How to benchmark the RAM in a QNAP NAS

Benchmark and Quality, available in the MemHound window in StorageDiag Workbench, measures memory bandwidth and variations in latency. It also checks for bit errors in the test area under full load. The results help you assess whether the installed memory configuration delivers the expected performance. To look for faults across as much RAM as possible, use the separate RAM test.

Preparethe NAS by stopping tasks that put it under load
Runphases P0 to P8. The test takes several minutes
Readthe verdict, bandwidth in GB/s, latency and error count
Checkthe channel count, memory speed and whether the modules match
Comparethe result with the previous run and the RAM board

01Benchmark or RAM test

Benchmark and Quality against the RAM test
	Benchmark and Quality	RAM test (Quick, Standard, Deep, Hammer, Boot)
Question	Does the memory deliver the expected bandwidth, with steady latency and no errors, under full load?	Are there errors in the tested memory area?
How much RAM it tests	A test area of at most 256 MB per processor thread (Arena in the window); the result also shows how much memory remains untested	Free RAM (Quick, Standard), most RAM (Deep, Hammer), nearly all RAM when booted from a USB drive (Boot)
NAS operation during the test	Continues running, but may be slower	Continues running (Quick, Standard) or stops its services (Deep, Hammer)
Result	PASS, MARGINAL, FAIL, INCONCLUSIVE or PASS-UNQUALIFIED, with bandwidth in GB/s and latency	PASS or FAIL, with the address of each detected error

A PASS in the benchmark does not rule out faults elsewhere in RAM. A RAM test that finds no errors does not establish how fast the memory is. If applications freeze or crash, or the kernel log reports memory errors, start with the RAM test.

02Before you start

Open MemHound: Start menu → Benchmarks → memhound (the Utilities desktop), or click the desktop icon with the same name.
Stop tasks that put the NAS under load, such as backups, large file transfers and virtual machines. The benchmark does not stop services itself, and their activity affects the measurement.

Figure 1. MemHound before a test. The Benchmark and Quality button is in the title bar. The cards below it are for testing RAM for errors.

Clicking Benchmark and Quality starts the test without further confirmation. The first run cannot be interrupted; its button reads Runs to completion. After your first complete run, you can stop later runs in the same browser: it remembers that the first run has finished. A PASS, PASS-UNQUALIFIED or MARGINAL result from a matched set of modules on an idle NAS is published to the public RAM board. The published details include bandwidth, memory configuration (type, speed and channels) and NAS model. The window has no option to disable publication.

03Run the benchmark

Click Benchmark and Quality in the title bar.
If the NAS is short of free memory, for example while applications restart after a CPU Lab run, the window displays Waiting for the NAS to go quiet. It waits up to five minutes, then starts the test whether or not conditions have improved. Stop cancels the wait.
Follow the Test sequence. P0 checks the configuration and modules. P1 to P3 write and verify patterns. P4 measures bandwidth with an increasing number of threads, from one to all available threads, then maintains full load for 60 seconds. P5 measures latency, and P6 generates a database-like load. P7 and P8 leave the data untouched for a while, then check it again.

During the test, the CPU governor is set to performance. The previous setting is restored when the test ends.

Figure 2. Phase P4 of a test (example configuration: two 32 GB DDR4 modules in two channels). The scaling test reaches 29.5 GB/s with eight threads. The tile above it compares the result with the JEDEC figure.

The Bandwidth tile marks 80 % as its pass line. The final verdict uses the threshold for the platform (section 05), so a result shown in amber on the tile may still pass.

04Read the verdict

When the test ends, a dialog displays the verdict and the result of each criterion. Click Close to leave the verdict card on screen with the detailed measurements.

Figure 3. The verdict dialog (example values).

Figure 4. The verdict card (example values). Each row shows a result indicator and a measurement. The notes below identify the untested portion of RAM and the change from the previous run.

