Research Engineer, Computer Architecture
Zurich, CH · hybrid
About ORIQX
Compute is fragmenting. For a decade the answer was one CPU and one class of GPU. It is now many: further GPU classes, gate quantum processors, annealers, analog and thermodynamic samplers, wafer-scale parts and custom accelerators, each an order of magnitude better or worse depending on the problem, and each with a different cost, energy and error profile. The decision of what to run where is currently made once, by hand, and is wrong as soon as prices, hardware or problem size change.
ORIQX is three things: a model, an engine and hardware. The ORIQX compute model is being built, starting now: a model that works in the vocabulary of a domain and states what to compute, never where. The ORIQX compute engine is live today: before anything executes, it predicts what a computation will cost in time, money, accuracy and energy on every class of chip we model, runs each part where it wins, and returns a report of predicted against actual. ORIQX hardware is being simulated now, and its blueprint follows in twelve months: the design of a heterogeneous rack built from other vendors' silicon, sized to a real portfolio of workloads.
The engine sits beneath the orchestrators rather than among them, and all three are meant for every vendor's silicon rather than one.
About the role
This seat owns the grading discipline underneath every hardware figure ORIQX publishes, and the inverse problem that turns a portfolio of workloads into a data-centre build specification: the missing rung between an architectural estimate and a measurement, and a first blueprint that exists today as a written formulation and not yet as software.
ORIQX models a broad catalogue of chip classes and predicts, before anything runs, what a computation will cost in time, money, accuracy and energy on each of them. Two things sit under that and both belong to this seat. Provenance is the rule by which every modelled parameter declares where it came from and what may be claimed of it, enforced in tests rather than in guidance: a figure that has not been measured may not be described as one, and a class whose parameters would have to be invented is declined a price rather than given a plausible one. The inverse problem is the costing machinery run backwards, to answer what fleet a portfolio of workloads warrants; it is set out as an optimisation over candidate machine mixes in the company's technical writing, is offered commercially as a licensable design deliverable, and has not been built. The appointee writes it as software, authors the first specification from simulated evidence at architectural provenance, and owns the rule by which any figure in it is ever allowed to be called a measurement, including what a cycle-level result produced on hardware emulating silicon nobody has built is permitted to be called. The founder sets architecture direction and reviews the appointee's pull requests; no one stands between the engineer and the work.
Responsibilities
- Naming, defending and landing the grade of provenance that does not exist. A cycle-level result for a device that has not been built is not an architectural estimate from a parameter vector and is not a measurement on silicon, and the grading in use has no term for it. The weaker grades are also not separated at the point of use, so a dated datasheet figure and an uncited assumption reach the decision that consumes them as the same claim. The term wants deciding in writing before the first such run rather than after it, and the deliverable is a type change, a promotion rule, a test that fails when the rule is broken, and an argument written down where the next engineer can disagree with it.
- Closing the distance between the graded confidence the company publishes and the grading its software operates. The published account describes an architectural estimate tightening through a corrected middle tier as related devices contribute telemetry, and converging on measurement. That middle tier is described more fully than it is produced: it has a representation and a documented band, and nothing outside its own tests puts a figure into it. Either the middle tier earns a producer or the published description of it changes, and this seat decides which, because a public claim about how ORIQX grades its own predictions should not be settled by whoever edits the document next.
- Owning the boundary at which a measurement is allowed to become modelled physics for an entire device class. Calibration runs against a reference configuration projected from the modelled parameters, and only devices for which real physics has been measured may displace those figures; every other difference is rejected deliberately, so that a modelled number cannot be smuggled across the boundary wearing a measurement's clothes. Widening the set of devices admitted is not an edit to a list. It is an instrumented campaign for every class admitted, and then the argument about what the campaign entitles anyone to write down. Other seats run the campaigns and produce the readings; this seat writes the rule those readings have to satisfy, and is the one that says no.
- Writing the first blueprint as software rather than as prose. The published formulation is a minimisation over candidate machine mixes under capital, power and floor constraints, with the per-candidate evaluation delegated to costing machinery that is in production, and it states its own defect plainly: the inner optima are independent per job, so nothing in the objective prevents the whole portfolio from claiming the same best device, and the recommended composition degenerates towards a single unit of each per-job winner. The independent form screens classes; only a form carrying concurrency commitments sizes them. The screening form is not yet software and the sizing form does not exist at all, and the work has so far been written up rather than built.
