ROCm 7.14 vs ROCm 10.1 with vllm-radiance

Hardware inventory

This inventory combines the collected journal and handoff notes with live verification on 18 October 2026. It distinguishes facts recorded at different times and does not infer missing component models.

Component Recorded specification Confidence / context
Host platform Minisforum MS-01; Intel Core i5-12600H; 32 GB DDR5 Host configuration supplied for the current two-card setup. Live DIMM, board and BIOS details are recorded below.
System memory 32 GB DDR5 installed; approximately 30–30.6 GiB visible to Linux; 8 GiB swap; vm.swappiness=1 recorded Host RAM did not increase when GPU 2 was added.
GPU A AMD Radeon AI PRO R9700, 32 GB, gfx1201 / RDNA4 Present in the single-card phase and retained in the dual-card phase.
GPU B AMD Radeon AI PRO R9700, 32 GB, gfx1201 / RDNA4 Added by 2026-10-16; TP=2 deployment.
GPU A link Current setup: PCIe 4.0 x8 Earlier single-card phases used OCuLink x4 and later a direct Gen5 path; archived measurements differ by phase.
GPU B link Current setup: PCIe 4.0 x4 through a redriver-equipped NVMe-to-OCuLink adapter This adapter/redriver path is required for the second card to operate in the current host. It is the narrower link; archived CPU→GPU bandwidth was about 6.8 GB/s.
Storage NVMe-backed root volume and an optional /kv-cache filesystem tier A later note identifies the root device as nvme1n1p2; the current root device and both NVMe models are verified below.
Power continuity A 650 W UPS was discussed during the two-card load phase This is UPS capacity, not a GPU PSU rating. PSU model and power budget are not recorded.
Archived software images ROCm 7.14 profile: vLLM 0.30.0, AMD PyTorch 2.12, Triton 3.7.1, AITER 0.1.20, libr4d 0.5.0. Later image: ROCm 10.1 with matching component pins. The older build evidence records a ROCm 7.14.0 base and PyTorch +rocm7.14. The separately built ROCm 10.1 profile and result scope are compared in ROCm version evidence.

Recorded hardware verification — 18 October 2026

Read-only SSH inspection used lscpu, filtered SMBIOS records, lsblk, lspci -tv/-vv, DRM sysfs and the running container. Serial numbers, network identifiers and credentials are omitted.

Component Verified value
CPU Intel Core i5-12600H, family 6/model 154/stepping 3; 12 cores, 16 logical CPUs; reported maximum 4.5 GHz; 18 MiB L3
Board Shenzhen Meigao Electronic Equipment Co.,Ltd; AHWSA, revision 1.0
BIOS American Megatrends International, LLC.; version 1.27; SMBIOS date 04/03/2025 (3 April 2025)
DIMM A 16 GB DDR5 SODIMM, Crucial Technology CT16G56C46S5.M8G1, single rank
DIMM B 16 GB DDR5 SODIMM, Micron Technology CT16G56C46S5.C8D, single rank
Memory speed Both modules rated 5600 MT/s, configured at 4800 MT/s, 1.1 V; no ECC
Linux memory 32,833,904,640 bytes (30.58 GiB); swap 8,589,930,496 bytes (approximately 8 GiB)
GPU A / GPU B AMD Navi 48, PCI ID 1002:7551, revision c0, at 03:00.0 / 06:00.0; Radeon AI PRO R9700; XTX branding supplied by the operator
GPU memory Each DRM device exposes 34,208,743,424 usable VRAM bytes (31.86 GiB), consistent with the nominal 32 GB cards
Host-side GPU links 00:01.0: 16 GT/s ×8; 00:06.0: 16 GT/s ×4, both PCIe 4.0
GPU endpoint links Each endpoint reports 32 GT/s ×16; these downstream readings do not remove the upstream ×8/×4 bottlenecks
Root NVMe Kingston OM3PGP41024P-A0, 953.9 GiB; root filesystem currently on its second partition
Other NVMe VS000960KWXFN, 894.3 GiB; PCI controller identified as SK hynix PE6010
Ethernet controllers Intel X710 dual 10GbE SFP+, I226-V and I226-LM; controller inventory only, no link-speed test
Host software Ubuntu 26.04.1 LTS; Linux 7.0.0-999-generic; Podman 5.7.0
Running image software vLLM 0.30.0, PyTorch 2.12.0+rocm10.1, Triton 3.7.1+gitf0b55c07

The PCI tree contains intermediate bridges on both GPU paths. The host ports above establish the limiting negotiated widths; endpoint sysfs alone would misleadingly suggest two Gen5 ×16 host connections. Adapter/redriver model numbers and PSU specifications remain unverified. These current observations do not retroactively identify the hardware of every historical run.

Topology timeline

  1. One GPU, early attachment: one R9700 connected through OCuLink. The early measured CPU↔GPU path was approximately 6.6 GB/s.
  2. One GPU, direct link: the card was rewired to a direct path (the later inspection records an upstream PCIe 4.0 ×8 limit); later bandwidth tests recorded approximately 12.9–13.8 GB/s, depending on test and direction.
  3. Two GPUs: GPU B was added on an x4 path, measuring approximately 6.8 GB/s. Tensor parallelism uses both devices, so the link asymmetry became a system-level constraint.

The live inspection above fills gaps in the historical inventory. PSU and adapter part numbers remain unknown. Device serials are intentionally omitted.

Home · Overview · ROCm 10.1 results · Charts