文章/memory-latency-and-bandwidth-comparison
#灌水1 分钟阅读
内存性能大比较！
对比多套硬件的内存延迟、带宽和 NUMA 表现。
Intel® Memory Latency Checker v3.11b
AMD Ryzen 7 9700X
-> % sudo lshw -c memory -short H/W path Device Class Description ==================================================================== /0/0 memory 64KiB BIOS /0/d memory 640KiB L1 cache /0/e memory 8MiB L2 cache /0/f memory 32MiB L3 cache /0/13 memory 48GiB System Memory /0/13/0 memory [empty] /0/13/1 memory 24GiB DIMM Synchronous Unbuffered (Unregistered) 5600 MHz (0.2 ns) /0/13/2 memory [empty] /0/13/3 memory 24GiB DIMM Synchronous Unbuffered (Unregistered) 5600 MHz (0.2 ns)
-> % sudo ./mlc
Intel(R) Memory Latency Checker - v3.11b
Measuring idle latencies for random access (in ns)...
Numa node
Numa node 0
0 80.3
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 58929.8
3:1 Reads-Writes : 66580.4
2:1 Reads-Writes : 66096.6
1:1 Reads-Writes : 55213.5
Stream-triad like: 67975.4
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 58812.5
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 178.12 58921.6
00002 175.72 59148.9
00008 173.88 58971.0
00015 172.73 58965.9
00050 185.50 59092.4
00100 193.09 58822.9
00200 185.97 58934.0
00300 130.40 47925.9
00400 111.70 37966.8
00500 105.12 31714.0
00700 102.12 23952.2
01000 100.80 17469.9
01300 99.64 13890.6
01700 96.24 10925.5
02500 94.55 7768.6
03500 93.30 5820.1
05000 93.43 4298.0
09000 93.51 2705.6
20000 92.08 1611.9
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 15.7
Local Socket L2->L2 HITM latency 15.8
AMD AI MAX 395+
-> % sudo lshw -c memory -short H/W path Device Class Description ============================================================ /0/0 memory 64KiB BIOS /0/b memory 1280KiB L1 cache /0/c memory 16MiB L2 cache /0/d memory 64MiB L3 cache /0/11 memory 128GiB System Memory /0/11/0 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/1 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/2 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/3 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/4 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/5 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/6 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns) /0/11/7 memory 16GiB Synchronous Unbuffered (Unregistered) 8532 MHz (0.1 ns)
-> % sudo ./mlc
Intel(R) Memory Latency Checker - v3.11b
Measuring idle latencies for random access (in ns)...
Numa node
Numa node 0
0 154.4
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 117081.5
3:1 Reads-Writes : 151323.4
2:1 Reads-Writes : 160109.6
1:1 Reads-Writes : 179796.4
Stream-triad like: 134033.9
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 117020.3
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 206.43 116980.0
00002 205.88 116874.9
00008 203.66 115473.1
00015 204.44 115052.1
00050 207.31 118860.5
00100 203.07 115144.3
00200 217.25 95711.7
00300 196.43 89988.4
00400 180.71 71684.1
00500 173.23 59571.4
00700 169.98 44459.9
01000 163.40 32452.3
01300 161.46 25534.4
01700 160.37 19887.2
02500 159.66 13889.4
03500 159.96 10142.6
05000 158.44 7281.4
09000 160.18 4242.9
20000 159.33 2144.0
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 17.5
Local Socket L2->L2 HITM latency 17.5
Intel Ultra 9 285H
-> % sudo lshw -c memory -short H/W path Device Class Description ============================================================ /0/0 memory 64KiB BIOS /0/12 memory 64GiB System Memory /0/12/0 memory 32GiB SODIMM Synchronous 5600 MHz (0.2 ns) /0/12/1 memory 32GiB SODIMM Synchronous 5600 MHz (0.2 ns) /0/1d memory 288KiB L1 cache /0/1e memory 384KiB L1 cache /0/1f memory 18MiB L2 cache /0/20 memory 24MiB L3 cache /0/21 memory 320KiB L1 cache /0/22 memory 640KiB L1 cache /0/23 memory 10MiB L2 cache /0/24 memory 24MiB L3 cache /0/100/14.2 memory RAM memory
-> % sudo ./Linux/mlc
[sudo] password for user:
Intel(R) Memory Latency Checker - v3.11b
Measuring idle latencies for sequential access (in ns)...
Numa node
Numa node 0
0 270.1
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 78440.3
3:1 Reads-Writes : 76891.5
2:1 Reads-Writes : 77205.5
1:1 Reads-Writes : 77033.2
Stream-triad like: 77061.9
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 78578.3
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 251.56 76195.1
00002 249.19 75876.3
00008 288.68 80638.9
00015 275.13 79408.2
00050 246.77 75827.7
00100 209.39 61399.6
00200 193.59 46888.4
00300 193.61 34877.7
00400 186.20 27353.6
00500 183.14 22485.5
00700 181.26 17103.6
01000 179.27 12473.8
01300 178.00 9898.2
01700 177.66 7782.8
02500 177.11 5550.5
03500 177.62 4089.2
05000 177.50 3024.4
09000 178.82 1829.0
20000 180.87 1019.9
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 91.0
Local Socket L2->L2 HITM latency 92.1
AMD EPYC 7K62
-> % sudo lshw -c memory -short [sudo] password for user: H/W path Device Class Description ================================================================= /0/0 memory 64KiB BIOS /0/21 memory 224GiB System Memory /0/21/0 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/1 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/2 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/3 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/4 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/5 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/6 memory 32GiB DIMM DDR4 Synchronous Registered (Buffered) 3200 MHz (0.3 ns) /0/21/7 memory [empty] /0/24 memory 3MiB L1 cache /0/25 memory 24MiB L2 cache /0/26 memory 192MiB L3 cache
-> % ./mlc
Intel(R) Memory Latency Checker - v3.11b
Measuring idle latencies for random access (in ns)...
