Nameplate Analytics LLC
Specimen sandbox

One bill, three ways of reading the cluster.

The same synthetic bill and the same detectors, three times. In the first view the cluster is read once. In the second it is sampled for one week, the length of a Diagnostic. In the third it is sampled across the whole billing period. Switch between them and watch which dollars the report is willing to call proven. Nothing on this page is live: it takes no input and sends nothing.

01What we are willing to claim

$100,101 per month

That is what this view of the report would put its name to: $1,201,213 a year at the same run rate. The cluster was read once. A reading observes an instant, so every dollar drawn from it is LIKELY, and the part of the finding below that rests on the cluster contributes nothing to PROVEN.

PROVEN
$1,976per month, counted once per machine
LIKELY
$98,125per month, one stated assumption each
UNPROVEN
$213,121per month, observed and not claimed

Figures are rounded to the dollar, so the bands can differ from the total by a dollar. The unproven band is never in the total.

$100,101 per month

That is what this view of the report would put its name to: $1,201,213 a year at the same run rate. The cluster was sampled for one week of the billing period, the length of a Diagnostic. The week's hours are proven, so the cluster finding's share for those hours moves into PROVEN. The rest of the month stays LIKELY, and the total does not change.

PROVEN
$16,077per month, counted once per machine
LIKELY
$84,024per month, one stated assumption each
UNPROVEN
$213,121per month, observed and not claimed

Figures are rounded to the dollar, so the bands can differ from the total by a dollar. The unproven band is never in the total.

$100,101 per month

That is what this view of the report would put its name to: $1,201,213 a year at the same run rate. The cluster was sampled across the whole billing period. The hours observed are proven, so the part of the finding below that rests on the cluster moves into PROVEN, and the total does not change.

PROVEN
$62,408per month, counted once per machine
LIKELY
$37,694per month, one stated assumption each
UNPROVEN
$213,121per month, observed and not claimed

Figures are rounded to the dollar, so the bands can differ from the total by a dollar. The unproven band is never in the total.

02How to read a finding

PROVEN

The evidence covers the hours it claims, so there is no assumption to accept or reject. A finding reaches this rating one of two ways: commitment dollars reconciled against a closed billing period, or cluster allocation for the hours the readings covered, rather than one reading carried across the period.

LIKELY

The billing data proves the cost. The waste follows from one stated assumption, printed with the finding, which you can accept or reject line by line.

UNPROVEN

There is a cost and a suspicion. The data that would settle it does not exist yet. These dollars are listed and then excluded from the recoverable total.

03Findings

  • PROVENCommitment bought and not consumed
  • LIKELYNon-production GPUs running around the clock, inferred from names
  • LIKELYNon-production GPUs running around the clock, by tag
  • LIKELYGPUs no pod requested where sampled, carried across the hours not sampled$60,432 per month from 0 sampled hours
  • LIKELYStorage with no compute attached
  • LIKELYSteady GPU usage with no commitment covering it
  • UNPROVENGPUs claimed by pods, with no profiling metrics
  • UNPROVENSpend on superseded accelerators
  • UNPROVENGPU spend with no utilization data

The proven part of the zero-allocation finding in this view is $0.

  • PROVENGPUs no pod requested in any reading across the window$14,101 per month from 168 sampled hours
  • PROVENCommitment bought and not consumed
  • LIKELYNon-production GPUs running around the clock, inferred from names
  • LIKELYNon-production GPUs running around the clock, by tag
  • LIKELYGPUs no pod requested where sampled, carried across the hours not sampled$46,331 per month from 168 sampled hours
  • LIKELYStorage with no compute attached
  • LIKELYSteady GPU usage with no commitment covering it
  • UNPROVENGPUs claimed by pods, with no profiling metrics
  • UNPROVENSpend on superseded accelerators
  • UNPROVENGPU spend with no utilization data
  • PROVENGPUs no pod requested in any reading across the window$60,432 per month from 720 sampled hours
  • PROVENCommitment bought and not consumed
  • LIKELYNon-production GPUs running around the clock, inferred from names
  • LIKELYNon-production GPUs running around the clock, by tag
  • LIKELYStorage with no compute attached
  • LIKELYSteady GPU usage with no commitment covering it
  • UNPROVENGPUs claimed by pods, with no profiling metrics
  • UNPROVENSpend on superseded accelerators
  • UNPROVENGPU spend with no utilization data

