Seven practices, one engineering team. We design, build, benchmark and operate the compute layer beneath your hardest problems — on-premise, hybrid or sovereign cloud.
A genomics pipeline and a trading engine are not the same machine. We start from your regulatory perimeter, your data gravity and your deadlines — then design backwards.
Servers & Workstations
One 8-way accelerator node draws around 10 kW under sustained load. Put eight in a rack and the number in front of you is 80 kW, in a hall that was probably designed for 6 kW a rack. So we model the power draw and the heat-rejection path before we select a single component, because that is what decides whether the hardware you paid for can run at full clock.
The Constraint
Accelerator power per socket has roughly quadrupled in a decade while rack footprints have not changed. The result is that most GPU procurement fails at the facility, not at the server — quietly, as sustained clock throttling nobody attributes to cooling.
Thermal limits rarely announce themselves. The job completes, the dashboard is green, and the accelerators are sitting several hundred megahertz below their rated clock because inlet air is too warm or the rear of the rack is recirculating. The loss is spread across every run and attributed to the model.
We measure it directly. Junction temperature, clock residency, power capping events and fan duty are logged during burn-in and again during on-site commissioning. If the facility cannot sustain the configuration, we say so while there is still time to change the design, the cooling, or the density.
Density is a choice, not a virtue. Four nodes in two racks with a comfortable thermal margin often deliver more useful throughput over three years than eight nodes in one rack that spend the summer capped.
Specification Method
Small, medium and large are pricing constructs. They describe a vendor's inventory, not your job. Two teams with identical budgets and identical GPU counts can need entirely different machines, and usually do.
We profile the real workload, not a proxy. Per-process memory high-water mark, memory bandwidth saturation, PCIe and NVLink traffic, storage queue depth and read size distribution, single-thread frequency sensitivity, and how much of the wall clock is spent waiting rather than computing.
More often than not the accelerator is not the limit. A training loop starved by a single-threaded data loader, a solver bound by memory bandwidth rather than cores, a render farm waiting on 4 kB random reads from the wrong storage tier. Buying more GPUs makes several of those worse.
Whether a job needs an NVLink domain or is content with PCIe changes the chassis, the price and the thermal envelope. Multi-node training needs a rail-optimised fabric and GPUDirect paths. Independent inference replicas need neither, and paying for both is a common and expensive error.
Sustained and peak draw per node, aggregated per rack with headroom. PDU and busway sizing with the derating rule applied. Redundancy topology. Airflow or liquid loop capacity, and the heat-rejection path traced end to end to the chiller or the dry cooler on the roof.
Capital cost is the smaller half. We model energy at your tariff, cooling overhead, rack space, support, and the residual value at disposal. It is common for the configuration with the higher purchase price to be the cheaper machine by the end of year two.
The benchmarks that will decide whether the build is accepted are agreed in writing first, with the method and the pass criteria. That removes the argument at delivery and it keeps us honest, because we are measured against your workload rather than a synthetic score.
Rack Design
Every density band has a cooling method that works and several that do not. The band you land in is decided by the node count and the accelerator choice, so we settle it on a spreadsheet before anything is ordered.
| Rack density | Workable cooling method | What usually goes wrong |
|---|---|---|
| Up to 10 kW | Conventional raised floor or perimeter units. Blanking panels and cable management are sufficient. | Missing blanking panels and an unmanaged rear bundle recirculating warm air to the intake. |
| 10–25 kW | Hot-aisle or cold-aisle containment, floor-tile airflow measured and balanced rather than assumed. | Containment fitted but tile placement never re-measured, so one rack in the row starves. |
| 25–50 kW | In-row cooling or rear-door heat exchangers on a chilled-water loop, with contained aisles. | Chilled-water return temperature too high for the coil to reject the load at full flow. |
| 50–100 kW | Rear-door heat exchangers at high flow, or direct-to-chip liquid cooling on the accelerators with air for the remainder. | No coolant distribution unit budgeted, and no leak detection or service procedure agreed with facilities. |
| Above 100 kW | Direct-to-chip liquid cooling with a CDU per rack or per row. Immersion where the facility and the vendor both support it. | Floor loading, busway capacity and the service clearance for the CDU discovered after the order is placed. |
Sustained draw, not nameplate, sized with the derating your electrical contractor applies. Three-phase balance checked across the rack rather than per node. A/B feed topology matched to the node's own redundancy so a single feed loss does not take the rack. Metered PDUs at outlet level, because per-rack totals hide the node that is capping.
