[Build] Turning a dual Xeon workstation (Z9PE-D8 WS) into a low-budget ComfyUI beast using Process Lasso!

[drim300875-cloud]

Hey everyone!

I wanted to share my recent low-budget experiment to get into local AI generation. I managed to recycle an old server motherboard into a highly optimized, budget-friendly ComfyUI workstation, and the results are surprisingly amazing.

Here is the exact setup and how I optimized it to bypass the classic multi-CPU bottlenecks:

The Hardware Stack

Motherboard: ASUS Z9PE-D8 WS (Dual Socket LGA2011)

CPUs: 2x Intel Xeon E5-2687W v2 (8 cores / 16 threads each)

OS Drive: 1x SATA SSD

Storage (Models/Outputs): 1x NVMe SSD installed via a PCIe adapter on the 3rd blue PCIe slot.

GPU: A modern GPU installed on the 1st blue PCIe slot.

Note on PCIe: Both the GPU and the NVMe SSD are running on PCIe slots controlled directly by CPU 1 (NUMA Node 0), ensuring maximum bandwidth and direct access.

The Multi-CPU Problem & The Fix

On dual-socket systems, if Windows balance tasks across both CPUs randomly, you suffer huge performance drops due to UPI/QPI latency (data traveling between the two physical processors).

To fix this and make it a dedicated ComfyUI machine, I used Process Lasso to isolate the workloads completely:

CPU 1 (NUMA Node 0): Dedicated entirely to running the ComfyUI backend (Python/PyTorch) and managing the GPU/NVMe. It has direct, low-latency access to the hardware.

CPU 2 (NUMA Node 1): Dedicated entirely to a specialized browser (Thorium Browser).

Why Thorium + Process Lasso?

By pinning the Thorium Browser to CPU 2, the heavy ComfyUI web interface (which can be quite CPU-intensive when rendering complex node workflows and live previews) runs entirely on the second processor.

This means 0% interference with the actual generation backend running on CPU 1. Python gets all the focus it needs, and the interface remains buttery smooth even during heavy batches.

Verdict

If you have an old dual Xeon workstation laying around, don't throw it away! With a recent GPU, an NVMe PCIe adapter, and a bit of Process Lasso tuning, you can get a super stable, highly responsive ComfyUI rig for a fraction of the price of a modern platform.

Happy generating! Let me know if you have any questions about the Process Lasso rules or the PCIe routing on this specific board.

[misakaai-dev]

How did you verify which PCIe slots are wired to CPU 1 on that board? That seems worth checking before comparing affinity settings.

[drim300875-cloud]

Les 4 premiers ports en partant du haut sont cpu1 ns0 les 3 derniers en descendant sont sur cpu2 ns 1....jai placé mon gpu et mon ssd sur port 1 bleu et port 3 bleu czblé en 3x16 Le jeu. 5 nov. 2026, 20:33, misakaai-dev ***@***.***> a écrit :

…

How did you verify which PCIe slots are wired to CPU 1 on that board? That seems worth checking before comparing affinity settings. — Reply to this email directly, view it on GitHub <#17530?email_source=notifications&email_token=CTWMPLFS76O6EPLMW2HVFZL5XTIZVA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJUGI2TINRRUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-19425461>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CTWMPLHOH6H6ADBVQZFD3LD5XTIZVAVCNFSNUABIKJSXA33TNF2G64TZHM2TQOJYGMYTOMJYHNCGS43DOVZXG2LPNY5TCMJSGIZTAOJVUF3AE> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CTWMPLCJPPFK33NYYIYSKED5XTIZVA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJUGI2TINRRUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVJTG633UMVZF62LPOM> and Android <https://github.com/notifications/mobile/android/CTWMPLA5TXS7CCFK4KXHWGL5XTIZVA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJUGI2TINRRUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVZTG633UMVZF6YLOMRZG62LE>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>