The verdicts of Benchmark and Quality
Verdict	What it means
`PASS`	Bandwidth meets or exceeds the pass threshold, jitter is within the limit, no errors were detected, and successive measurement rounds are stable.
`PASS-UNQUALIFIED`	No errors were detected and measurements are stable, but the firmware's memory table does not provide enough information about speed, bus width or channel layout. There is therefore no JEDEC figure for comparison. This applies to most virtual machines and some ARM devices. Bandwidth in GB/s is still shown.
`MARGINAL`	Bandwidth is between the fail and pass thresholds, measurement rounds are unstable, or a corrected ECC error was detected.
`FAIL`	Bandwidth is below the fail threshold, jitter exceeds the limit, a bit error, uncorrected ECC error or critical kernel event was detected, or the modules form a mixed set (section 06).
`INCONCLUSIVE`	Bandwidth or jitter missed the thresholds while the NAS was busy, short of free memory or subject to CPU throttling. The same verdict appears when other processes made the measurement rounds unstable, and when a test was stopped. The measurement conditions prevent a conclusive result.

Bandwidth. The highest sustained read or write bandwidth, expressed as a share of the JEDEC figure, with the pass and fail thresholds for this NAS. If writes were faster, the row identifies that measurement and also reports the verified read result.
Jitter. The ratio of the 99th percentile of latency to the average. A value above 2.50 exceeds the test's limit for latency variation under random load. This measurement alone does not identify a faulty component.
Zero-tolerance. Bit errors, ECC events (CE: corrected; UE: uncorrected) and critical kernel events during the test. One bit error, one UE or one critical event causes a FAIL. Corrected ECC errors lead to MARGINAL.
Stability. The spread across full-load measurement rounds must be at most 5 %, and the ratio of the final quarter's result to the first quarter's must be at least 0.97. This row helps identify bandwidth that falls as the test progresses. If the CPU throttled, the Environment row reports it.
Untested. The portion of RAM outside the benchmark's test area. This run cannot detect a fault there.

05Expected bandwidth: channels and memory speed

In this guide, the JEDEC figure means the theoretical bandwidth of the installed memory. It is calculated from the data bus width and the memory's actual transfer rate. MT/s means millions of transfers per second, not clock frequency in MHz. For 64-bit channels, equivalent to 8 bytes, the formula is channels × MT/s × 8 bytes. This gives a result in MB/s; divide by 1000 to express it in GB/s. A standard DDR5 module also has a total data width of 64 bits, split into two independent 32-bit subchannels. In this formula, 8 bytes represents their combined width, not a single DDR5 subchannel. The JEDEC panel shows the bus, memory speed and calculated result, while the Topology table lists the individual modules.

Worked examples (DDR4 at 2400 MT/s)
Modules	Channels	JEDEC figure at 2400 MT/s	About 77 % of that figure, as in figure 4
2 × 32 GB, one per channel	2	2 × 2400 × 8 B = 38,400 MB/s = 38.4 GB/s	about 29.5 GB/s
1 × 32 GB	1	2400 × 8 B = 19,200 MB/s = 19.2 GB/s	about 14.8 GB/s

One module and one DDR4 channel. MemHound counts only populated channels, so a single module is assessed against the figure for one channel. At the same memory speed, this is half the theoretical bandwidth of two such channels; the measured benchmark result need not fall by exactly half. The percentage alone therefore does not show that one module gives fewer GB/s than two. Also compare the GB/s result with RAM board entries for the same NAS model using two modules.
A lower speed than the label. A ▾ beside a module's MT/s value means it is running below its rated speed, for example at 2400 rather than 2667 MT/s. The JEDEC figure uses the speed the module actually runs at. The lower speed therefore does not lower the percentage, but it does lower the bandwidth in GB/s. All modules running at the same reduced speed can reflect platform limits, an SPD profile or firmware settings. Check the NAS model's specifications and module compatibility.
Soldered memory. LPDDR channels are 16 or 32 bits wide. MemHound reads the width from the firmware's memory table. If the table does not provide enough information about speed, bus width or channel layout, the program cannot calculate a JEDEC figure. The best possible verdict is then PASS-UNQUALIFIED.

