What an inspection checks.
Five areas of the GPU network, one bandwidth test, and a report that states its evidence and its limits. All access is read-only.
Physical inventory
- Question
- Is all the network equipment present, and on the collateral schedule?
- What we look at
- Switches, transceivers and cables, with serial numbers, matched against the schedule the loan relies on.
- Example finding
- 3 transceivers in service are not on the schedule (replaced under warranty).
Cabling vs. design
- Question
- Does every cable connect what the design says it connects?
- What we look at
- The topology discovered from the fabric itself, compared link by link with the borrower's cabling plan. Miswired, swapped, missing and unexpected links.
- Example finding
- 14 links miswired: GPU rails 3 and 4 swapped on 7 nodes.
Link health and history
- Question
- Are the links healthy, and have they stayed healthy?
- What we look at
- Error counters over months: links that keep failing and recovering, rising signal errors, counters reset or stuck, links running at reduced width or speed.
- Example finding
- 6 links running at half width (2 of 4 lanes), with no alarm raised.
Configuration
- Question
- Is the network built and configured the way the borrower claims?
- What we look at
- Topology type and oversubscription, firmware consistency across switches, and changes made outside approved windows.
- Example finding
- 2 leaf switches one firmware release behind the rest.
Takeover readiness
- Question
- Could a new operator take the cluster over after a default?
- What we look at
- Topology, configurations, firmware list, licences and runbooks: present, current, and transferable to the lender or its nominee.
- Example finding
- Fabric-manager licence is in the borrower's name and not assignable.
One active test, at a time you choose.
A cluster-wide all-reduce benchmark measures the bandwidth the network actually delivers, at several job sizes, against the design target. It runs only in an agreed maintenance window. Everything else in the inspection is passive.
Fingerprinted inputs
Every file the inspection relies on is listed in the report with its SHA-256 fingerprint, so anyone can confirm the report was built from those exact files.
Versioned method
Each report states the version of the method it was produced under.
Limits stated
Every report lists what was not checked and where access was limited, as plainly as it lists findings.
Read-only, and never customer workloads.
We need read-only access to the network's management and fabric tools, or the operator's own saved output from standard tools. We never need, and never see, customer workloads or data.
InfiniBand fabrics first. For other network types, ask us.
Before a closing, sale or acquisition
A full inspection with a scored verdict and closing conditions.
Error-history review
A review of the error history the network has already recorded, for a view across months rather than one day.
Periodic re-inspection
The same checks repeated on an agreed schedule, compared with the previous report.