Why Science? | National Supercomputing Centre Singapore

Official NSCC sources describe a National Research Infrastructure supporting research programmes, institutions, agencies and companies. It provides computing platforms, expertise and partnerships rather than scientific conclusions by itself.

High-performance computing earns its value by shortening a scientifically meaningful calculation or enabling a scale that smaller systems cannot manage. A larger machine does not repair an unclear model or weak data. The practical capability is institutional literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “NSCC is Singapore’s national supercomputing infrastructure” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Access to a national supercomputer does not validate a model or result.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Next chapter

HPC is useful when resolution, ensemble size, data volume or model complexity exceed practical smaller systems. Teams should define the scientific output and decision before requesting resources. A simpler model or experiment may sometimes answer the question better.

Parallel computing divides work across processors while communication, memory and storage impose limits. Performance depends on the algorithm and data movement as much as on the advertised number of cores or accelerators. The practical capability is problem framing. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Begin with a computationally meaningful question” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A large resource request is not evidence of scientific importance.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Data or tasks can be distributed across processors, nodes or accelerators. Dependencies require communication and synchronisation. Good decomposition balances work while minimising movement and idle time. Some algorithms expose more parallelism than others.

A supercomputer is a shared research instrument. Schedulers, allocations, accounting and support services help many teams use expensive national resources responsibly and reproducibly. The practical capability is parallel-computing reasoning. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Parallelism divides work with dependencies” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: An algorithm that is fast on one core may scale poorly across many nodes.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

CPUs provide flexible cores and large memory systems for varied control-heavy tasks. GPUs offer many parallel units suited to regular numerical work and AI operations. Performance depends on code structure, precision, memory access and libraries.

Computational evidence needs provenance: code, environment, inputs, parameters, hardware, random seeds, logs and analysis. Without that record, speed can produce answers that are difficult to inspect or repeat. The practical capability is hardware literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “CPUs and GPUs serve different workload patterns” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A GPU is not automatically faster for every scientific program.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Registers, caches, device memory and main memory differ in size and speed. Data layout and access patterns affect how well computation uses the hardware. Exceeding available memory can force communication or storage traffic that dominates runtime.

High-performance computing earns its value by shortening a scientifically meaningful calculation or enabling a scale that smaller systems cannot manage. A larger machine does not repair an unclear model or weak data. The practical capability is memory-performance reasoning. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Memory hierarchy shapes performance” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Peak arithmetic performance does not describe memory-limited workloads.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Nodes exchange messages over high-speed networks. Latency, bandwidth and communication pattern affect strong and weak scaling. Collective operations can become bottlenecks as node count rises. Network topology should inform decomposition.

Parallel computing divides work across processors while communication, memory and storage impose limits. Performance depends on the algorithm and data movement as much as on the advertised number of cores or accelerators. The practical capability is distributed-systems literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Interconnects make a cluster one system” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Many fast nodes do not guarantee fast coordination.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Simulations and AI workflows can create large inputs, checkpoints and outputs. Parallel filesystems, metadata operations and transfer paths have limits. Retention, compression and analysis strategy should be planned so storage does not become an unexamined cost.

A supercomputer is a shared research instrument. Schedulers, allocations, accounting and support services help many teams use expensive national resources responsibly and reproducibly. The practical capability is research-data planning. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Storage and data movement need their own plan” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Writing every intermediate value may slow work without improving evidence.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Compilers, libraries, drivers, containers and modules influence numerical behaviour and performance. Environment records and versioned workflows support reruns. Portable code may still require hardware-specific tuning.

Computational evidence needs provenance: code, environment, inputs, parameters, hardware, random seeds, logs and analysis. Without that record, speed can produce answers that are difficult to inspect or repeat. The practical capability is computational reproducibility. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Software environments preserve dependencies” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A container alone does not capture external data, scheduler or hardware context.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Users submit jobs with requested time, processors, accelerators and memory. The scheduler balances policy, priority and efficient packing. Accurate requests reduce queue delays and wasted allocations. Checkpointing protects long jobs from interruption.

High-performance computing earns its value by shortening a scientifically meaningful calculation or enabling a scale that smaller systems cannot manage. A larger machine does not repair an unclear model or weak data. The practical capability is shared-resource literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Schedulers coordinate shared resources” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Queue time is not a direct measure of a project’s scientific merit.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Strong scaling holds problem size constant while adding resources; weak scaling grows the problem with resources. Speedup, efficiency and cost should be measured over realistic cases. Profiling identifies serial sections, imbalance and communication.

