How do we plan growth and cost?
Installed GPU count is not the capacity you can promise. Some devices are unavailable, reserved, or unsuitable for a job.
Count usable GPUs without overlap
- 64 installed
Fictional identical devices
- 8 unavailable
56 remain
- 8 separate recovery reserves
48 available for normal allocation
Count usable capacity
Start with installed devices. Subtract unavailable capacity and separately designated reserves without counting anything twice.
Then describe the remaining GPUs by type, memory, and placement. An eight-GPU job may still queue when eight scattered devices are free.
Price useful results
In a fictional example, daily cost is 1,000 currency units and output is 50,000 acceptable tasks. Cost per acceptable task is 0.02.
If failures halve useful output while cost stays fixed, that doubles to 0.04.
Use comparable cost boundaries: hardware or rental, supporting infrastructure, energy, and operations as appropriate.
Try it yourself
A fleet has 64 identical GPUs, eight unavailable and eight different GPUs reserved for recovery. How many remain for normal allocation?
Check your answer
48. If those categories overlap, subtracting sixteen is wrong. Recovery reserves are intentional capacity, not automatically waste. Check which commitments must survive a rack failure before reducing them.