- Extending the grading to measurements that are real but are not ours. Static draw as a fraction of board power is modelled coarsely, set conservatively rather than interpolated, with per-class figures for some classes taken from a peer-reviewed measurement of the parts those classes model. Those are filed as estimates, on the stated ground that the measurement is the paper's, on its silicon, and not ours on ours. That is the right answer under the grading as it stands, and it is also an admission that the grading cannot express a third party's measurement at all, which is the same missing-rung problem arriving from the other direction. Energy is one of the quantities ORIQX predicts before a run, so this is not bookkeeping.
- Establishing what a cycle-level model is and is not accurate with respect to, before anything built on one is published. The strongest published result on the question is that cycle-exact models of real machines landed well below the hardware they modelled on memory microbenchmarks, because the parameters they needed, cache latencies, translation buffer sizes, load-store queue depths, bank counts, were never published; an enormous gain in simulation speed bought a model that was wrong by a multiple, and every correction was software-side calibration against real silicon. A cycle-level run on an emulated rack therefore cannot be trusted merely because it is cycle-level. The method has to be validated where a real machine exists before it is relied upon where none does, which is a measurement campaign and a written stopping rule, not an assertion.
You may be a good fit if you
- The successful candidate will have produced a hardware specification, reference design or capacity plan that another party procured against, and can state which of its figures were measured, which were extrapolated, and which part of it later proved wrong.
- The engineer has built or driven a cycle-level or register-transfer-level model of a system, with Verilator or comparable tooling, and can say precisely what such a model is accurate with respect to, which is rarely the silicon, and what has to be calibrated against real hardware before it can be believed.
- The candidate thinks in roofline terms and is exact about them: peak per precision lane, sustained bandwidth, the split between static and dynamic power, and what one measurement at one shape entitles anybody to claim about an entire device class.
- Rust is read and written comfortably, alongside working familiarity with at least one of C++, Python or Go, and the candidate is at ease where the deliverable is a parameter table and the argument recorded beside it rather than a new abstraction.
- The engineer understands a rack as a system rather than as a parts list: the coupling fabric and what it costs per handoff, the power and thermal envelope, the difference between a placement problem and a capacity problem, and why a portfolio of jobs each choosing its own best device does not add up to a fleet.
- The candidate has designed or run a measurement campaign on real machines and treats the infrastructure as a source of error in its own right, having at least once found that the defect lay in the harness and not in the model.
- Comments are written to carry the argument, the counter-argument and the incident that produced the fence. On this surface a claim without its provenance written beside it is the defect, and terse code with thin comments reads as a regression here regardless of its quality.
- Coding agents are used as a matter of course and what they produce is verified; that is how this platform was built and how it will continue to be built.
Strong candidates may also
- Experience of a design or reference-architecture deliverable that a third party licensed, together with an account of what the licensee needed in the pack beyond the specification itself.
- Familiarity with embodied and operational carbon accounting at rack scale, including where the standard methodologies stop being defensible.
- A published or refereed validation of a performance model against hardware, particularly one that reported the regime in which the model was worst rather than a single aggregate figure.
- Reading knowledge of the discrete-event simulator literature on fidelity tiers, and working familiarity with the open accelerator toolchains around it, such as MLIR, TVM, Triton or Pallas, sufficient to argue about what a modelled device would have to support for real code to approach its peak and which external reference is worth standing up beside ORIQX's own.
Logistics
- Education
- A degree in computer science, electrical engineering, physics or mathematics is one way to be ready for this work. Equivalent experience is another, and the two are weighted equally. The assessment is what the candidate has built.
- Location
- Based in Zurich, or hybrid. Relocation support.
Apply even if you do not meet every line above. Few strong candidates meet all of them, and the ones who exclude themselves early are disproportionately the people we would want to have met.
Come work with us
- What you can do
- We care about what you can do, not where you learned to do it.
- Compensation
- Every seat has a real band, and it is given on the first call: competitive for a company at this stage, and not matched to a frontier laboratory's cash.
- Equity
- Equity is part of every offer. The instrument, the vesting and the Swiss tax treatment are settled in the offer letter, with the candidate. No range is published, because none is agreed for a seat until then.
- Where
- Based in Zurich, or hybrid. Relocation support.