Does data-centric computing mean processing everything inside memory?

No. Processing-in-memory is one related technique; the broader approach includes placement, reuse and movement-aware software and hardware design.

Is moving computation always better than moving data?

No. Compute capability, data reuse, transfer costs and coordination determine the tradeoff.

Is it the same as data-centric AI?

No. The terms emphasize different issues: system organization here, and typically dataset improvement in data-centric AI.

AI Glossary / Data engineering and architecture / Intermediate

Data-centric computing

Definition

Data-centric computing is an approach to computer-system design that treats the location, movement and access patterns of data as central concerns. It seeks to organize computation and resources around those concerns, often reducing unnecessary transfers or placing suitable work closer to the data. The term covers a family of approaches rather than one specific architecture.

What it means

A processor cannot perform useful work on information it cannot obtain quickly enough. In some workloads, moving data through memory, storage or networks consumes substantial time and energy relative to the computation itself. Adding arithmetic capacity alone may therefore deliver little improvement.

Data-centric computing asks where the information is, how often it is reused and which operations should occur at which location. It is distinct from data-centric AI, which commonly emphasizes improving the datasets used to develop models. The two can coexist, but one concerns computing-system organization while the other concerns a model-development emphasis.

How it works

Engineers examine the path between data and computation. They can reduce transfers through filtering, aggregation or reuse; improve placement and scheduling; or introduce processing capabilities near memory or storage. The appropriate technique depends on the operation and the system.

Moving computation closer to data is one option, not an absolute rule. A nearby processor may be less capable, while a central accelerator may process a sufficiently demanding task much faster after transfer. Reuse, synchronization and shared access also affect the decision.

The Livia research project illustrates this nuance within a memory hierarchy. Its proposed architecture schedules tasks and data at suitable locations rather than assuming all computation belongs at a single fixed point. More broadly, data-centric design combines hardware and software choices to reduce avoidable movement while preserving correct execution. It still requires computation, communication and resource management.

Three practical examples

1. Large-data filtering: an analytical system applies an appropriate filter near stored data and transfers only matching records to a later stage. Engineers check that the filter preserves the intended result.

2. Distributed AI training: a team studies how data placement, worker scheduling and caching affect repeated dataset access. Its goal is to reduce waiting and unnecessary transfers while maintaining consistent input versions.

3. Memory-intensive analysis: researchers explore processing near the memory holding a large graph. They evaluate irregular access patterns and coordination costs rather than assuming faster arithmetic will solve the bottleneck.

Think of it this way

Think of inspecting goods in a warehouse before shipping them to a workshop. If most items will be rejected by a simple inspection, checking them locally can avoid transporting unnecessary material. For a specialized repair, however, moving an item to the workshop may still be sensible.

The analogy captures placement tradeoffs. Computer systems add cache reuse, concurrency and precise consistency requirements; physical proximity alone does not determine the best execution location.

Worked example

Suppose a dataset occupies 1 terabyte, or 1,000 gigabytes in decimal units. Transferring it over a link that sustains 1 gigabyte per second takes approximately 1,000 seconds, excluding overhead.

Assume a source-side filter takes 200 seconds and retains 1% of the data, or 10 gigabytes. Transferring the filtered result takes approximately 10 seconds. In a simplified sequential comparison, filtering and then transferring take 210 seconds rather than 1,000 seconds of raw transfer, a difference of 790 seconds.

This is not a universal speedup. It excludes destination processing, overlap between stages and resource contention. If the filter retains almost everything or runs very slowly at the source, the advantage can disappear. The example shows how selectivity and placement affect movement cost; it does not claim a measured system result.

Illustrative case study

Fictional case: NearField Analytics repeatedly downloads a large event collection to calculate a few summary counts. Its team initially plans a faster central processor, then measures the workflow and finds that transfer dominates elapsed time.

The team tests computing the summaries where the data already reside, sending the smaller results onward. It verifies that grouping rules, missing-value treatment and access controls remain correct. For a different workload requiring detailed records, it keeps the original transfer path because summaries would discard needed information.

The case illustrates workload-specific placement. NearField Analytics is invented, and the example does not imply every system supports arbitrary computation inside its storage service.

Common misconceptions and mistakes

Do not equate data-centric computing with eliminating data movement entirely. Useful systems still communicate; the aim is to make movement appropriate to the workload.

Do not assume processing-in-memory, near-data processing and every other data-centric approach are identical. They describe related but distinct design choices.

Do not infer improved model accuracy from a more efficient architecture. The architecture may change cost or performance while leaving the algorithm and predictions unchanged, unless other numerical or implementation changes are introduced.

Business use

The approach is relevant when measurements show that data access or movement limits a valuable workload. It can inform placement, storage organization, caching and hardware selection, but specialized solutions can also introduce programming and operational complexity.

A business evaluation should compare complete workflows, including preprocessing, transfer, execution and recovery. Use representative data sizes and access patterns. Record whether reduced movement changes data fidelity or available detail, and separate performance claims from energy claims unless both are measured. A faster benchmark on one workload does not establish a benefit for every analytical task.

Check your understanding

Which two assumptions drive the apparent benefit in the worked example?

Answer: The local filter finishes in 200 seconds and reduces the transferred data to 1% of the original. Different processing cost or selectivity can change the conclusion.

Related terms

Distributed computing
Data locality
near-data processing
processing-in-memory
memory hierarchy
bandwidth
caching
computational storage
hardware-software co-design
data movement

Learn next

Learn about data locality and memory hierarchies next. Break one workload into time spent obtaining data and time spent computing. Identify which transfers are necessary, which repeat avoidably and whether an earlier operation could reduce the amount moved without changing the required result.

Sources

Lockerman and colleagues, Livia: Data-Centric Computing Throughout the Memory Hierarchy, primary research: https://people.csail.mit.edu/sanchez/papers/2020.livia.asplos.pdf

IEEE Micro special-issue scope, published by ACM SIGARCH: https://www.sigarch.org/call-contributions/ieee-micro-special-issue-on-data-centric-computing/

Primary research and field-scope material consulted. The article describes a broad architectural emphasis; calculations and the fictional case are original illustrations rather than published performance results.

Previous: Data catalog Next: Data-Centric ML