FIELD STUDY 02 · MULTI-GPU HOST
Dual RTX 5090 at PCIe x8/x8: Power, Heat and Vast.ai Host Limits
A real dual-RTX 5090 host configuration covering PCIe lanes, Ryzen 9 3900X, 128 GB, NVMe, network, power and temperature without inventing an x8/x16 performance delta.
This is Mick’s setup, using an AORUS X570 Pro Wi‑Fi, which he confirmed on September 11, 2026. It does not establish the performance difference versus x16: no comparative test was supplied. Before copying the build, check slot spacing, power delivery and heat removal.
Observed configuration
- Processor
- Ryzen 9 3900X · 12C/24T
- Motherboard
- AORUS X570 Pro Wi‑Fi
- Memory
- 128 GB DDR4
- GPUs
- 2× RTX 5090 · 32 GB each
- PCIe
- x8 / x8 observed
- Working storage
- 1 TB NVMe · XFS / Docker
- Measured network
- ≈3 Gbps up/down on a 10G service
PCIe x8/x8 is the host state, not an x8-versus-x16 benchmark. No performance-loss percentage is claimed without a reproducible protocol.
Power and heat
GPU draw and temperatures are Mick’s reported readings for a sustained workload. They do not measure the complete host at the wall. Other renters may draw almost no power, create short spikes or maintain a heavy load. Cooling should therefore handle a plausible worst case, not the average of one day.
The bottlenecks that actually need checking
Physical slot count does not guarantee electrical lanes or clearance between thick cards.
Continuous capacity, circuits, dedicated cables and connectors matter as much as the total wattage number.
A GPU offer can lose value when images, data or preprocessing are constrained by memory, disk or available cores.
Throughput, stability, ports and restarts affect the renter even when the GPUs are fast.
Before adding a third RTX 5090
- Confirm the CPU and motherboard lane map, not only slot count.
- Measure circuit capacity and whole-host wall power.
- Verify spacing, intake and heat exhaust with every card loaded.
- Calculate whether measured demand justifies the new platform, RAM, storage and cooling.
Why Mick is now preparing a server platform
In his September 18, 2026 account, Mick explains that he wants to add several GPUs without rebuilding a complete computer for every expansion. He accepts a higher initial cost to reduce management across multiple machines and provide sufficient PCIe lanes for future workloads. This is his design objective; no workload results from the new platform are published here yet.
The transferable lesson is to compare expansion paths: remaining slots, electrically wired lanes, power, cooling and management. An x16-length connector alone does not establish its electrical link or sharing with other ports. The motherboard manual and a check on the machine are needed before assigning a speed to future customers.