Proxmox Host: Lenovo P8 (Threadripper PRO 9965WX)

Identifier: AMD Ryzen Threadripper PRO 9965WX on Lenovo ThinkStation P8 Form factor: workstation tower running Proxmox with GPU passthrough Status: active

A Lenovo ThinkStation P8 workstation repurposed as a Proxmox host with GPU passthrough to a Windows 11 VM. Confirmed by user notes on the test results (“Lenovo P8 Proxmox VM”). This is the only rig in the collection running a Threadripper PRO and the only one tested with the NVIDIA RTX PRO 6000 Blackwell Max-Q.

Build & Cost

ComponentSpecCost
SystemLenovo ThinkStation P8 (30HHCTO1WW), Win 11 Pro, 1400W PSU, 512 GB NVMe$5,045.72 (2025-10-20)
CPUAMD Threadripper PRO 9965WX (Zen 5, 24c/48t, 4.20–5.40 GHz)included
RAM (base)16 GB DDR5-6400 ECC RDIMM (1×16) — Lenovo factory configincluded
RAM upgradeNEMIX 256 GB (4×64 GB) DDR5-4800 ECC RDIMM (PC5-38400)$1,636.99 (2025, ~Aug)
GPU (config A)NVIDIA RTX PRO 6000 Blackwell Max-Q 96 GB$9,145.00 (2025)
GPU (config B)NVIDIA RTX 3090 24 GB$1,132.45 (2025)
GPU (config C)NVIDIA RTX 5090 32 GB$1,999.00 (2025)
System base + RAM upgrade$6,682.71

Beyond Geekbench: vLLM + FP8 Throughput

Geekbench AI numbers reflect single-request throughput. The real ceiling for this rig comes from running vLLM with FP8 quantization on the Blackwell tensor cores — see Alex Ziskind’s benchmark on the same RTX PRO 6000 class: ~5,800–6,000 tokens/sec sustained across 256 concurrent users with Quen 3 Coder 30B FP8. The Max-Q variant in this rig will be slightly lower (lower power envelope) but in the same order of magnitude. This is the practical answer to “why pay $9,145 for the PRO 6000.”

Geekbench AI Runs

All runs are GPU passthrough into a Windows 11 Pro VM (16 GB allocated to guest).

Config A — RTX PRO 6000 Blackwell

DateSingleHalfQuantResult
2026-03-27+44,50067,34132,427470442

Config B — RTX 5090

The same Lenovo P8 host was also tested with an RTX 5090 in passthrough — confirmed by the CPU ID reporting as Threadripper PRO 9965WX in this run (the Ultra 7 host runs the rest of the 5090 tests).

DateSingleHalfQuantResult
2026-03-27+44,34565,43732,958470447

Config C — RTX 3090

DateSingleHalfQuantResult
2026-01-2526,52141,57315,689470438
earlier24,94337,64115,099428699

AI Capability

ConfigurationBest SingleBest HalfBest QuantACITierGPU costGPU $/ACI
+ RTX PRO 6000 Blackwell44,50067,34132,427978.8S$9,145$9.34
+ RTX 509044,34565,43732,958970.5S$1,999$2.06
+ RTX 309026,52141,57315,689580.2A$1,132$1.95 ⭐

The RTX 3090 is the best $/ACI of any compute card in the entire collection (1.95) — but the RTX 5090 on the same host gets within 1% of the PRO 6000’s score for 22% of the GPU cost. The PRO 6000 Blackwell only justifies its premium when a workload needs the 96 GB VRAM. For anything that fits in 32 GB, the 5090 is dramatically more efficient per dollar; for anything that fits in 24 GB, the 3090 is the value champion.

Total invested for the P8 with all three cards + RAM upgrade: 1,636.99 + 1,999 + 18,959.16**

Notes: the NEMIX RAM kit is technically marketed for the ASUS Pro WS WRX90E SAGE SE (also a Threadripper PRO board) but the same DDR5-4800 ECC RDIMM modules work in any compatible Threadripper PRO platform — the P8’s official spec sheet supports DDR5 ECC RDIMM up to 8 channels. Confirmed installed via the P8’s actual host OS, not from Geekbench (which only sees the VM’s allocated 16 GB).

Notes

  • Workstation board identification needed — confirm Lenovo P8 motherboard model and host RAM via dmidecode on the Proxmox host.
  • Threadripper PRO 9965WX is a 24-core Zen 5 part — capable as a CPU compute fallback, though no CPU-only Geekbench AI run was captured for this host.
  • Test variance is much smaller here than on the Ultra 7 host — the Threadripper platform plus stable workstation cooling may be why.

See Also