New PC Performance Checklist: XMP, EXPO, Resizable BAR, SSD Slots and GPU PCIe Links

This guide covers each check, what the vendor documentation supports, and where the original advice goes too far.

1. Turn on XMP or EXPO (and know what you're doing)​

Khan's starting point is sound. Memory modules boot at conservative baseline settings, and a rated kit speed only applies after you select a profile in firmware. Intel's XMP page confirms it: XMP modules first boot with default JEDEC settings, so the first boot succeeds even if the system doesn't support overclocking.

That quote is from Intel's page, but the citation markers above are placeholders. The point stands on Intel's own description. XMP profiles hold preset speed, timing and voltage values, and you enable them in the BIOS.

The steps, as a general pattern:

1. Restart and enter firmware setup. Delete or F2 is common, but your board may differ.
2. Switch to the advanced view if the board offers one.
3. Find the memory or AI Tweaker-style section. The name varies by vendor.
4. Select the XMP profile on Intel platforms or the EXPO profile on AMD platforms. Profile 1 is typically the kit's rated profile.
5. Save, reboot, and confirm the speed in firmware or in Windows.

Caveats the original glosses over:

• It is overclocking. Intel itself calls XMP a type of memory overclocking. It warns that using it beyond Intel's specifications may void the processor warranty. Most builders accept that trade, but it should be a conscious choice.
• Support isn't guaranteed. Intel advises checking that your motherboard supports XMP and that your kit is certified. Intel also says compatible CPUs are the unlocked Core Ultra and Core K-series parts on boards that support the feature, and it names Z and B series boards. If a profile doesn't boot or the system is unstable, go back to defaults and troubleshoot before assuming the kit is faulty.
• Menu names vary. Check the board vendor's manual for the exact labels.

The four-DIMM claim is wrong as stated​

Khan says filling all four slots can cut speeds by 36% and "crowds out" dual-channel operation. The research doesn't back that up. ASUS's installation guidance lists the correct population for each module count:

• One module goes in A2.
• Two modules go in A2 and B2.
• Four modules go in A1, A2, B1 and B2.

So four sticks can still run in dual-channel mode. What does happen on some CPU and board combinations is that four modules put more electrical load on the memory controller, and the maximum stable speed can drop. That is a compatibility limit, not a loss of dual-channel operation. I found no source for the 36% figure, so treat it as unverified.

The practical advice still holds. For two sticks, use the slots your manual recommends, which is usually A2 and B2. Use all four only if you need the capacity, and expect to lower the speed if XMP or EXPO proves unstable.

2. Enable Resizable BAR (if your platform supports it)​

Microsoft's driver documentation explains the background. Discrete GPUs commonly expose only a small slice of their frame buffer over PCI, typically a 256 MB region kept for compatibility with 32-bit operating systems. On GPUs that support resizable BAR, Windows renegotiates the size after firmware initialization under WDDM 2 and later. Once that succeeds, the driver sees the BAR resized to expose the GPU's entire VRAM. That is the mechanism behind Khan's description.

What the mechanism doesn't prove is a big speedup. Microsoft describes how the feature works, not what it does for frame rates. Khan's claim that it "significantly improves gaming performance" is too strong as a blanket statement. One guide notes that NVIDIA turns the feature on per game from its own list, so the gain differs: it helps in some games and does nothing in others. Expect results to vary by title.

The typical sequence, with prerequisites most short guides skip:

1. Update your motherboard BIOS and graphics driver first. Several guides stress this, and early RTX 30-series cards may need a VBIOS update from the card maker.
2. In firmware, enable Above 4G Decoding. Khan's "PCI Subsystem Settings" path is one common example, and it differs by vendor. Some guides list alternatives such as PCI Configuration or IO Ports, and MSI boards may label it as Above 4G memory.
3. Enable the Re-Size BAR Support option, set to Enabled or Auto. It often only appears after step 2.
4. Many guides also say to disable CSM, the legacy compatibility mode. Check the order in your manual, because the system must boot in native UEFI mode.
5. Boot Windows and verify the result.

How to verify:

• On NVIDIA cards, open NVIDIA Control Panel, go to System Information, and check that the Resizable BAR line says "Yes". This is vendor-native and a good first check.
• GPU-Z, which Khan recommends, also reports the status. It is a useful second opinion, and it can hint at a cause if the feature shows as unavailable.

