Common Misconceptions

These five misconceptions come up constantly in build discussions, forums, and homelab planning conversations. None of them are unreasonable to believe — they just happen to be wrong.

1. "PCIe Lanes Use CPU Cores"

This one is surprisingly widespread.

The idea is that a GPU at full bandwidth is somehow consuming CPU core capacity — that the CPU is "doing work" to push data through the lane. That is not how it works.

PCIe lanes are managed by the uncore portion of the CPU die. The uncore handles I/O, memory controllers, and interconnects. It is separate hardware from the compute cores.

CPU Die (simplified):
 
  ┌────────────────────────────────────────────┐
  |  Core 0  Core 1  Core 2  Core 3            |  ← compute, runs your code
  |  (completely separate from I/O)            |
  |                                            |
  |  Memory controller  PCIe controller  DMI   |  ← uncore, handles lanes
  └────────────────────────────────────────────┘
 
A GPU saturating a x16 PCIe lane has zero effect on Core 0-3.

What people sometimes confuse this with is I/O wait — a software state where a CPU core is idle, waiting for a slow I/O operation to complete. I/O wait is a scheduling concept. It does not mean the lane is consuming core capacity; it means the software asked for data and the core has nothing else to do while it waits.

2. "All M.2 Slots Are The Same"

This is the one that catches people mid-build when they buy a second NVMe drive expecting the same performance as their first.

Most consumer motherboards have one CPU-direct M.2 slot and one or more chipset-attached slots. The chipset-attached slots:

go through the DMI bus before reaching the CPU
share bandwidth with SATA, USB, and other chipset devices
may run at x2 instead of x4 on budget boards

What the spec sheet says: "4× M.2 NVMe slots"
What it often means:
 
  M.2_1  →  CPU-direct, PCIe 4.0 x4  (7.9 GB/s, low latency)
  M.2_2  →  Chipset, PCIe 4.0 x4    (7.9 GB/s, but shared DMI bandwidth)
  M.2_3  →  Chipset, PCIe 3.0 x4    (3.9 GB/s, older generation)
  M.2_4  →  Chipset, PCIe 3.0 x2    (1.9 GB/s, lane-constrained)

For a Proxmox host or NAS, this matters. VM storage on a chipset-attached slot under sustained I/O will show different behavior under load than the same drive on a CPU-direct slot.

Check the board manual. It will spell this out exactly.

3. "Bifurcating x16 to x8+x8 Halves GPU Performance"

This concern comes up whenever someone is adding a second GPU for AI or compute tasks.

Bifurcation splits the x16 slot into two x8 connections. Each GPU goes from having 16 lanes to 8. That is half the lanes.

But here is what the benchmarks actually show:

RTX 4090 gaming at 4K:
  x16 Gen 4  →  baseline
  x8  Gen 4  →  ~97–99% of baseline  (1–3% difference)
  x4  Gen 4  →  ~80–90% of baseline  (noticeable)
 
Why x8 is fine for most scenarios:
  x8 Gen 4 provides ~16 GB/s
  Most gaming workloads use ~8–12 GB/s of actual PCIe bandwidth
  The GPU is not the bottleneck — the GPU's internal memory bus is

GPUs have their own high-bandwidth VRAM. The PCIe link primarily handles transferring data between system RAM and VRAM. For gaming, that data transfer is not the bottleneck — the rendering computation is.

For AI training with large models and large batches, PCIe bandwidth does matter more. Even then, x8 Gen 4 is rarely the first bottleneck.

4. "You Need Gen 5 NVMe To Have Fast Storage"

Gen 5 NVMe can hit around 14–15 GB/s sequential read. That is an impressive number.

But sequential read speed is one of the least representative metrics for real workloads.

Most storage operations in a running system look like this:

Database queries:     small random reads/writes, latency-sensitive
VM disk I/O:          mixed random, queue depth matters more than peak speed
OS operations:        many small files, file system overhead dominates
Backup/restore:       sequential, but network or CPU is usually the limit
 
Only these workloads genuinely push sequential speed:
  - large video file transfers
  - disk-to-disk imaging
  - direct AI dataset loading into GPU memory

Gen 3 NVMe at ~3.9 GB/s already outpaces what most software can produce. Gen 4 at ~7.9 GB/s is the sensible upgrade for anyone who moves large files regularly. Gen 5 is overhead for everything else.

If you are running Proxmox and your VMs feel slow, upgrading from Gen 4 to Gen 5 NVMe will not fix it. The bottleneck is almost certainly memory, CPU, or network — not storage bandwidth.

5. "USB Is A Separate Bus From PCIe"

USB feels independent because it has its own name, its own cable, its own ecosystem of devices. Internally it is not.

USB controllers are PCIe devices. They attach to the chipset over PCIe lanes, go through the DMI bus, and reach the CPU the same way everything else does.

The path for a USB storage device read:
 
  USB drive  →  USB port  →  USB 3.2 controller (PCIe device)
                                    |
                            PCIe x1 or x4 (chipset lane)
                                    |
                              Chipset / PCH
                                    |
                               DMI bus
                                    |
                                  CPU
                                    |
                               System RAM

The consequence that catches people: USB bandwidth and NVMe bandwidth are not independent on most consumer boards. They share chipset lanes and DMI bandwidth. Saturating USB transfers while running heavy NVMe I/O can create contention — not because the devices are physically connected, but because they share the path to the CPU.

On workstation and HEDT platforms with more CPU-direct lanes and wider DMI connections, this is less of a concern. On a typical consumer ATX board, it is worth knowing.