losing performance on memory bottleneck?

[kracker]

Got a new (very heavily memory-bottlenecked) system, with a core i5-8400 with dual channel DDR4 2666 MHz ram... I know after I hit the bottleneck I won't be getting much more... but is it normal to actually lose iter/s? Little confused..

Prime95 64-bit version 29.4, RdtscTiming=1
Timings for 4096K FFT length (1 core, 1 worker): 15.10 ms.  Throughput: 66.21 iter/sec.
Timings for 4096K FFT length (2 cores, 1 worker):  8.01 ms.  Throughput: 124.81 iter/sec.
Timings for 4096K FFT length (3 cores, 1 worker):  5.87 ms.  Throughput: 170.40 iter/sec.
Timings for 4096K FFT length (4 cores, 1 worker):  5.17 ms.  Throughput: 193.28 iter/sec.
Timings for 4096K FFT length (5 cores, 1 worker):  5.19 ms.  Throughput: 192.50 iter/sec.
Timings for 4096K FFT length (6 cores, 1 worker):  5.33 ms.  Throughput: 187.52 iter/sec.
Timings for 4096K FFT length (6 cores, 6 workers): 33.34, 34.10, 33.87, 34.15, 33.98, 33.88 ms.  Throughput: 177.08 iter/sec.

[Mark Rose]

It can be. Having more active cores can lower the max clock of the cores to keep the cpu in its thermal budget.

[NookieN]

Yeah, this (https://en.wikichip.org/wiki/intel/c...00#Frequencies) indicates that max freq will drop from 3.9GHz with 4 cores to 3.8GHz with 5/6 cores on that chip. Turbo is dynamic based on the thermal headroom though, so if your cooling is decent it could still be switching between 3.8 and 3.9 with 5 cores active.

[kracker]

Originally Posted by NookieN

Yeah, this (https://en.wikichip.org/wiki/intel/c...00#Frequencies) indicates that max freq will drop from 3.9GHz with 4 cores to 3.8GHz with 5/6 cores on that chip. Turbo is dynamic based on the thermal headroom though, so if your cooling is decent it could still be switching between 3.8 and 3.9 with 5 cores active.

You're probably right... although even with a clock speed drop I didn't expect the overall speed to decrease... Are there work options on Prime95 which isn't so memory-intensive?

[GP2]

Originally Posted by kracker

You're probably right... although even with a clock speed drop I didn't expect the overall speed to decrease... Are there work options on Prime95 which isn't so memory-intensive?

This post here mentions ECM work as being light on memory bandwidth (although obviously it may use a fair bit of memory space). Whereas LL testing doesn't use much memory but needs lots of memory bandwidth.

[Prime95]

TF does not use memory bandwidth.

[lycorn]

Originally Posted by Prime95

TF does not use memory bandwidth.

True, but... TFing on a CPU?