Parallel computing divides work across processors while communication, memory and storage impose limits. Performance depends on the algorithm and data movement as much as on the advertised number of cores or accelerators. The practical capability is performance evaluation. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Scaling tests reveal useful size” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A short benchmark may not represent production input and I/O.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Teams estimate core-hours, GPU-hours, memory, storage and wall time from pilots or comparable workloads. NSCC’s current project guidance evaluates computational complexity and use track record alongside national priorities. Estimates should include analysis and failed-run contingencies without padding.

A supercomputer is a shared research instrument. Schedulers, allocations, accounting and support services help many teams use expensive national resources responsibly and reproducibly. The practical capability is resource-estimation reasoning. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Resource estimates connect science with stewardship” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: An allocation is permission to use resources, not a guarantee of scientific success.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Verification asks whether equations or algorithms were implemented and solved correctly. Tests include known solutions, conservation checks, code comparison, convergence and sensitivity to numerical parameters. Floating-point arithmetic and parallel ordering can affect results.

Computational evidence needs provenance: code, environment, inputs, parameters, hardware, random seeds, logs and analysis. Without that record, speed can produce answers that are difficult to inspect or repeat. The practical capability is numerical verification. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Numerical verification checks the computation” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A program that runs without error may still solve the wrong equations incorrectly.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Validation compares model outputs with measurements or established observations for the intended use. Parameter fitting and validation data should be separated where possible. Uncertainty from inputs, structure and measurement remains visible.

High-performance computing earns its value by shortening a scientifically meaningful calculation or enabling a scale that smaller systems cannot manage. A larger machine does not repair an unclear model or weak data. The practical capability is model-validation literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Validation checks the model against reality” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Agreement in one regime does not validate extrapolation to another.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Climate, fluid flow, materials, engineering, genomics and biomedicine use different models and data. Each domain has its own validation standards and ethical constraints. NSCC lists many of these fields as national HPC applications.

Parallel computing divides work across processors while communication, memory and storage impose limits. Performance depends on the algorithm and data movement as much as on the advertised number of cores or accelerators. The practical capability is interdisciplinary computing. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “HPC supports simulation across domains” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A common computing platform does not make domain evidence interchangeable.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Large datasets and accelerators support training and inference. Data provenance, leakage, baselines, calibration, subgroup performance and external validation matter. Scientific models should produce testable insights rather than only fit benchmarks.

A supercomputer is a shared research instrument. Schedulers, allocations, accounting and support services help many teams use expensive national resources responsibly and reproducibly. The practical capability is AI evaluation literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “AI workloads still need scientific evaluation” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A larger model or lower training loss does not establish scientific truth.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Sensitive data may require approved environments, access controls, logging and de-identification. Computing also consumes electricity and cooling resources. Efficient algorithms, right-sized runs and reuse of validated outputs support responsible stewardship.

Computational evidence needs provenance: code, environment, inputs, parameters, hardware, random seeds, logs and analysis. Without that record, speed can produce answers that are difficult to inspect or repeat. The practical capability is responsible-computing literacy. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Data governance and energy are part of responsible HPC” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Removing direct identifiers does not eliminate every privacy risk.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Archive question, model, code commit, environment, inputs, licences, parameters, random seeds, hardware, allocation, job scripts, logs, profiling, failures, outputs, analysis and uncertainty. Students can learn parallel concepts on small systems while authorised users control national infrastructure.

High-performance computing earns its value by shortening a scientifically meaningful calculation or enabling a scale that smaller systems cannot manage. A larger machine does not repair an unclear model or weak data. The practical capability is HPC career preparation. Keep the question, method, controls, observation, uncertainty, review and next decision in one evidence receipt. That record makes progress visible without enlarging the claim beyond the stage reached. NSCC Singapore controls current systems, allocation policies and access; project owners, institutions, data custodians and regulators control research use and conclusions.

Put this into practice by turning the idea into a small, reviewable action: Build a dated evidence card for “Keep an HPC evidence receipt” with the question, method, observation, limitation and next test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Only current NSCC policies and institutional approvals determine access and data use.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter