AI Glossary / AI and machine learning / Intermediate

SUPS

Definition

SUPS stands for synaptic updates per second, a throughput measure used in neural simulation and some neural computing contexts. It counts how many operations defined as synaptic updates are performed in a second, so its meaning depends on the model, implementation, and measurement convention.

What it means

A simulated synapse represents a connection through which one modeled neuron influences another. Processing activity at that connection can require updating a state, delivering an effect, or applying a learning rule, depending on the system.

SUPS expresses the rate of those defined updates. It is not a general measure of intelligence, task accuracy, or biological realism. Two systems can report the same rate while doing substantially different work per update.

The denominator also matters. A report may discuss updates required per second of simulated biological time or updates achieved per second of actual elapsed computing time. Those quantities should not be confused.

How it works

First define what counts as one update. Identify whether the benchmark includes event handling, learning-related changes, communication, and other relevant work.

Next count the qualifying operations over a measured interval and divide by elapsed time. For actual throughput, use wall-clock seconds and clearly state what initialization or data movement is included.

In a time-driven simulation, updates may occur at fixed intervals. In an event-driven design, activity can depend on incoming events. Network size, connectivity, firing activity, and model complexity can therefore change both the required work and measured performance.

Report the workload alongside the number. Useful context includes simulated duration, model details, precision, hardware, power measurement boundaries, and whether the result is a sustained rate or a theoretical peak.

Three practical examples

1. Simulator benchmarking: Researchers compare two implementations of the same neural model using an identical event trace and update definition.

2. Capacity planning: A laboratory estimates whether available computing resources can process its chosen network activity close to simulated real time.

3. Hardware evaluation: Engineers compare energy use and achieved update throughput for a controlled workload before selecting a platform for a prototype.

Think of it this way

Think of counting parcels processed per second in a sorting facility. The number is meaningful only if “processed” means the same thing and parcels require comparable work.

The analogy is limited because simulated synaptic updates can vary in mathematical complexity, memory access, and communication. A higher count can reflect simpler operations rather than a more capable system. Throughput must be interpreted alongside what each update accomplishes.

Worked example

Suppose a benchmark counts 120 million qualifying synaptic updates while the simulation runs for 3 wall-clock seconds. Its achieved throughput is 120 million / 3 = 40 million SUPS.

If those updates represent 2 seconds of simulated activity, the workload requires an average of 60 million updates per simulated second. Completing it in 3 real seconds means the simulation runs at two-thirds of real time, or takes 1.5 real seconds per simulated second.

Now suppose the measured system consumes an average of 20 watts over the same interval. Under that measurement boundary, 40 million SUPS / 20 watts = 2 million updates per joule.

These figures say nothing by themselves about model accuracy or whether another benchmark's update definition is comparable. They also exclude any energy outside the stated measurement boundary.

Illustrative case study

Fictional case: Pinecrest Simulation Lab receives two performance reports for neural computing platforms. One lists a higher SUPS value, but the reports use different synapse models and activity rates.

The team runs a common workload and records elapsed time, counted updates, and power under a shared measurement procedure. It separately documents setup costs and communication overhead.

The resulting comparison is more useful for the lab's task than the headline rates. Researchers still evaluate numerical behavior and scientific suitability before choosing a platform. This fictional example illustrates benchmark design; it does not claim that one real product outperforms another.

Common misconceptions and mistakes

SUPS is not interchangeable with floating-point operations per second. One synaptic update may involve multiple operations and memory accesses, with no universal conversion factor.

An update does not necessarily mean a learned weight changed. Some conventions count processing or state updates associated with synaptic activity.

High throughput does not demonstrate human-like intelligence or a faithful brain simulation.

Do not compare simulated-time demand with wall-clock throughput as though both were achieved computing rates. Likewise, distinguish peak specifications from sustained application measurements.

Business use

SUPS is most useful when evaluating a well-defined neural simulation or neuromorphic workload. Procurement teams should ask for the exact counting convention and a reproducible benchmark matching their intended application.

Consider latency, memory capacity, communication, energy, software support, and model fidelity alongside throughput. An implementation that processes fewer but more detailed updates may be the right choice for a scientific question. For business applications, task performance and operating cost remain more directly relevant than a standalone synaptic update rate.

People also ask

Does one SUPS equal one neural network prediction per second?

No. A prediction can involve many updates, and the relationship depends on the model and implementation.

Can SUPS measure learning speed?

It can describe defined update throughput during learning, but it does not directly measure how quickly useful behavior is learned.

Why specify the firing rate or event activity?

Event-driven workloads can require very different amounts of work as activity changes, affecting both demand and observed performance.

Check your understanding

A system processes 90 million defined updates in 6 wall-clock seconds. What is its throughput, and what extra information is needed for a fair comparison?

Answer: Its throughput is 15 million SUPS. Comparison requires at least a matching update definition and workload, plus relevant model and measurement details.

Related terms

• Latency
• Neuromorphic computing
• spiking neural network
• synapse
• event-driven simulation
• throughput
• FLOPS
• energy efficiency

Learn next

Learn the difference between time-driven and event-driven neural simulation. Then practice separating simulated time from elapsed time in a benchmark report. Compare update throughput and energy only after writing down precisely which operations and system components are included.

More in AI and machine learning

• Supervised Clustering - Supervised clustering groups objects using guidance from known labels.
• Supervised learning - Supervised learning trains a model using examples that pair inputs with known target outputs.
• Support vector machine - A support vector machine, or SVM, is a learning method that can build a decision boundary between classes while balancing separation against errors.
• Surrogate model - A surrogate model is a simpler or cheaper model used to approximate the behavior of another system, model, experiment, or computational process.
• Switching Hidden Markov Model - A switching hidden Markov model describes a sequence by switching among the outputs of several hidden Markov models.
• Symbol level - The symbol level is a way of describing a computing system in terms of its representations and the operations that manipulate them.

Sources

Dharmendra S. Modha, “FAQs and Summary: Rat-scale, near Real-time Cortical Simulations”: https://modha.org/2007/11/faqs-and-summary-rat-scale-near-real-time-cortical-simulations/

Independent review read the clock-driven neuron/event-driven synapse description, computation workload discussion, and model-time versus elapsed-time example. All benchmark numbers and the fictional laboratory case here are original teaching examples, not measurements reproduced from that research.

Previous: Supremum (or least upper bound) Next: Surely expanded nodes