Why Does Nvidia Report Less GPU Memory Than Specified?

Background

In production the author used two Nvidia cards, the L40S and the A6000, both listed by Nvidia as having 48 GB of VRAM. However, running nvidia‑smi inside a container displayed only 45 GB for the L40S and 48 GB for the A6000, a discrepancy of about 3 GB.

Observation

Running nvidia‑smi on the L40S returned 46068MiB (≈45 GB), while the A6000 reported 49140MiB (≈48 GB). The author linked the official datasheets for both cards and captured screenshots of the command output.

Analysis

Inspecting the nvidia‑smi output revealed that the L40S has ECC set to 0 (enabled), whereas the A6000 shows ECC as off. The author notes that Nvidia’s enterprise‑grade GPUs enable Volatile Uncorr. ECC by default, which consumes a portion of VRAM, while consumer‑grade GPUs keep ECC disabled.

Verification

To confirm the hypothesis, the author used a Tesla‑T4 (an enterprise card) and performed the following steps:

With ECC enabled, nvidia‑smi reported 15360MiB (≈15 GB) of usable memory.

Disabled ECC with $ nvidia-smi -i 0 -e 0 and observed the memory remained the same.

Enabled ECC with $ nvidia-smi -i 0 -e 1, then rebooted the machine. After reboot, nvidia‑smi showed 16384MiB (≈16 GB), confirming that ECC activation changes the reported usable VRAM.

Conclusion

The apparent missing memory on the L40S is not a manufacturing defect but the result of ECC memory being active, which reserves a few gigabytes of VRAM. Enterprise‑grade Nvidia GPUs like the L40S have ECC on by default, while consumer cards such as the RTX A6000 keep it off, explaining the observed differences.