (4.46x)
76.1s 8 (4.76x)

i7-4770K @ 4.3GHz

time threads
415.8s 1
212.6s 2 (1.96x)
142.5s 3 (2.92x)
107.0s 4 (3.89x)
119.8s 5 (3.47x) [yes, this is slower]
99.6s 6 (4.17x)
95.8s 7 (4.34x)
90.4s 8 (4.60x)

Looks like scaling through 4 threads works pretty well. 8 threads is faster but not much gain for the effort from hyperthreading. It's still faster in absolute terms.

With 8 threads I also ran with an empty merits file, so the test prints out gaps it found (that are not close to being records). It seems to run just slightly faster (75.6s and 90.0s).

Hi Danaj. Thank you for running the test. Is it possible to redesign the gap finding program to use the threads really efficiently? FYI I also carried out my run with an empty merits.txt file.

[mackerel]

To me storage is like any other part, you pick the right one for the intended use case. I'd disagree with LaurV's viewpoint, but my use case may well be very different leading to a different optimisation.

For basically any modern Windows system I can't imagine seriously using it without a SSD of any kind. Even cheap ones. Even small ones. I'm running one system off an ancient Intel 40GB SSD which is still going strong after, I don't know how many years, and it is far more responsive than any HD based system. I've even played with 15k drives, they're not even close to cheap SSD performance. A cheap SSD will still destroy a top end HD in performance.

Unless you know you are continuously writing a lot to the drives, endurance is largely a non-issue. A tech site once tested various drives to destruction, and found their actual endurance was far above the manufacturer's rated life. For consumers, it is not a concern.

Are SSDs 100% reliable? No, but not electronic devices are. My sample size will not be significant, but I own SSDs numbering into the double digits, and over the years have had two sudden death failures. These aren't endurance, but some controller or other failure. One unbranded 60GB, one Sandisk 480GB. It happens. Over the years I've had hundreds of HDs, and a fair bunch of them failed too, the most common scenario being growing bad sectors over time.

Right now I still use a mix of HDs and SSDs, with the HDs more for bulk storage where performance isn't a concern and the increased cost per capacity of a SSD is not justified.

[henryzz]

SSHDs are often a good compromise IMO.

[Mark Rose]

Originally Posted by Mark Rose

The GT 710 is based on the Kepler architecture from 2012.

I messed up. It's some GT 730 that are based on Fermi, not the 710. Still, the Kepler architecture of the 710 won't be around forever. Fermi support is being dropped in January.

https://www.anandtech.com/show/12624...bit-os-support