Architecture & Engine
From existing code to parallel placement.
A plain technical view of how ORIQX connects beneath your stack, partitions workloads across five steps, and respects clear operational boundaries.
The problem
Hardware capability climbs. Capacity sits idle.
Static placement pins mixed workloads onto one processor type. Data moves unnecessarily between nodes while expensive accelerators stall waiting on sequential code—and AI agent loops multiply that waste.
Fixed placement waste
Mixed math trapped on one-size hardware
A single job mixes sequential logic, sparse solves, memory transfers, and parallel matrix math. Running all of it on one chip class leaves processors underused and stretches wall time.
The agentic multiplier
Short prompts trigger hidden heavy compute
Autonomous agents launch retries, parallel sweeps, and heavy tool calls beneath the surface. Without pre-run pricing and placement, compute bills climb unseen.
Stack fit
Sits beneath your existing stack.
ORIQX is a frictionless solution—no code rewrites are required.
- Application code, job submission paths, and access boundaries stay as they are.
- Reach CPUs, GPUs, TPUs, NPUs, and QPUs across data centres, HPC clusters, and cloud accounts.
- Works underneath existing orchestrators, queues, and agent frameworks without replacing them.
Your code & workloads
Keep solvers, scripts, and applications intact
Queues & agent frameworks
Existing orchestrators keep control of approvals
ORIQX compute engine
Partitions blocks and chooses chip placement
Any reachable compute pool
Fully managed or self-managed across owned clusters and clouds
Options scored for time, cost, energy, and accuracy before dispatch.
CPU, GPU, TPU, NPU, and QPU execution
How it works
Five engine steps from code to chips.
Rather than running an entire job monolithically on one processor class, ORIQX breaks programs into compute blocks, predicts each block per chip, and searches for the split that meets your goal.
Find compute blocks & data
Reads the program and finds its compute blocks, the kinds of math it does (dense linear algebra, sparse solves, FFTs, attention, graph search), and how data moves between them.
Predict per chip class
Predicts each block's time, cost, energy, and accuracy on each chip class it can use.
Search the split
Searches for the split that best meets your goal across all blocks, counting the cost of moving data between chips.
Prepare & run in parallel
Prepares each block for its chip and runs blocks in parallel across your infrastructure.
Measure & sharpen
Measures what actually happened and uses it to sharpen the next prediction.
Compute language & zero-rewrite intake
Bring code as it is—into the ORIQX compute language
The ORIQX compute language describes what to compute rather than where to run it—allowing the engine to choose and combine chip classes freely.
Existing programs written in Python, C++, Rust, Julia, Go, and more enter as they are without rewriting code around specific hardware targets.
Supported hardware
CPUs, GPUs, TPUs, NPUs, and QPUs—plus more as needed
Workloads run on CPUs, GPUs, TPUs, NPUs, and QPUs. ORIQX adds new chip classes as customers need them.
Explicit boundaries
Five plain questions for technical teams.
Every deployment agrees service placement, approval authority, and infrastructure ownership upfront.
What changes in application code?
Nothing. Existing programs and pipelines stay unchanged; ORIQX operates beneath your workflow entrypoints.
Where do workloads execute?
Either on compute ORIQX provides and runs (fully managed) or inside your data centre, HPC cluster, or cloud accounts (self-managed).
Who approves compute spend before execution?
Preflight presents predicted time, cost, energy, and accuracy options first—nothing runs or spends until you or your policy goal approve.
Does Estimate touch or run on unowned hardware?
No. Estimate scores program work against ORIQX chip-class performance profiles with zero candidate hardware access and zero execution.
Which chip classes are supported?
CPUs, GPUs, TPUs, NPUs, and QPUs. ORIQX adds new chip classes as customers need them.