Numa node
Numa node 0
0 110.7
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 17503.5
3:1 Reads-Writes : 16796.5
2:1 Reads-Writes : 16591.0
1:1 Reads-Writes : 16305.2
Stream-triad like: 16432.0
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 17127.6
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 501.66 17519.5
00002 504.00 17527.5
00008 532.26 17561.6
00015 528.77 17552.4
00050 538.20 17536.7
00100 531.86 17565.1
00200 532.29 17566.0
00300 532.23 17560.9
00400 532.54 17570.5
00500 533.94 17542.5
00700 526.42 17573.1
01000 519.98 17563.5
01300 498.84 17526.6
01700 386.58 17527.4
02500 174.14 15569.8
03500 145.14 11328.5
05000 135.52 8108.2
09000 124.54 4806.6
20000 119.60 2478.4
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 22.4
Local Socket L2->L2 HITM latency 22.7
[Sat Aug 30 11:21:54 2025] [Hardware Error]: event severity: fatal [Sat Aug 30 11:21:54 2025] [Hardware Error]: Error 0, type: fatal [Sat Aug 30 11:21:54 2025] [Hardware Error]: fru_text: DIMM Locate: P0H0 [Sat Aug 30 11:21:54 2025] [Hardware Error]: section_type: memory error [Sat Aug 30 11:21:54 2025] [Hardware Error]: error_status: Storage error in DRAM memory (0x0000000000040400) [Sat Aug 30 11:21:54 2025] [Hardware Error]: node:0 card:4 module:0 [Sat Aug 30 11:21:54 2025] [Hardware Error]: error_type: 15, physical memory map-out event
真奇怪，又 g 了，明明修好了才是
重启了一下：
[Sat Aug 30 14:17:19 2025] BERT: Error records from previous boot: [Sat Aug 30 14:17:19 2025] [Hardware Error]: It has been corrected by h/w and requires no further action [Sat Aug 30 14:17:19 2025] [Hardware Error]: event severity: corrected [Sat Aug 30 14:17:19 2025] [Hardware Error]: Error 0, type: corrected [Sat Aug 30 14:17:19 2025] [Hardware Error]: fru_text: DIMM Locate: P0H0 [Sat Aug 30 14:17:19 2025] [Hardware Error]: section_type: memory error [Sat Aug 30 14:17:19 2025] [Hardware Error]: error_status: Storage error in DRAM memory (0x0000000000040400) [Sat Aug 30 14:17:19 2025] [Hardware Error]: physical_address: 0x0000003f56549180 [Sat Aug 30 14:17:19 2025] [Hardware Error]: node:0 card:4 module:0 rank:0 bank:6 row:63538 column:528 [Sat Aug 30 14:17:19 2025] BERT: Total records found: 1
-> % free -mh
total used free shared buff/cache available
Mem: 251Gi 4.2Gi 248Gi 5.0Mi 645Mi 247Gi
Swap: 4.0Gi 0B 4.0Gi
居然自己好了。
-> % sudo mlc
Intel(R) Memory Latency Checker - v3.11b
Measuring idle latencies for random access (in ns)...
Numa node
Numa node 0
0 119.2
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 136823.2
3:1 Reads-Writes : 134545.9
2:1 Reads-Writes : 136703.3
1:1 Reads-Writes : 135919.7
Stream-triad like: 136468.7
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 136846.2
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 680.01 136402.3
00002 680.09 136451.1
00008 603.64 130970.7
00015 380.77 105365.8
00050 383.47 105430.4
00100 390.13 105452.3
00200 400.41 105823.8
00300 216.26 107479.2
00400 148.23 88973.3
00500 139.61 73057.0
00700 133.20 53527.2
01000 124.36 38585.8
01300 117.81 29940.8
01700 116.19 23132.7
02500 114.84 15983.0
03500 114.62 11608.9
05000 114.74 8304.9
09000 115.69 4871.7
20000 117.22 2491.8
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 22.4
Local Socket L2->L2 HITM latency 22.7
AMD Ryzen 7 9800X3D
16GB * 2, DDR5 6000MT/s
(0) > .\Windows\mlc.exe
Intel(R) Memory Latency Checker - v3.11b
Measuring idle latencies for random access (in ns)...
Numa node
Numa node 0
0 78.6
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 56582.8
3:1 Reads-Writes : 62128.6
2:1 Reads-Writes : 61323.4
1:1 Reads-Writes : 52797.4
Stream-triad like: 64569.7
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 55632.8
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 220.10 56594.2
00002 248.74 56311.8
00008 337.96 56143.3
00015 286.18 55749.9
00050 246.07 56395.8
00100 219.09 56282.8
00200 177.87 53891.6
00300 113.36 41023.7
00400 123.68 31897.4
00500 117.87 25665.6
00700 117.57 18838.9
01000 110.93 13795.8
01300 114.69 10782.7
01700 106.25 8418.5
02500 108.61 5935.5
03500 134.22 4178.1
05000 120.32 3221.0
09000 99.62 2164.5
20000 101.26 1314.9
Measuring cache-to-cache transfer latency (in ns)...
Unable to enable large page allocation
Using small pages for allocating buffers
Local Socket L2->L2 HIT latency 19.1
Local Socket L2->L2 HITM latency 19.5
讨论
邮箱用于身份识别和回复通知，并由 Waline 存储。