# FIXME
# FIXME performance counters might be wrong. This will only affect the Benchmark model
# FIXME bandwidth measurements need validation
# FIXME
kerncraft version: 0.7.2
model name: Intel(R) Xeon(R) Gold 5122 CPU @ 3.60GHz
model type: Intel Core Skylake SP
sockets: 2
cores per socket: 4
threads per core: 2
NUMA domains per socket: 1
cores per NUMA domain: 4
clock: 3.6 GHz
FLOPs per cycle:
SP:
total: 64
FMA: 64
ADD: 32
MUL: 32
DP:
total: 32
FMA: 32
ADD: 16
MUL: 16
micro-architecture: SKX
compiler:
!!omap
- icc: -O3 -fno-alias -xCORE-AVX512
- clang: -O3 -march=skylake-avx512 -D_POSIX_C_SOURCE=200112L
- gcc: -O3 -march=skylake-avx512
cacheline size: 64 B
overlapping model:
ports: ["0", "0DV", "1", "2", "3", "4", "5", "6", "7"]
performance counter metric:
Max(UOPS_DISPATCHED_PORT_PORT_0:PMC[0-3],
UOPS_DISPATCHED_PORT_PORT_1:PMC[0-3],
UOPS_DISPATCHED_PORT_PORT_4:PMC[0-3],
UOPS_DISPATCHED_PORT_PORT_5:PMC[0-3],
UOPS_DISPATCHED_PORT_PORT_6:PMC[0-3],
UOPS_DISPATCHED_PORT_PORT_7:PMC[0-3])
non-overlapping model:
ports: ["2D", "3D"]
performance counter metric: T_OL + T_L1L2 + T_L2L3 + T_L3MEM
memory hierarchy:
- level: L1
performance counter metrics:
accesses: MEM_INST_RETIRED_ALL_LOADS:PMC[0-3]
misses: L1D_REPLACEMENT:PMC[0-3]
evicts: L2_TRANS_L1D_WB:PMC[0-3]
cache per group:
sets: 64
ways: 8
cl_size: 64
replacement_policy: 'LRU'
write_allocate: True
write_back: True
load_from: L2
store_to: L2
size per group: 32.00 kB
groups: 8
cores per group: 1
threads per group: 2
- level: L2
non-overlap upstream throughput: [64 B/cy, 'half-duplex']
performance counter metrics:
accesses: L1D_REPLACEMENT:PMC[0-3]
misses: L2_LINES_IN_ALL:PMC[0-3]
evicts: L2_TRANS_L2_WB:PMC[0-3]
cache per group:
sets: 1024
ways: 16
cl_size: 64
replacement_policy: 'LRU'
write_allocate: True
write_back: True
load_from: null # L3 is a victim cache, thus unless a hit in L3, misses get forwarded to MEM
victims_to: L3 # all victims, modified or not are passed onto L3
store_to: L3
size per group: 1.00 MB
groups: 8
cores per group: 1
threads per group: 2
- level: L3
non-overlap upstream throughput: [16 B/cy, 'full-duplex']
performance counter metrics:
accesses: L2_LINES_IN_ALL:PMC[0-3]
# FIXME not all misses in L2 lead to loads from L3, only the hits do
misses: (CAS_COUNT_RD:MBOX0C[01] + CAS_COUNT_WR:MBOX0C[01] +
CAS_COUNT_RD:MBOX1C[01] + CAS_COUNT_WR:MBOX1C[01] +
CAS_COUNT_RD:MBOX2C[01] + CAS_COUNT_WR:MBOX2C[01] +
CAS_COUNT_RD:MBOX3C[01] + CAS_COUNT_WR:MBOX3C[01] +
CAS_COUNT_RD:MBOX4C[01] + CAS_COUNT_WR:MBOX4C[01] +
CAS_COUNT_RD:MBOX5C[01] + CAS_COUNT_WR:MBOX5C[01])
evicts: L2_TRANS_L2_WB:PMC[0-3]
cache per group:
sets: 16896
# TODO is actuall something else, but necessary to get to 16.5 MB
ways: 16
# TODO is actually 11, but pycachesim only supports powers of two
cl_size: 64
replacement_policy: 'LRU'
