Services

GPU clusters and HPC

From hardware selection to scheduling and monitoring, the deliverable is a cluster that can be operated; delivery ends with documentation and a handover.

Engineering field guide

Design the cluster around its jobs.

GPU count is only part of the configuration. Data delivery, node communication and job recovery determine whether the system is useful.

Start with the workload

Describe hardware needs through real jobs, rather than individual GPU specifications.

Distributed training & tuning
Provide the model, precision, sequence length, parallelism plan and checkpoint cadence. Evaluate local interconnect, cross-node communication and checkpoint writes together.
Serving & batch compute
Separate latency-sensitive serving from queueable batch work. Define capacity, isolation, scheduling priority and recovery behavior.
Shared research & HPC
Specify users, job types, software stacks and storage access. Define queueing, quotas, reproducible environments and operational permissions.

Compute, data and control together

Connect the physical topology to operations so every component has a defined responsibility.

  1. Compute plane

    Size GPU, CPU, host memory and local interconnect for the job. Record driver and runtime versions and retain a single-node baseline.

  2. Data plane

    Plan dataset reads, caches, checkpoints and archives separately. Test concurrent access and recovery; network bandwidth alone does not validate data delivery.

  3. Control plane

    Configure queues, quotas, monitoring and alerts. Define node draining, failed-job retries, maintenance windows and administrator access.

Accept the workload, not just the hardware

Agree test cases and thresholds before implementation; record hardware, software and test conditions.

Nodes & interconnect
Check device discovery, topology and link state. Record local and cross-node communication tests and investigate abnormal links.
Representative jobs
Run agreed training, serving or compute workloads; inspect throughput, waiting, memory use, data reads and failures.
Recovery & handover
Exercise interrupted jobs, checkpoint recovery, node maintenance and alerts. Deliver topology, configuration, runbooks and ownership boundaries.

Before planning the build

The cluster must fit both the site and the team that will operate it.

Can existing hardware be reused?

Inventory models, memory, interconnect, storage, drivers and support status, then test the target job. Mixed hardware depends on scheduler and framework constraints.

Who operates the cluster after handover?

The customer team takes over by default. Ongoing operations, on-call coverage, hardware maintenance and spares need a separate agreement.

What does planning need?

Example jobs and growth expectations, existing assets, power and cooling, network and storage constraints, procurement scope and the receiving team.

4ROUTER / SYSTEMS RESEARCH

Technical foundations of the service

Understand how operators, KV state and scheduling shape an inference path.

Read the systems notes in Resources
STATE / CAPACITY

KV cache: from tensor shapes to capacity budgets

Context length is a state budget that grows with concurrency.

TIERS / TRANSFER

LMCache: placing KV state in a storage hierarchy

Avoiding prefill requires paying for state access.

SCHEDULING / SLO

From prefill to decode: designing the scheduling experiment

When throughput rises, know which requests are waiting.

Let’s talk infrastructure.

Connect to the platform, or discuss models, performance and deployment.