04Cluster allocation

One reading of the cluster. Pods asked for most of the accelerators, and some nodes had nothing asking for theirs at that moment. A reading cannot say how long that had been true, so the report carries it across the period and rates it LIKELY.

Hours sampled: 0 of 720 in the billing period.

Readings across one week of the period. A GPU counts as unrequested only if it was unrequested in every reading that saw its node. The week proves its own hours; the rest of the month is carried from them and rated LIKELY.

Hours sampled: 168 of 720 in the billing period.

Readings across the whole period. A GPU counts as unrequested only if it was unrequested in every reading that saw its node, so one pod asking for it once takes it off the list. What remains is the idle capacity the window proves.

Hours sampled: 720 of 720 in the billing period.

05What this report cannot tell you

Counting each machine once, and setting aside what is already claimed, $213,121 per month is observed and not claimed. None of it is in the recoverable figure. Closing these gaps is the work of the next engagement, and the honest reason to do it is that the answer might be that nothing is wrong.

Counting each machine once, and setting aside what is already claimed, $213,121 per month is observed and not claimed. None of it is in the recoverable figure. Closing these gaps is the work of the next engagement, and the honest reason to do it is that the answer might be that nothing is wrong.

Counting each machine once, and setting aside what is already claimed, $213,121 per month is observed and not claimed. None of it is in the recoverable figure. Closing these gaps is the work of the next engagement, and the honest reason to do it is that the answer might be that nothing is wrong.

06Order of work

PROVEN findings first, then LIKELY, each by the dollars it holds.

  1. PROVEN Commitment bought and not consumed
  2. LIKELY GPUs no pod requested where sampled, carried across the hours not sampled
  3. LIKELY Steady GPU usage with no commitment covering it
  4. LIKELY Non-production GPUs running around the clock, inferred from names
  5. LIKELY Non-production GPUs running around the clock, by tag
  6. LIKELY Storage with no compute attached
  1. PROVEN GPUs no pod requested in any reading across the window
  2. PROVEN Commitment bought and not consumed
  3. LIKELY GPUs no pod requested where sampled, carried across the hours not sampled
  4. LIKELY Steady GPU usage with no commitment covering it
  5. LIKELY Non-production GPUs running around the clock, inferred from names
  6. LIKELY Non-production GPUs running around the clock, by tag
  7. LIKELY Storage with no compute attached
  1. PROVEN GPUs no pod requested in any reading across the window
  2. PROVEN Commitment bought and not consumed
  3. LIKELY Steady GPU usage with no commitment covering it
  4. LIKELY Non-production GPUs running around the clock, inferred from names
  5. LIKELY Non-production GPUs running around the clock, by tag
  6. LIKELY Storage with no compute attached

07Method

Northwind Systems, Inc. is a fictional company. The bill is a synthetic Cost and Usage Report built by the same generator that produces the published specimen report, with the same recipe and seed, so it is not a customer's and not a disguised one. The cluster is synthetic as well.

The week and month views are built from one synthetic reading repeated at hourly timestamps and merged by the same code a real set of readings goes through, because the hours are counted from the timestamps and nothing else. A real cluster changes between readings and this one does not, so what those views show is how the arithmetic treats a window, not what a cluster does.

The measurements the practice rests on are real and documented on the method page. These figures are not measurements of anything. A shorter window proves proportionally less, and the rest stays LIKELY. The total does not move between views; the ratings do, and the dollars you can defend to a skeptical engineer move with them.