write_allocate: False
write_back: True
size per group: 16.50 MB
groups: 2
cores per group: 4
threads per group: 8
- level: MEM
cores per group: 4
threads per group: 8
non-overlap upstream throughput: ['full socket memory bandwidth', 'half-duplex']
penalty cycles per read stream: 0
size per group:
benchmarks:
kernels:
load:
read streams:
streams: 1
bytes: 8.00 B
read+write streams:
streams: 0
bytes: 0.00 B
write streams:
streams: 0
bytes: 0.00 B
FLOPs per iteration: 0
copy:
read streams:
streams: 1
bytes: 8.00 B
read+write streams:
streams: 0
bytes: 0.00 B
write streams:
streams: 1
bytes: 8.00 B
FLOPs per iteration: 0
update:
read streams:
streams: 1
bytes: 8.00 B
read+write streams:
streams: 1
bytes: 8.00 B
write streams:
streams: 1
bytes: 8.00 B
FLOPs per iteration: 0
triad:
read streams:
streams: 3
bytes: 24.00 B
read+write streams:
streams: 0
bytes: 0.00 B
write streams:
streams: 1
bytes: 8.00 B
FLOPs per iteration: 2
daxpy:
read streams:
streams: 2
bytes: 16.00 B
read+write streams:
streams: 1
bytes: 8.00 B
write streams:
streams: 1
bytes: 8.00 B
FLOPs per iteration: 2
measurements:
L1:
1:
threads per core: 1
cores:
- 1
- 2
- 3
- 4
threads:
- 1
- 2
- 3
- 4
size per core:
- 21.12 kB
- 21.12 kB
- 21.12 kB
- 21.12 kB
size per thread:
- 21.12 kB
- 21.12 kB
- 21.12 kB
- 21.12 kB
total size:
- 21.12 kB
- 42.24 kB
- 63.36 kB
- 84.48 kB
results:
load:
- 42.98 GB/s
- 85.08 GB/s
- 127.45 GB/s
- 169.92 GB/s
copy:
- 56.07 GB/s
- 111.50 GB/s
- 164.90 GB/s
- 221.50 GB/s
update:
- 56.54 GB/s
- 112.25 GB/s
- 168.50 GB/s
- 224.75 GB/s
triad:
- 45.90 GB/s
- 89.81 GB/s
- 127.29 GB/s
- 169.57 GB/s
daxpy:
- 36.62 GB/s
- 71.30 GB/s
- 103.52 GB/s
- 135.26 GB/s
2:
threads per core: 2
cores:
- 1
- 2
- 3
- 4
threads:
- 2
- 4
- 6
- 8
size per core:
- 21.12 kB
- 21.12 kB
- 21.12 kB
- 21.12 kB
size per thread:
- 10.56 kB
- 10.56 kB
- 10.56 kB
- 10.56 kB
total size:
- 21.12 kB
- 42.24 kB
- 63.36 kB
- 84.48 kB
results:
load:
- 49.61 GB/s
- 98.80 GB/s
- 147.98 GB/s
- 198.22 GB/s
copy:
- 55.98 GB/s
- 111.56 GB/s
- 167.08 GB/s
- 220.42 GB/s
update:
- 56.53 GB/s
- 112.72 GB/s
- 168.95 GB/s
- 225.31 GB/s
triad:
- 54.01 GB/s
- 104.58 GB/s
- 153.02 GB/s
- 200.93 GB/s
daxpy:
- 41.11 GB/s
- 80.28 GB/s
- 115.71 GB/s
- 152.81 GB/s
L2:
1:
threads per core: 1
cores:
- 1
- 2
- 3
- 4
threads:
- 1
- 2
- 3
- 4
size per core:
- 660.00 kB
- 660.00 kB
- 660.00 kB
- 660.00 kB
size per thread:
- 660.00 kB
- 660.00 kB
- 660.00 kB
- 660.00 kB
total size:
- 660.00 kB
- 1.32 MB
- 1.98 MB
- 2.64 MB
results:
load:
- 27.15 GB/s
- 54.09 GB/s
- 80.61 GB/s
- 106.41 GB/s
copy:
- 43.53 GB/s
- 90.07 GB/s
- 127.73 GB/s
- 171.81 GB/s
update:
- 50.38 GB/s
- 98.47 GB/s
- 147.91 GB/s
- 197.20 GB/s
triad:
- 43.38 GB/s
- 83.72 GB/s
- 124.83 GB/s
- 166.04 GB/s
daxpy:
- 36.29 GB/s
- 71.29 GB/s
- 103.33 GB/s