If the option never appears, guides point to an outdated BIOS, an unsupported CPU or board, or Above 4G Decoding still being off. One guide advises that if the PC won't boot after the change, you clear the BIOS settings with the reset jumper or by briefly removing the battery. Know how to do this before you start.

3. Use the M.2 slot wired to the CPU​

Khan's advice is to put your main NVMe SSD in the slot that connects directly to the CPU rather than through the chipset. PassMark's own documentation supports the principle. It tells users to consult the motherboard manual for the best NVMe slot and says that slot is usually the one closest to the CPU.

Note the word "usually". Physical proximity is a clue. The manual's slot and lane tables are the authority.

What the PassMark numbers actually show​

Khan cites a 25.2% loss for a Samsung 970 EVO. PassMark's own page frames the same data differently. The figures below come from a Samsung 970 EVO 500GB across ten runs, comparing the chipset slot with the main CPU slot:

The two percentages describe the same gap from different baselines. A drop from 106.83 to 79.96 is about a 25% loss, while a rise from 79.96 to 106.83 is about 34%. Both are arithmetically correct.

The bigger point is that the gap is concentrated in one small random-I/O test. Sequential transfers differed by roughly 2.4%. For large file copies you would be unlikely to notice. For latency-sensitive random reads, such as OS and application responsiveness, the difference is more plausible. It is one drive on one board, not a universal prediction.

Check the lane sharing before you move anything​

PassMark notes that a drive running with fewer lanes than it supports may be in a slot wired for fewer lanes, or one that shares lanes with another M.2 slot, a PCIe slot or SATA ports. On many boards, using one slot reduces the lanes available to another. Khan's warning about SATA ports being disabled or the GPU slot dropping to x8 is plausible but board-specific. Before relocating a drive, read the "shared bandwidth" table in your manual so you don't trade an SSD gain for a lost SATA port or a narrower GPU slot.

PassMark also points out that a Gen 4 SSD in a Gen 3 slot will be bottlenecked, but a Gen 5 drive in a Gen 4 slot generally won't hit the limit, because those drives are constrained by their controllers.

4. Check your GPU's PCIe link​

The fourth item is a diagnostic rather than a setting. GPU-Z's Bus Interface field shows two values: what the card supports, and what it's running now. Khan's example reads PCIe x16 4.0 @ x4 4.0. He describes the part before the @ as the card's maximum and the part after as the current link.

I couldn't find GPU-Z documentation to confirm that reading, so treat it as the author's description. It matches how the tool is commonly used.

Practical points:

• Wake the GPU first. Some cards lower their link speed at idle to save power. Khan advises running a benchmark or game before reading the field, and the idle case is a well-known source of false alarms.
• A narrower link isn't automatically a fault. The expected width depends on the GPU, the slot and the board. A budget card may be physically designed with fewer lanes. Compare against your GPU's specifications and your manual.
• Check which slot you're using. Khan says the first slot runs at x16 and the second at x4. That is typical but board-specific, so verify it.
• Add-in cards and M.2 drives can steal lanes. This is the same lane-sharing issue from section 3. Removing a card as a test is reasonable, but check the manual first rather than pulling hardware blindly.
• Skip the DDU advice. Khan suggests running Display Driver Uninstaller after a GPU upgrade. Driver cleanup can help with genuine driver problems, but it doesn't change physical PCIe lane allocation. I found no support for it as a fix for a narrow link.

The verdict: a good checklist, not a magic bullet​

The original article's real value is as a post-build checklist. In order of confidence:

1. XMP or EXPO: high value for most builders. Know that it is overclocking, and be ready to fall back to defaults if the system is unstable.
2. M.2 slot choice: low-risk and worth doing during assembly, when you are already reading the manual. The measured benefit in PassMark's test was concentrated in small random reads.
3. Resizable BAR: supported by Microsoft's platform design, but gains vary by title. Verify the status rather than assuming it.
4. GPU link check: a quick sanity test that costs nothing.

The article's claims that four DIMMs lose dual-channel, that a 36% drop applies, and that these tweaks deliver double-digit gains in demanding workloads aren't backed by the evidence I found. Your results depend on your exact CPU, GPU, RAM kit, board, BIOS version and workload. Whatever you change, write down the original firmware values first. Then run a stability test, such as a memory test or a long gaming session, before you call the build finished.