Compute
Operations, execution resources, and the useful work performed during inference.
Arithmetic workDepends on representation & scheduling
The computational frontier
ENGINEERING
THE NEXT ERA.
Frontier research in AI infrastructure and computational efficiency.
We investigate how intelligence is computed—and how its foundations can make it more efficient, affordable, and accessible.
A study of computational dependenciesTransformer execution / schematic
- Input sequence
- Embedding & Q / K / V
- Causal attention & KV cache
- Feed-forward & token scores
Intelligence has
a computational cost.
That cost is not immutable.
Capability alone does not make intelligence accessible. The resources needed to execute it determine who can use it—and at what cost.
Memory capacity, bandwidth, computational throughput, execution scheduling, and hardware constraints shape the economics of intelligent systems. We treat these constraints as research opportunities.
Forera is an independent research-driven engineering venture. We study how computational work is organized, measured, and executed, seeking deeper understanding that can inform practical improvements in AI infrastructure.
Our ambition is to change the computational economics of intelligent systems. It is a direction for investigation, not a claim of a breakthrough.
Efficiency is a
systems problem.
Six dimensions. One interdependent system.
An improvement in one place can move the cost elsewhere.
Memory
The space for model state, and the rate at which that state can be accessed.
Capacity & bandwidthConstrains concurrency & movement
Movement
Transfers between storage, memory, and compute can dominate execution time.
Data across boundariesCouples memory to compute
Scheduling
How work is ordered, overlapped, and distributed across available resources.
Order & concurrencyTrades throughput for memory pressure
Representation
The numerical formats and structures used for weights and intermediate states.
Structure & precisionChanges storage, arithmetic & quality
Measurement
Instrumentation that separates a real bottleneck from an attractive assumption.
Evidence & interpretationEvaluates every other dimension
Precision is a tradeoff. Lower numerical precision can reduce memory demand while changing accuracy or hardware efficiency.
Concurrency has a cost. Overlapping more work can improve throughput while increasing memory pressure or latency.
Questions worth
computing.
Our current research interests.
Open directions, investigated through experiments and engineering.
01
Inference architecture
Investigating how model execution is structured, and where alternative architectures may change performance, resource needs, and scale.
How should model execution be organized?
02
Computational efficiency
Studying useful computational work in relation to hardware utilization, memory behavior, and the cost of execution.
Where does useful work become overhead?
03
Experimental systems
Developing instrumentation, profiling methods, and reproducible environments for controlled computational experiments.
What evidence makes a comparison meaningful?
04
Intelligent orchestration
Exploring infrastructure for composing, coordinating, and observing intelligent computational workflows.
How do complex processes remain accountable?
05
Resource-aware computing
Investigating how systems adapt to hardware constraints and make better use of the resources already available.
What can intelligence do within a budget?
06
Emerging computational techniques
Exploring hypotheses and architectural alternatives that challenge assumptions in established AI execution systems.
Which assumptions deserve another experiment?
Measure first.
Understand deeply.
Engineer deliberately.
Research proceeds in loops, not declarations. Every explanation is open to revision by the next experiment.
Clear baselines. Disclosed hardware and software configurations. Controlled variables. Correctness and quality checks. Reproducible comparisons. Transparent limitations.
- 1
Observe
Instrument behavior. Identify what can be measured.
- 2
Model
Explain mechanisms. Form a testable hypothesis.
- 3
Experiment
Isolate variables. Compare controlled alternatives.
- 4
Validate
Check correctness, quality, and reproducibility.
- 5
Engineer
Build reusable infrastructure from validated findings.
- 6
Iterate
Let new measurements revise the explanation.
New measurements return to observation.
Research,
taking shape.
The initial Forera ecosystem.
Experimental projects that connect questions to working systems.
Project / 01
Cachalot
ExperimentalComputational research / AI inference
An experimental project investigating how large language models execute. Cachalot examines inference mechanisms, resource consumption, and opportunities for alternative execution strategies.
- Model execution
- Memory behavior
- Architecture experiments
Project / 02
Cachalot Lab
Research & developmentInstrumentation / Experimental systems
The macOS observability and experimentation environment associated with Cachalot. Lab makes runtime activity visible; its research direction includes deeper profiling, structured comparisons, and reproducible experiments.
- Runtime instrumentation
- Profiling visualization
- Experimental comparison
Project / 03
Crewshal
Under developmentIntelligent systems / Agent orchestration
Crewshal explores reusable, project-independent infrastructure for coordinating intelligent workflows and computational agents. The focus is modular execution, adaptable architecture, and reliable coordination.
- Agent orchestration
- Explicit execution models
- Execution observability
Anatomy of
computational cost.
Faster compute is only useful when data can keep up. Explore the relationship between arithmetic work and memory bandwidth.
Illustrative computational model
P ≤ min(Ppeak, B × I)Change the workload.
Watch the constraint.
Operations per byte transferred from memory.
An illustrative hardware resource, held independent of compute.
Performance upper bound400 GFLOP/s
Active constraintMemory bandwidth
More arithmetic per byte or more memory bandwidth raises this bound. The compute ceiling still applies.
A simplified upper bound, not a benchmark or a prediction for Cachalot. Real performance also depends on latency, scheduling, cache behavior, numerical formats, and correctness. Units use decimal GB and GFLOP.
Read the roofline.
Before the ridge, the memory system limits the bound: bandwidth multiplied by operations per byte. Beyond it, peak compute throughput sets the ceiling. Raising one resource does not remove every constraint.
Roofline model · Berkeley LabHow we choose
to engineer.
Rigor is a practice.
Not an aesthetic.
Understanding before optimization
Investigate the mechanism before changing its behavior.
Measurement over assumption
Claims need reproducible experiments and clearly defined baselines.
Architecture matters
How work is organized matters as much as the speed of individual operations.
Efficiency with integrity
Evaluate performance alongside correctness, quality, reliability, and practical constraints.
Research with practical consequences
Seek knowledge that can inform real systems and usable infrastructure.
Independent thinking
Established techniques are foundations for investigation, not its boundaries.
From computational understanding
to practical infrastructure.
Our research starts with questions about computational behavior and architectural efficiency. Promising findings may become experimental implementations, open-source tools, reusable libraries, infrastructure components, or developer platforms.
Commercial technologies and technical collaborations can put those ideas to work—and expose new questions for research. Scientific investigation and practical engineering reinforce one another.
- Research questions
- Experimental models
- Validated findings
- Engineering prototypes
- Reusable technology
- Practical applications
A conceptual path, not a record of completed stages. Progress depends on evidence.
For researchers, engineers, and organizations.
Explore the computational
frontier with us.
We welcome meaningful technical conversations about advancing the efficiency and accessibility of intelligent computation.
Start a conversation