wubbery://ai-infrastructure — training fleets, inference platforms, neoclouds
Your decode workload reaches 2.2% of the compute you installed. Buying more returns nothing.
Not a claim about your team — arithmetic about the workload. Autoregressive decode moves more bytes than it does maths, so the accelerator waits on memory and the utilisation graph reports a machine that is busy doing nothing. WUBBERY measures which resource actually binds, and names the configuration that reaches the silicon you have already paid for.
What actually binds
Utilisation is not the number you think it is
A GPU reported at 95% can be stalled on memory for most of that time. Arithmetic intensity — FLOPs per byte moved — has a ridge point equal to peak FLOPs over peak bandwidth, and decode sits far below it on every modern part. Below the ridge, additional compute returns exactly nothing, and the commercial pressure of the entire industry runs the other way.
A GPU count is not a placement
Forty free GPUs across twenty nodes cannot run an eight-GPU job, and the dashboard shows five times the capacity needed while the queue does not move. An allreduce-heavy job split across NVLink domains runs at cross-domain bandwidth throughout, and the shortfall gets attributed to the model rather than the placement.
At fleet scale, memory failure is routine rather than exceptional
A published frontier run recorded 72 interruptions from uncorrectable memory errors across 54 days on 16,384 accelerators. On a synchronous job every one of those stops the entire run, so one device's memory health is a property of the whole cluster's throughput — and devices signal before they fail.
Measured, on a named alternative
Guards are reported blocked-of-attempted, never as a percentage — “zero violations” without the attempt count is a claim about nothing. Each is posed against a case that passes the check the industry already runs, because a guard that only catches obvious mistakes measures nothing.
Throughput is low, so the plan is more accelerators. At an arithmetic intensity of 2 against an 89.3 ridge point, that purchase returns nothing.
Declines the purchase that returns nothing on this workload, and names the one that does.
'Do not buy compute' is a diagnosis. It does not get anyone a faster cluster on Monday.
Returns a runnable configuration that reaches far more of the silicon you already own, with nothing bought.
The fleet reports forty free GPUs. No node holds eight. The queue does not move and the capacity is real, paid for and unplaceable.
Quantifies the capacity that is paid for and unplaceable, and names the single node to act on.
Two training jobs each hold half the fleet, each waiting for the other, neither taking a step. Utilisation reads 100%.
Prevents the deadlock outright, and when one has already formed, names the eviction that costs least.
A long-context request fits at admission and not at completion. It dies near the end of a generation the customer has already paid for and waited through.
Refuses the request that would die mid-generation, and returns the largest context that will complete.
Eight GPUs are free, so eight are allocated. Split across NVLink domains, every gradient exchange runs at cross-domain speed.
Places a collective-heavy job where it runs at full speed, and refuses to split it rather than silently halving throughput.
A device is degrading. Draining costs capacity now; not draining costs the entire synchronous run at an unknown moment.
Returns a drain-or-hold decision for that device on that job, with the cost of being wrong either way.
Prefill/decode disaggregation is the industry default, so it gets adopted. On sparse traffic it makes the workload slower.
Says whether disaggregation helps or hurts your traffic, and gives the threshold at which the answer flips.
Least-loaded routing sends a 32k prompt with a 30k cached prefix to an idle cold node, and recomputes the lot.
Routes each request to the node that starts generating soonest, congestion included.
Reproduce every number above
One command, one seed, the same answer — on your hardware, in front of you. Almost no infrastructure vendor can do that, and it is the only reason to believe any of the figures on this page.
POST https://api.wubbery.com/v1/bench/all-modules
{ "seed": 1 }What we will not claim: the bounding analysis is first-order: it assumes bandwidth and compute are the only limits and that both reach peak, and it ignores cache reuse, launch overhead, and the fact that sustained bandwidth is typically 70–90% of specification. Use it to decide WHICH resource binds, not to predict absolute throughput. The scheduling analysis takes reuse distances and failure rates as supplied — a wrong input produces a confidently wrong placement, and every module states that in its own output.
Figures measured at seed 1 on the WUBBERY substrate engine and reproducible with the command above. WUBBERY is independent; company and product names are used for identification only and imply no endorsement.