Whitepaper · Version 0.1
A clearer plan for GPU work.
Kernveil turns a stated workload and available capacity into an estimated schedule. This paper explains the current model, its limits and the path toward compute coordination.
01 / What you can do today
The Capacity Studio accepts four inputs: the number of tasks, the baseline GPU time for one task, the number of available GPUs and an expected utilization percentage. It calculates batch completion time, sustained theoretical throughput and the wait before the final wave begins.
These calculations run in your browser. The current prototype does not connect to GPUs, reserve capacity, dispatch jobs or verify execution. The queue visualization shows a schedule derived from your assumptions.
02 / Why the schedule matters
A GPU count alone cannot tell you when a workload will finish. Task duration and usable capacity change the result. A partly filled final wave still needs time to run.
Kernveil makes that relationship visible. You can change one assumption, inspect the resulting waves and decide which configuration deserves a real benchmark.
03 / The model
Every task is independent and takes the same time. Each task occupies one GPU. The GPUs are identical, exclusive to this batch and available when the batch begins. All tasks enter the queue together. A new wave starts when the previous one finishes.
Utilization is a simplified effective-speed factor that you provide. It is applied uniformly to each GPU. It is not a hardware measurement, and it does not reproduce the behavior of a specific accelerator or model.
- N · Tasks
- A positive whole-number count of tasks.
- t · Baseline task time
- Positive seconds for one task at full utilization.
- G · Available GPUs
- A positive whole-number count of identical GPUs.
- u · Expected utilization
- A percentage from 1 through 100, inclusive.
04 / From inputs to a schedule
First, calculate the effective time for one task and the number of waves needed to fit the batch:
d = t / (u / 100) seconds per taskw = ceil(N / G) waves
The batch finishes after the last wave completes:
Completion time = w × d seconds
The final wave waits while the earlier waves run:
Final-wave wait = (w − 1) × d seconds
With a continuously full queue, the model's sustained theoretical throughput is:
Throughput = G × (u / 100) × 3600 / t tasks per hour
Sustained throughput is not the realized rate of a short batch. If there are fewer tasks than GPUs, each task still needs its full effective duration. Idle GPUs do not accelerate a task in this model.
The Studio rejects invalid, zero, negative and non-finite inputs. Task and GPU counts must be whole numbers. Display rounding does not change the calculation.
05 / A worked example
Consider 120 tasks, each taking 30 baseline seconds, with 8 GPUs at 75% expected utilization.
- Each task takes an effective 40 seconds.
- The batch fits into 15 waves.
- The estimated completion time is 10 minutes.
- The last wave starts after 9 minutes 20 seconds.
- Sustained theoretical throughput is 720 tasks per hour.
This is a hypothetical calculation, not a measured benchmark or service guarantee.
06 / Where the model stops
The estimator does not account for accelerator type, GPU memory, model size, data movement, cold starts, shared resources, variable task lengths, task dependencies, failures, retries or communication between GPUs. It does not model tasks that need several GPUs at once.
It does not calculate prices, energy use or a delivery guarantee. A real workload may behave differently even when the arithmetic is correct. Use the result for early planning and benchmark on the target system before relying on a schedule.
07 / The network direction
The current stage is local planning. Later research may cover provider registration, device declarations, capacity matching, workload dispatch and execution reporting. Each requires real infrastructure and evidence before it can be described as available.
A future settlement layer is a separate research stage. BOT Chain may be explored for records or settlement; this prototype has no deployed registry, contract or payment route. It makes no claim of an operational partnership, token issuance, provider rewards or earnings.
08 / Your inputs and the result
The calculator runs locally and does not require an account or wallet. An optional wallet connection reads the selected public address and chain ID. It does not sign messages, send transactions or change access to the Studio.
A separate optional points-upgrade preview shows a fixed amount of 0.44 BOT on BOT Chain. Its recipient is not configured and confirmation is disabled. The demo cannot submit this payment or issue points. The calculator remains free to use; the preview is not a compute rental price, token purchase or financial-return offer.
A Kernveil estimate means that the displayed assumptions produce the displayed schedule under this model. It does not mean compute has been reserved or that any job has run.