The achievable share of theoretical bandwidth depends in part on the memory controller. MemHound therefore uses thresholds specific to the platform. The first matching row applies. Rows marked “two or more modules” require that many modules to be installed.

Pass and fail marks, share of the JEDEC figure
Platform	PASS from	FAIL below
LPDDR (soldered memory), any processor	68 %	48 %
AMD EPYC 7002 or 7003 with DDR4, two or more modules	50 %	35 %
Xeon or EPYC with DDR5, two or more modules	82 %	62 %
Xeon or EPYC with DDR4, two or more modules	80 %	60 %
Core i5, i7, i9 or Ryzen with DDR5, two or more modules	80 %	60 %
Core i5, i7, i9 or Ryzen with DDR4, two or more modules	75 %	55 %
Celeron, Pentium or Atom	65 %	45 %
ARM processors	55 %	40 %
Any other processor	80 %	60 %

If the NAS meets none of the earlier conditions, the last row applies. This includes Core, Ryzen, Xeon and EPYC platforms with a single module. The verdict card then notes that the built-in default thresholds were used. A single DDR4 module on a Ryzen NAS scoring 77 % therefore receives MARGINAL.

06Mixed modules

Benchmark and Quality is designed to qualify a matched set of memory modules. During P0, it checks the modules and rejects the configuration if:

the modules have different part numbers or, if the firmware does not supply them, different capacities, speeds, types or ECC support;
three modules are installed, or an odd number greater than four;
the channels have different memory capacities.

The window displays SETUP NOT ELIGIBLE FOR QA with the reason. The test continues, but the measurements are for information only. The verdict is FAIL, and the result is neither saved nor published.

Figure 5. A mixed pair of modules (example configuration: 32 GB and 16 GB modules from two manufacturers). The measurements remain on screen for information only.

A FAIL for this reason means the configuration does not meet the qualification rules; it does not confirm a memory fault. A mixed set may work correctly. Use the RAM test to check it for errors. To run the qualification, install a matched pair or set and repeat the benchmark.

07Compare the results

The same NAS before and after a change. MemHound keeps the result of each test with a matched set. The verdict card includes a comparison with the previous run, for example vs last run (2026-10-17): read +0.4%. Run the benchmark before replacing modules, changing slots or updating the firmware, then repeat it and compare the results.
Other NAS units. Start menu → Benchmarks → RAM board opens the public board in a browser tab. It lists bandwidth in GB/s, NAS model, memory configuration, speed and channel count. Compare results from the same model with the same memory configuration.
Comparable conditions. Compare runs made without other workloads on the NAS. If the difference is small, run the benchmark two or three times before drawing a conclusion.

08When the result points to hardware

FAIL due to errors. A bit error or uncorrected ECC error indicates a problem revealed under load. A critical kernel event also causes a FAIL, but its log entry needs to be examined; the event alone does not establish that a RAM module is faulty. Run the RAM test (Deep, then Boot) to confirm errors and identify their addresses. Replace a module if diagnostics confirm that it is faulty. A test that detects bit errors or an uncorrected ECC error reports the module model to the public fault list.
MARGINAL due to corrected errors. ECC corrected the detected errors, but that result is insufficient to pass the qualification. Repeat the benchmark. If CE events recur, run the RAM test and identify the affected module.
FAIL or MARGINAL due to bandwidth. Check the configuration first: a single module in a model supporting two channels, modules in the wrong slots, or a lower memory speed than expected. If the set is matched and running at the expected speed, compare the result with the same NAS model on the RAM board.
INCONCLUSIVE. Stop tasks that put the NAS under load and repeat the test. The note below the verdict identifies the factor that affected the measurement.