{"slug": "4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003", "title": "4x 32GB and 4x 64GB RAM modules on Epyc 7003", "summary": "A hands-on test on an AMD Epyc 7003 system found that upgrading from 256GB to 512GB of RAM using 8x64GB modules raised LLM prefill speed 1.6–1.9× and decode speed 2.2–2.6× with the Maya/DeepSeek V4.1 Flash UD-Q8_K_XL model. A mixed-capacity 384GB configuration (4x32GB plus 4x64GB) cut performance 40% versus the 256GB setup, which the tester attributed to the memory controller running in \"limp mode.\" Intel Memory Latency Checker v3.13 measured all-reads bandwidth of 160,158.8 MB/sec on the 8x64GB 512GB setup versus 67,980.1 MB/sec on the mixed 384GB configuration.", "body_md": "My assumption was far too optimistic, with a mixed configuration, the memory controller effectively runs in “limp mode.”\n\nUpgrading from 256GB to 512GB, using Maya/DeepSeek V4.1 Flash UD-Q8_K_XL, results in 1.6–1.9× the original prefill speed and 2.2–2.6× the original decode speed.\n\nWith a mixed-capacity configuration (384GB), performance dropped by 40% compared to the 256GB setup, even though it was no longer necessary to stream data from NVMe storage.\n\n```\n> #################### RAM 384GB 4x32GB - 4x64GB\n> \n> Intel(R) Memory Latency Checker - v3.13\n> Measuring idle latencies for random access (in ns)...\n>                 Numa node\n> Numa node            0\n>        0         100.0\n> \n> Measuring Peak Injection Memory Bandwidths for the system\n> Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)\n> Using all the threads from each core if Hyper-threading is enabled\n> Using traffic with the following read-write ratios\n> ALL Reads        :       67980.1\n> 3:1 Reads-Writes :       48768.9\n> 2:1 Reads-Writes :       45117.0\n> 1:1 Reads-Writes :       39321.8\n> Stream-triad like:       51610.1\n> \n> Measuring Memory Bandwidths between nodes within system \n> Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)\n> Using all the threads from each core if Hyper-threading is enabled\n> Using Read-only traffic type\n>                 Numa node\n> Numa node              0\n>        0         65961.3\n> \n> Measuring Loaded Latencies for the system\n> Using all the threads from each core if Hyper-threading is enabled\n> Using Read-only traffic type\n> Inject  Latency Bandwidth\n> Delay   (ns)    MB/sec\n> ==========================\n>  00000  234.22   67511.6\n>  00002  234.02   67527.3\n>  00008  231.23   67564.5\n>  00015  232.57   67648.1\n>  00050  233.64   67813.9\n>  00100  232.22   67781.7\n>  00200  207.79   66955.0\n>  00300  168.28   62819.4\n>  00400  143.20   50386.9\n>  00500  135.45   41090.1\n>  00700  125.29   30000.6\n>  01000  121.16   21470.8\n>  01300  119.52   16765.2\n>  01700  118.08   13019.1\n>  02500  117.01    9125.7\n>  03500  116.25    6700.3\n>  05000  115.61    4868.7\n>  09000  114.69    2960.9\n>  20000  113.81    1644.8\n> \n> Measuring cache-to-cache transfer latency (in ns)...\n> Local Socket L2->L2 HIT  latency        20.8\n> Local Socket L2->L2 HITM latency        21.3\n> root@ubu03:~/intel/Linux# \n> #################################\n> \n> ######### RAM 512GB 8x64GB\n> \n> Intel(R) Memory Latency Checker - v3.13\n> Measuring idle latencies for random access (in ns)...\n>                 Numa node\n> Numa node            0\n>        0          98.4\n> \n> Measuring Peak Injection Memory Bandwidths for the system\n> Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)\n> Using all the threads from each core if Hyper-threading is enabled\n> Using traffic with the following read-write ratios\n> ALL Reads        :      160158.8\n> 3:1 Reads-Writes :      151815.6\n> 2:1 Reads-Writes :      153173.7\n> 1:1 Reads-Writes :      151911.0\n> Stream-triad like:      153978.4\n> \n> Measuring Memory Bandwidths between nodes within system \n> Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)\n> Using all the threads from each core if Hyper-threading is enabled\n> Using Read-only traffic type\n>                 Numa node\n> Numa node              0\n>        0        160257.5\n> \n> Measuring Loaded Latencies for the system\n> Using all the threads from each core if Hyper-threading is enabled\n> Using Read-only traffic type\n> Inject  Latency Bandwidth\n> Delay   (ns)    MB/sec\n> ==========================\n>  00000  370.40  160125.7\n>  00002  369.04  160047.5\n>  00008  360.81  159587.4\n>  00015  362.97  159895.5\n>  00050  283.48  161500.5\n>  00100  188.85  157706.7\n>  00200  129.62   94350.9\n>  00300  122.49   65979.2\n>  00400  119.18   50762.0\n>  00500  114.81   41302.6\n>  00700  112.78   30271.8\n>  01000  112.04   21656.5\n>  01300  111.01   16884.5\n>  01700  109.51   13078.9\n>  02500  110.11    9164.6\n>  03500  109.66    6732.3\n>  05000  109.13    4899.6\n>  09000  109.22    2989.8\n>  20000  109.08    1670.5\n> \n> Measuring cache-to-cache transfer latency (in ns)...\n> Local Socket L2->L2 HIT  latency        20.7\n> Local Socket L2->L2 HITM latency        21.2\n```\n\n", "url": "https://wpnews.pro/news/4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003", "canonical_source": "https://forum.level1techs.com/t/4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003/260592#post_5", "published_at": "2026-11-18 21:27:04+00:00", "updated_at": "2026-11-18 21:44:26.829565+00:00", "lang": "en", "topics": ["ai-infrastructure", "large-language-models"], "entities": ["AMD", "Epyc 7003", "DeepSeek V4.1 Flash", "Maya", "Intel Memory Latency Checker"], "also_reported_by": [], "takeaway": "Populate all eight Epyc 7003 memory channels with identical 64GB modules: 8x64GB delivered 160,158.8 MB/sec all-reads bandwidth and 2.2–2.6× decode speed, while a mixed 4x32GB/4x64GB 384GB setup dropped 40% versus 256GB.", "recommendation": {"advice": "Fill every memory channel with matched-capacity DIMMs when building an LLM inference host; the mixed 4x32GB/4x64GB 384GB config lost 40% versus 256GB, while 8x64GB more than doubled bandwidth.", "playbook": "inference-serving", "playbook_url": "https://wpnews.pro/playbooks/inference-serving"}, "alternates": {"html": "https://wpnews.pro/news/4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003", "markdown": "https://wpnews.pro/news/4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003.md", "text": "https://wpnews.pro/news/4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003.txt", "jsonld": "https://wpnews.pro/news/4x-32gb-and-4x-64gb-ram-modules-on-epyc-7003.jsonld"}}