memory bandwidth for many channels x86 systems

I'm testing the memory bandwidth on a desktop and a server.

Sklyake desktop 4 cores/8 hardware threads
Skylake server Xeon 8168 dual socket 48 cores (24 per socket) / 96 hardware threads

The peak bandwidth of the system is

Peak bandwidth desktop = 2-channels*8*2400 = 38.4 GB/s
Peak bandwidth server  = 6-channels*2-sockets*8*2666 = 255.94 GB/s

I'm using my own triad function from STREAM to measure the bandwidth (full code later)

void triad(double *a, double *b, double *c, double scalar, size_t n) {
  #pragma omp parallel for
  for(int i=0; i<n; i++) a[i] = b[i] + scalar*c[i];
}

Here are results I get

         Bandwidth (GB/s)
threads  Desktop  Server         
1             28      16
2(24)         29     146
4(48)         25     177
8(96)         24     189 

For 1 thread I don't understand why the desktop is so much faster than the server. According to this answer https://stackoverflow.com/a/18159503/2542702 SSE is sufficient to get the full bandwidth of a dual channel system. That's what I observe on the desktop. Two threads only helps slightly and 4 and 8 threads give a worse result But on the server the single threaded bandwidth is much less. Why is this?

On the server I get the best results using 96 threads. I would have thought it would be saturated with far fewer threads. Why are so many threads necessary to saturate the bandwidth on the server? There is a large margin of error in my results and I don't include an error estimate. I took the best result of several runs.

The code

//gcc -O3 -march=native triad.c -fopenmp
//gcc -O3 -march=skylake-avx512 -mprefer-vector-width=512 triad.c -fopenmp
#include <stdio.h>
#include <omp.h>
#include <x86intrin.h>

void triad_init(double *a, double *b, double *c, double k, size_t n) {
  #pragma omp parallel for
  for(size_t i=0; i<n; i++) a[i] = k, b[i] = k, c[i] = k;
}

void triad(double *a, double *b, double *c, double scalar, size_t n) {
  #pragma omp parallel for
  for(size_t i=0; i<n; i++) a[i] = b[i] + scalar*c[i];
}

void triad_stream(double *a, double *b, double *c, double scalar, size_t n) {
#if defined ( __AVX512F__ ) || defined ( __AVX512__ )
  __m512d scalarv = _mm512_set1_pd(scalar);
  #pragma omp parallel for
  for(size_t i=0; i<n/8; i++) {
    __m512d bv = _mm512_load_pd(&b[8*i]), cv = _mm512_load_pd(&c[8*i]);
    _mm512_stream_pd(&a[8*i], _mm512_add_pd(bv, _mm512_mul_pd(scalarv, cv)));
  }		   
#else
  __m256d scalarv = _mm256_set1_pd(scalar);
  #pragma omp parallel for
  for(size_t i=0; i<n/4; i++) {
    __m256d bv = _mm256_load_pd(&b[4*i]), cv = _mm256_load_pd(&c[4*i]);
    _mm256_stream_pd(&a[4*i], _mm256_add_pd(bv, _mm256_mul_pd(scalarv, cv)));
  }		   
#endif
}

int main(void) {
  size_t n = 1LL << 31LL; 
  double *a = _mm_malloc(sizeof *a * n, 64), *b = _mm_malloc(sizeof *b * n, 64), *c = _mm_malloc(sizeof *c * n, 64);
  //double peak_bw = 2*8*2400*1E-3; // 2-channels*8-bits/byte*2400MHz
  double peak_bw = 2*6*8*2666*1E-3; // 2-sockets*6-channels*8-bits/byte*2666MHz
  double dtime, mem, bw;
  printf("peak bandwidth %.2f GB/s\n", peak_bw);

  triad_init(a, b, c, 3.14159, n);
  dtime = -omp_get_wtime();
  triad(a, b, c, 3.14159, n);  
  dtime += omp_get_wtime();
  mem = 4*sizeof(double)*n*1E-9, bw = mem/dtime;
  printf("triad:       %3.2f GB, %3.2f s, %8.2f GB/s, bw/peak_bw %8.2f %%\n", mem, dtime, bw, 100*bw/peak_bw);

  triad_init(a, b, c, 3.14159, n);
  dtime = -omp_get_wtime();
  triad_stream(a, b, c, 3.14159, n);  
  dtime += omp_get_wtime();
  mem = 3*sizeof(double)*n*1E-9, bw = mem/dtime;
  printf("triads:      %3.2f GB, %3.2f s, %8.2f GB/s, bw/peak_bw %8.2f %%\n", mem, dtime, bw, 100*bw/peak_bw);
}

The hardware prefetcher is tuned differently on server vs workstation CPUs. Servers are expected to handle many threads, so the prefetcher will request smaller chunks from RAM. Here is a paper that goes into detail about the issue you're experiencing, but from the other side of the coin:

Hardware Prefetcher Aggressiveness Controllers: Do We Need Them All the Time?