#1 - Page 3

axn · 2019-12-14, 13:43

I hope you give a good look towards AMD Zen 2 platform as they provide the most core / $. Should be very good for NFS PP.

masser · 2019-12-15, 17:18

Here is an alternate build with a 12 core Zen 2 cpu and 64 GB of FAST ram (upgradable to 128 GB).

Some of the 16 core options from the Zen+ line seem more cost-efficient. However, if there's a big jump in performance between Zen+ and Zen 2, the comparison might be a wash.

Do the factoring tasks benefit from fast ram the way prime95 LL testing does?

VBCurtis · 2019-12-15, 18:37

Ram bandwidth does matter for matrix solving, but not for sieving nor poly select.
If Zen 2 had 4 channel ram, I think I'd buy one next rather than another used Xeon.

lycorn · 2019-12-15, 19:16

I reckon zen 2 builds are nice to ECM on small exponents due to the relatively large caches, where the FFTs are supposed to fit. But for LL testing on large exponents the 4 channels will certainly make a difference.
I am thinking of getting a 3950x next year and then I`ll post some results.

Xyzzy · 2019-12-15, 23:57

https://www.amazon.com/dp/B074CBJHCT

$200

Good deal?

Xyzzy · 2019-12-15, 23:59

Will too many cores saturate the memory bandwidth?

That CPU we just posted is a 12 core 24 thread part.

Xyzzy · 2019-12-16, 00:09

Also, should we bother with hyperthreading?

Our guess is it might be better to prioritize real cores.

We doubt a consumer-level setup can support more than eight real cores, but we could be wrong.

Example:

8 cores/8 threads might cost the same as 6 cores/12 threads.

CRGreathouse · 2019-12-16, 07:23

Quote:

Originally Posted by Xyzzy

Also, should we bother with hyperthreading?

No. I'd take 8 cores over 6 cores with 12 threads any day (ceteris paribus).

axn · 2019-12-16, 09:42

Quote:

Originally Posted by CRGreathouse

No. I'd take 8 cores over 6 cores with 12 threads any day (ceteris paribus).

AMD SMT is slightly more capable than intel HT, so depending on which way you go, it might be useful.

/IIUC

masser · 2019-12-17, 19:11

Quote:

Originally Posted by Xyzzy

https://www.amazon.com/dp/B074CBJHCT

$200

Good deal?

The 1920X at $200 appears to be a very good deal. Here's a build based off of that chip, with 64GB ram upgradeable to 128 GB. Note that the 1920X is from the original Zen lineup, so the ipc won't be as good as Zen 2. It supports quad channel ram, but only up to 2666.

On the power supply front, if you are going to be running a machine near 100% utilization fulltime for three or more years, platinum PSUs become worth the upfront cost, IMO.

HTH

kriesel · 2019-12-17, 20:42

Quote:

Originally Posted by masser

The 1920X at $200 appears to be a very good deal. Here's a build based off of that chip, with 64GB ram upgradeable to 128 GB. Note that the 1920X is from the original Zen lineup, so the ipc won't be as good as Zen 2. It supports quad channel ram, but only up to 2666.

On the power supply front, if you are going to be running a machine near 100% utilization fulltime for three or more years, platinum PSUs become worth the upfront cost, IMO.

HTH

And oversize the PSU. Max efficiency is near 50% of max output.

2019-12-14, 13:43	#23
axn Jun 2003 5464₁₀ Posts	I hope you give a good look towards AMD Zen 2 platform as they provide the most core / $. Should be very good for NFS PP.

2019-12-15, 17:18	#24
masser Jul 2003 Behind BB 2·7·11·13 Posts	Here is an alternate build with a 12 core Zen 2 cpu and 64 GB of FAST ram (upgradable to 128 GB). Some of the 16 core options from the Zen+ line seem more cost-efficient. However, if there's a big jump in performance between Zen+ and Zen 2, the comparison might be a wash. Do the factoring tasks benefit from fast ram the way prime95 LL testing does?

2019-12-15, 18:37	#25
VBCurtis "Curtis" Feb 2005 Riverside, CA 2×2,927 Posts	Ram bandwidth does matter for matrix solving, but not for sieving nor poly select. If Zen 2 had 4 channel ram, I think I'd buy one next rather than another used Xeon.

2019-12-15, 19:16	#26
lycorn "GIMFS" Sep 2002 Oeiras, Portugal 2×7×113 Posts	I reckon zen 2 builds are nice to ECM on small exponents due to the relatively large caches, where the FFTs are supposed to fit. But for LL testing on large exponents the 4 channels will certainly make a difference. I am thinking of getting a 3950x next year and then I`ll post some results.

2019-12-15, 23:57	#27
Xyzzy Aug 2002 21D3₁₆ Posts	https://www.amazon.com/dp/B074CBJHCT $200 Good deal?

2019-12-15, 23:59	#28
Xyzzy Aug 2002 7·1,237 Posts	Will too many cores saturate the memory bandwidth? That CPU we just posted is a 12 core 24 thread part.

2019-12-16, 00:09	#29
Xyzzy Aug 2002 7×1,237 Posts	Also, should we bother with hyperthreading? Our guess is it might be better to prioritize real cores. We doubt a consumer-level setup can support more than eight real cores, but we could be wrong. Example: 8 cores/8 threads might cost the same as 6 cores/12 threads.