SpareGrid Compute · architecture

Architecture, and its limits

Ten questions a careful buyer should ask, answered at the maturity we can honestly claim. Every answer lists what would be done and where it falls short.

How to read the labels

Each answer below carries one of four labels. The count is how many of the ten answers sit at that level today.

Implemented0 of 10
Built and running in production.
Tested0 of 10
Built and exercised in a controlled test, not yet in production.
Planned7 of 10
A design we intend to build. No working code yet.
Conceptual3 of 10
A direction with open questions. It may change or prove unworkable.

Network answers by label: Implemented: 0. Tested: 0. Planned: 7. Conceptual: 3.

What exists today

Nothing in the network itself. Only this:

  • ImplementedThis website, its calculator and its demo dashboard.
  • ImplementedSign-up and compute-request forms: validated on the server, rate limited, and saved privately when storage is connected.
Jump to a question (10)

Question 01

How are workloads isolated from the device they run on?

PlannedDesign only. No agent has been built.

Each task would run inside a virtual machine with no access to the owner's files, accounts or home network.

Approach

  • A hardware-virtualised sandbox per task, not just a separate process.
  • No host filesystem mounts. Task data lives in a scratch volume that is wiped afterwards.
  • Outbound network limited to SpareGrid job endpoints. No access to other devices on the home network.
  • CPU, GPU, memory and temperature caps set by the device owner and enforced outside the sandbox.

Limits

  • Giving a sandbox direct access to a GPU widens the attack surface through the graphics driver.
  • We intend to complete and publish an independent security review before any third-party code runs on a contributor's device.

Question 02

What stops malicious or unreliable nodes?

PlannedEstablished techniques from volunteer computing. Not built here.

We assume some devices will cheat or fail. The design pays only for results that survive checking.

Approach

  • Replication: a share of tasks would be sent to more than one independent device and the results compared.
  • Known-answer tasks mixed into the queue to catch devices that return plausible garbage.
  • Reputation earned from history. New devices start on low-value work.
  • Hardware claims would be checked against measured throughput, so a device cannot simply declare itself a flagship GPU.

Limits

  • Colluding devices that return the same wrong answer are a known hard problem in volunteer computing.
  • Every check costs compute. Redundancy has to be paid for out of the buyer's price or the platform fee.

Question 03

How are results verified?

ConceptualThe approach for non-deterministic work is an open question.

It depends on the workload. Deterministic work could be compared exactly. Everything else needs tolerance checks or trusted re-execution.

Approach

  • Deterministic tasks: compare result hashes across replicas.
  • Numerical tasks: compare within a tolerance, with fixed seeds where the workload allows.
  • Generative tasks: sampled re-execution on hardware we control.
  • Payment would be released only for verified units of work.

Limits

  • Different GPUs do not produce bit-identical floating-point results.
  • Some workloads have no cheap check. We would rather decline them than pretend.

Question 04

What happens when a device disconnects?

PlannedDesign only.

Disconnection is the normal case. Tasks would be small and restartable, so a lost device should cost little. That is unmeasured.

Approach

  • Work would be cut into short, idempotent tasks.
  • Devices would send heartbeats. A missed heartbeat would return the task to the queue.
  • Longer tasks would checkpoint, so a replacement device resumes rather than restarts.
  • The owner using their device always wins: Jetpack would pause immediately.

Limits

  • Work that cannot be cut into restartable pieces is not a fit.

Question 05

How are failed jobs rescheduled?

PlannedDesign only.

The scheduler would reassign a failed task to another device and track the job, not the device.

Approach

  • Automatic retry on a different device, with a cap on attempts.
  • Hedging: near a deadline, slow tasks would be duplicated and the first verified result wins.
  • Job-level progress and partial results visible to the buyer.

Limits

  • Retries create a long tail. The last few per cent of a job can take disproportionately long.
  • No completion-time guarantees until pilots have measured real behaviour.

Question 06

How are confidential inputs protected?

ConceptualNo credible general answer exists for consumer hardware today.

Honestly: not well enough yet. Early workloads should use data you could afford to have seen.

Approach

  • Encryption in transit and at rest inside the sandbox.
  • Jobs split so that no single device holds the whole dataset.
  • Inputs and outputs deleted from the device when a task ends.

Limits

  • The device owner controls the hardware. A determined owner could inspect what runs on it.
  • Confidential-computing hardware is not generally available on consumer GPUs.
  • Splitting data reduces exposure. It is not confidentiality.

Question 07

How is performance normalised across different hardware?

PlannedDesign only. We have no benchmarks to publish.

Buyers would pay for verified work delivered, measured in normalised units, not for hours a device claims.

Approach

  • A benchmark on enrolment and at intervals, per workload family.
  • Scheduling by measured capability: VRAM, throughput, bandwidth, availability history.
  • Benchmarks cross-checked against real task throughput.

Limits

  • Thermal throttling and background use make a home device slower than its benchmark. Normalisation has to be continuous.

Question 08

How are geographic restrictions enforced?

PlannedDesign only.

Buyers could restrict jobs to chosen countries. Locating a home device is an estimate, not proof.

Approach

  • Location inferred from several signals: network address, payout country, latency to known points.
  • Devices with conflicting signals are excluded from restricted jobs.

Limits

  • A network address is not proof of location.
  • This should not be relied on to meet data-residency law.

Question 09

How do pricing and settlement work?

ConceptualCommercial terms are not set.

Buyers would pay in ordinary currency for verified work. Contributors would be paid from that, less a platform fee. There is no token.

Approach

  • Pilot pricing agreed directly with each buyer, per unit of verified work.
  • Contributors paid in their local currency. No cryptocurrency, no token, no staking.
  • The 15% platform fee in our calculator is a working assumption, not a decision.

Limits

  • There is no market price yet because there is no market.
  • Payout provider, minimum payout, tax reporting and the countries we can pay in are undecided.

Question 10

What should never run on this network?

PlannedPolicy position. Enforcement is not built.

Anything sensitive, urgent, stateful or abusive.

Approach

  • Regulated or personal data: health records, payment data, anything covered by privacy law.
  • Secrets, credentials or unreleased proprietary models.
  • Latency-critical or safety-critical services.
  • Databases and anything that must stay up.
  • Training that needs many GPUs acting as one.
  • Routing internet traffic through contributors' home connections. We will not sell residential proxying.
  • Cryptocurrency mining, password cracking, or anything unlawful.

Limits

  • An acceptable-use policy and the means to enforce it do not exist yet. We intend to have both before the first pilot.

Disagree with an answer, or need one we have not given?

Send us the question with your workload. The hard ones are the useful ones.

Tell us