36 lines
1.9 KiB
Markdown
36 lines
1.9 KiB
Markdown
# Validate vs Burn: Hardware Impact Policy
|
||
|
||
## Validate Tests (non-destructive)
|
||
|
||
Tests on the **Validate** page are purely diagnostic. They:
|
||
|
||
- **Do not write to disks** — no data is written to storage devices; SMART counters (power-on hours, load cycle count, reallocated sectors) are not incremented.
|
||
- **Do not run sustained high load** — commands complete quickly (seconds to minutes) and do not push hardware to thermal or electrical limits.
|
||
- **Do not increment hardware wear counters** — GPU memory ECC counters, NVMe wear leveling counters, and similar endurance metrics are unaffected.
|
||
- **Are safe to run repeatedly** — on new, production-bound, or already-deployed hardware without concern for reducing lifespan.
|
||
|
||
### What Validate tests actually do
|
||
|
||
| Test | What it runs |
|
||
|---|---|
|
||
| NVIDIA GPU | `nvidia-smi`, `dcgmi diag` (levels 1–4 read-only diagnostics) |
|
||
| Memory | `memtester` on a limited allocation; reads/writes to RAM only |
|
||
| Storage | `smartctl -a`, `nvme smart-log` — reads SMART data only |
|
||
| CPU | `stress-ng` for a bounded duration; CPU-only, no I/O |
|
||
| AMD GPU | `rocm-smi --showallinfo`, `dmidecode` — read-only queries |
|
||
|
||
## Burn Tests (hardware wear)
|
||
|
||
Tests on the **Burn** page run hardware at maximum or near-maximum load for extended durations. They:
|
||
|
||
- **Wear storage**: write-intensive patterns can reduce SSD endurance (P/E cycles).
|
||
- **Stress GPU memory**: extended ECC stress tests may surface latent defects but also exercise memory cells.
|
||
- **Accelerate thermal cycling**: repeated heat/cool cycles degrade solder joints and capacitors over time.
|
||
- **May increment wear counters**: GPU power-on hours, NVMe media wear indicator, and similar metrics will advance.
|
||
|
||
### Rule
|
||
|
||
> Run **Validate** freely on any server, at any time, before or after deployment.
|
||
> Run **Burn** only when explicitly required (e.g., initial acceptance after repair, or per customer SLA).
|
||
> Document when and why Burn tests were run.
|