Lexar NQ780 QLC SSD: Why Large Writes Slow to 350MB/s After the Cache Fills

He expected the move to take under three minutes. The first few seconds were fast, and then the drive slowed to speeds he called very un-SSD-like. The story is personal, but the lesson applies to any Windows PC, workstation or homelab that does large sustained writes. The advertised write speed describes a short burst, not a long copy.

The setup: a PCIe 4.0 drive in a PCIe 3.0 box​

The drive went into an HP Z6 G4, a workstation the author uses as a home server. Its M.2 slots are PCIe 3.0, so the PCIe 4.0 Lexar could never reach its full read speed there. The author says PCIe 3.0 tops out around 3.5 GB/s in practice. He wasn't worried about writes, because the rated 2,500 MB/s fits inside that ceiling.

The drive it replaced was a 256GB Samsung SM961. Notebookcheck's profile describes the SM961 as a PCIe 3.0 x4 drive with 2-bit MLC memory and Samsung's Polaris controller. It is rated at 3,100 MB/s read and 1,400 MB/s write. By the numbers on the box, the Lexar looked faster and cheaper.

The author's own figures put the SM961 at a steady 1.2–1.4 GB/s for this kind of move. On that basis he estimates it would have finished the job in about two minutes. That is his experience, not a controlled benchmark.

Why the copy fell off a cliff​

Bits per cell. SLC stores one bit per cell, MLC two, TLC three and QLC four. Each extra bit doubles the voltage states a cell must distinguish: 4 for MLC, 8 for TLC and 16 for QLC. That makes QLC cheaper per gigabyte, but writes are slower and the cells wear faster.

The SLC cache. Most TLC and QLC drives hide this weakness. They run some free cells in fast one-bit mode as a temporary landing zone, then repack the data into dense storage when the drive is idle. The headline write speed is the speed of that cache. The cache also comes out of free space, so a fuller drive has less of it.

When a write outruns the cache, the drive has to write straight to QLC while clearing the cache in the background. The author puts that at "a few hundred MB/s at best" on his 1TB model.

Independent testing agrees on the pattern​

Independent reviews of the same family show this behavior. TechPowerUp tested the 4TB model with a sequential write stream. Writes started at 5 GB/s and held until 928 GB had been written, meaning about 68% of the drive's capacity runs in SLC mode first. After the cache filled, speeds dropped to around 350 MB/s. TechPowerUp called that expected for a QLC SSD, and it rated the NQ780 better than some rival QLC drives.

PCWorld was kinder to the drive. It said the NQ780's native write rate after the cache is exhausted is two to four times what it has seen from other QLC SSDs. It still noted that the drive slows to 450MBps when writing natively.

These reviews cover other capacities and a different test setup, so the 1TB unit in this story may behave differently. The cache size scales with capacity and free space. The author's own estimate for a 1TB drive is far smaller than the 4TB figure above. The direction is the same, though. A big sequential write that outruns the cache falls to a few hundred MB/s.

DRAM and HMB: a red herring here​

The NQ780 has no onboard DRAM. It uses Host Memory Buffer, borrowing a small slice of system RAM (about 64MB, per the author) to hold its data map. TechPowerUp lists "No DRAM cache" among the drive's drawbacks. The author's view is that DRAM matters mostly for random and mixed workloads, such as several VMs running at once. It matters little for one big sequential copy.

I'd agree. The sharp slowdown in a single large transfer points to the QLC cache cliff, not the missing DRAM. A Proxmox host running many guests is a different story, and DRAM is worth having there.

Is the Lexar a bad drive?​

No, and the author says so himself. It worked as designed, and it stays in his server holding a library of open-weight LLM files that are written once and read many times. That is a good fit for capacity-focused QLC. Lexar's marketing leans on an 8-channel controller for "better sustained write performance". Reviewers found that claim is relative: the drive is good for QLC, not for a TLC drive's consistency.

The mismatch was the job. He used a budget drive for a large disk migration while expecting the headline write figure to hold.

What to check before you buy​

• Look for sustained-write testing. Look for reviews that write past the cache, not just peak CrystalDiskMark numbers.
• Confirm the exact capacity. Cache size, speeds and endurance change between 1TB, 2TB and 4TB models. Lexar's product sheet lists 600 TBW for 1TB, 1,200 for 2TB and 2,400 for 4TB.
• Leave free space. A nearly full drive has a smaller cache, so empty-drive results flatter it.
• Match the drive to the workload. QLC suits read-heavy storage. For frequent large writes, VM migrations or video scratch disks, TLC is the safer pick.
• Know your slot. A PCIe 4.0 drive in a PCIe 3.0 slot is limited by the slot. It doesn't explain a drop to a few hundred MB/s.

A migration safety note​

The author notes that Proxmox only deletes the source disk after a move finishes, so an interrupted job wouldn't cost him the VM. Proxmox's qm disk move documentation says the same: deletion of the source depends on the delete option, and without it the original stays as an unused disk. Even so, a half-copied mess still needs cleaning up. He also ran the job over a flaky SSH session without a terminal multiplexer. Running long jobs inside tmux or Zellij lets them survive a dropped connection. Check the VM and storage state afterward.

The takeaway​

SSD prices have climbed sharply, as the author notes, and cheap drives are tempting. A budget SSD isn't a trap, but the price buys you something specific: cheaper cells, which means slower sustained writes. Spec-sheet numbers show how a drive behaves in its best moments, so read them that way. Pay for TLC when you write big files regularly, and save QLC for the data you write once and read forever.