Both sparks are identical: GB10 sm_121, 64 KiB pages, 123.73 GiB
experiment
hardware
fabric
Measured baseline for the two-node cluster, including dual 200Gb RoCE links up and the peer reachable on both.
Question. What does an engine actually get to target?
uname -r; getconf PAGESIZE; grep -E 'MemTotal|Hugepagesize' /proc/meminfo; nproc
nvidia-smi --query-gpu=name,compute_cap,driver_version --format=csv,noheader
for i in enp1s0f1np1 enP2p1s0f1np1; do ethtool $i | awk '/Speed/{print $2}'; done
ping -c3 -W1 192.168.100.11 | tail -1| property | head | peer |
|---|---|---|
| GPU | GB10, cc 12.1 (sm_121), driver 580.173.02 |
same |
| kernel | 6.17.0-1031-nvidia-64k | same |
| page size | 65536 B | 65536 B |
| hugepage size | 512 MiB | 512 MiB |
| MemTotal | 129,735,232 kB (123.73 GiB) | 129,735,360 kB |
| CPU cores | 20 | 20 |
Memory is unified across CPU and GPU. nvidia-smi --query-gpu=memory.total returns [N/A] on this platform, so allocator telemetry is the only authority on GPU-visible capacity.
| fabric | value |
|---|---|
| links | 2, both up at 200000 Mb/s (192.168.100.0/24, 192.168.101.0/24) |
| RDMA devices | 4 RoCE (rocep1s0f0/f1, roceP2p1s0f0/f1) |
| peer RTT | 0.55 ms |
| root storage | 916 GB, 544 GB free |
Cluster totals: 247.5 GiB unified memory, ~50 GB/s theoretical dual-rail fabric.
Verdict. Accepted as the target baseline. Memory bandwidth is still a datasheet number (~273 GB/s) and has not been measured.
Reopen if.
- either node is re-imaged, or the kernel or driver changes