Dual RTX PRO 6000 workstation hits PCIe limits

Hot take: this is already deep into "overbuilt desktop" territory, and the real question is less raw capability than whether the current platform is leaving measurable performance on the table for the specific workload.

• –For inference on a single GPU, the CPU and memory-channel upgrade probably won’t matter much.
• –For multi-GPU tensor parallelism, larger models, or heavy host-to-device traffic, more PCIe lanes can help avoid bottlenecks.
• –More memory channels mainly matter when the workload is CPU-side throughput bound or when you need higher aggregate memory bandwidth for data prep, caching, and orchestration.
• –The current x8 PCIe 5.0 links are not inherently bad; for many LLM workloads they are still quite usable.
• –If the goal is to run both cards at full speed under sustained multi-GPU workloads, a workstation or server platform with more lanes would be the cleaner long-term fit.