- 136.48 GB/s
2:
threads per core: 2
cores:
- 1
- 2
- 3
- 4
threads:
- 2
- 4
- 6
- 8
size per core:
- 660.00 kB
- 660.00 kB
- 660.00 kB
- 660.00 kB
size per thread:
- 330.00 kB
- 330.00 kB
- 330.00 kB
- 330.00 kB
total size:
- 660.00 kB
- 1.32 MB
- 1.98 MB
- 2.64 MB
results:
load:
- 35.29 GB/s
- 70.28 GB/s
- 104.67 GB/s
- 139.63 GB/s
copy:
- 42.23 GB/s
- 83.70 GB/s
- 124.33 GB/s
- 167.50 GB/s
update:
- 50.09 GB/s
- 99.77 GB/s
- 149.87 GB/s
- 198.82 GB/s
triad:
- 52.38 GB/s
- 100.00 GB/s
- 147.40 GB/s
- 193.31 GB/s
daxpy:
- 41.14 GB/s
- 80.22 GB/s
- 116.23 GB/s
- 155.08 GB/s
L3:
1:
threads per core: 1
cores:
- 1
- 2
- 3
- 4
threads:
- 1
- 2
- 3
- 4
size per core:
- 10.56 MB
- 5.28 MB
- 3.52 MB
- 2.64 MB
size per thread:
- 10.56 MB
- 5.28 MB
- 3.52 MB
- 2.64 MB
total size:
- 10.56 MB
- 10.56 MB
- 10.56 MB
- 10.56 MB
results:
load:
- 22.40 GB/s
- 44.77 GB/s
- 65.71 GB/s
- 89.26 GB/s
copy:
- 25.32 GB/s
- 49.70 GB/s
- 72.89 GB/s
- 98.62 GB/s
update:
- 41.24 GB/s
- 81.14 GB/s
- 122.22 GB/s
- 166.44 GB/s
triad:
- 25.61 GB/s
- 50.02 GB/s
- 73.23 GB/s
- 98.95 GB/s
daxpy:
- 32.07 GB/s
- 62.65 GB/s
- 89.91 GB/s
- 120.65 GB/s
2:
threads per core: 2
cores:
- 1
- 2
- 3
- 4
threads:
- 2
- 4
- 6
- 8
size per core:
- 10.56 MB
- 5.28 MB
- 3.52 MB
- 2.64 MB
size per thread:
- 5.28 MB
- 2.64 MB
- 1.76 MB
- 1.32 MB
total size:
- 10.56 MB
- 10.56 MB
- 10.56 MB
- 10.56 MB
results:
load:
- 26.18 GB/s
- 51.85 GB/s
- 75.82 GB/s
- 101.39 GB/s
copy:
- 26.22 GB/s
- 51.83 GB/s
- 76.40 GB/s
- 102.84 GB/s
update:
- 43.51 GB/s
- 86.75 GB/s
- 129.86 GB/s
- 174.54 GB/s
triad:
- 26.39 GB/s
- 51.80 GB/s
- 76.27 GB/s
- 102.66 GB/s
daxpy:
- 37.43 GB/s
- 73.16 GB/s
- 106.53 GB/s
- 142.76 GB/s
MEM:
1:
threads per core: 1
cores:
- 1
- 2
- 3
- 4
threads:
- 1
- 2
- 3
- 4
size per core:
- 240.00 MB
- 120.00 MB
- 80.00 MB
- 60.00 MB
size per thread:
- 240.00 MB
- 120.00 MB
- 80.00 MB
- 60.00 MB
total size:
- 240.00 MB
- 240.00 MB
- 240.00 MB
- 240.00 MB
results:
load:
- 12.03 GB/s
- 24.38 GB/s
- 34.83 GB/s
- 45.05 GB/s
copy:
- 12.32 GB/s
- 24.40 GB/s
- 32.82 GB/s
- 37.00 GB/s
update:
- 20.83 GB/s
- 40.25 GB/s
- 48.81 GB/s
- 54.84 GB/s
triad:
- 11.64 GB/s
- 23.17 GB/s
- 34.78 GB/s
- 42.97 GB/s
daxpy:
- 17.69 GB/s
- 34.02 GB/s
- 48.12 GB/s
- 55.73 GB/s
2:
threads per core: 2
cores:
- 1
- 2
- 3
- 4
threads:
- 2
- 4
- 6
- 8
size per core:
- 240.00 MB
- 120.00 MB
- 80.00 MB
- 60.00 MB
size per thread:
- 120.00 MB
- 60.00 MB
- 40.00 MB
- 30.00 MB
total size:
- 240.00 MB
- 240.00 MB
- 240.00 MB
- 240.00 MB
results:
load:
- 15.33 GB/s
- 28.32 GB/s
- 41.34 GB/s
- 53.02 GB/s
copy:
- 13.96 GB/s
- 26.61 GB/s
- 34.39 GB/s
- 38.96 GB/s
update:
- 26.47 GB/s
- 47.82 GB/s
- 56.70 GB/s
- 62.78 GB/s
triad:
- 14.42 GB/s
- 26.66 GB/s
- 36.94 GB/s
- 44.01 GB/s
daxpy:
- 20.96 GB/s
- 39.12 GB/s
- 51.55 GB/s
- 58.37 GB/s