Blanking panels on every free U, brush strips at every cable entry, and rear cable management that does not obstruct the exhaust. Airflow measured at the tile and at the intake of the worst-placed node, not calculated from room totals. Where a row is mixed-density we model the hot spot, because the average is always comfortable and never true.
The most common practical step past 25 kW, and the least disruptive. The rack rejects its heat to water at the door rather than into the room, which leaves the rest of the hall unchanged. It needs loop capacity, a return temperature the coil can work with, and a fan failure mode your operations team has rehearsed.
Necessary at the top of the range and increasingly the default for dense accelerator nodes. Cold plates on the high-power components, a coolant distribution unit, quick-disconnect service procedures, leak detection and a maintenance regime facilities must sign up to. Roughly a fifth of the heat still leaves by air, so the room still matters.
Indicative Range
These are reference configurations we start conversations from, not a product catalogue. Every one of them changes once we have the workload profile — which is the point of publishing them as starting points rather than tiers.
| Attribute | Data science workstation | Visualisation workstation | 4U accelerator node | GPU-dense training rack |
|---|---|---|---|---|
| Intended work | Feature engineering, notebook iteration, single-GPU fine-tuning | Real-time 3D, CAD and DCC review, volumetric and medical imaging | Multi-GPU training, batch inference, GPU rendering | Distributed multi-node training and large-batch inference |
| CPU | 1 × 16–24 core workstation class, high single-thread clock | 1 × 24–32 core, tuned for viewport responsiveness | 2 × 32–64 core server class, selected for PCIe lane count | As the 4U node, replicated across eight chassis |
| Accelerators | 1 × 48 GB professional GPU with ECC | 2 × 48 GB professional GPU, NVLink where the application uses it | 4–8 × data-centre GPU, 80–141 GB HBM, full NVLink domain | 32–64 accelerators across the rack |
| System memory | 128–256 GB ECC DDR5 | 256–512 GB ECC DDR5 | 1–2 TB ECC DDR5, populated for full channel bandwidth | 8–16 TB aggregate |
| Local storage | 2 × 2 TB NVMe, mirrored | 4 × 4 TB NVMe, scratch separated from the OS | 8 × 7.68 TB NVMe scratch plus a boot mirror | Node-local NVMe scratch plus a parallel file system mount |
| Fabric | 10/25 GbE | 25 GbE, with a remote-display path sized for the codec | 2 × 200–400 Gb InfiniBand or RoCE, separate management network | Non-blocking rail-optimised fabric, NVMe-oF to storage |
| Sustained draw | 0.6–0.9 kW | 1.0–1.4 kW | 6–10 kW | 48–80 kW per rack |
| Cooling assumption | Office ambient, acoustically acceptable at full load | Office ambient with directed intake and managed desk placement | Contained aisle; direct-to-chip option on the accelerators | Rear-door heat exchanger or direct-to-chip with a CDU |
| Driver stack | Validated production branch, pinned with the framework build | ISV-certified branch for the applications in use | Validated branch, container images pinned to it | One validated branch across the fleet, rolled forward together |
| Warranty position | 3 year on-site, next business day | 3 year on-site, next business day | 3 year on-site, 4-hour response option in metro | 3 year on-site, 4-hour response plus an on-site spares kit |
Workstations
A specification sheet tells you very little about whether a researcher's Monday will work. What decides that is the combination of kernel, driver branch, runtime, framework build and application certification — and every one of those can break the others.
We ship workstations on a validated stack: a specific kernel, a production driver branch rather than the newest one, a matched CUDA or ROCm runtime, and container images pinned against it. The combination is recorded in the build document, and the fleet moves forward together rather than machine by machine.
This matters because the failure is subtle. A driver update changes a kernel selection heuristic, a training run produces a different loss curve, and a week disappears into a bug that was never in the code. In a clinical or engineering context the same change can invalidate a vendor's certification and, with it, your validation evidence.
Remote use is now the norm rather than the exception. A workstation accessed over a remote display protocol is limited by codec offload and the network path long before it is limited by the GPU, so we specify and test that path as part of the build.
Sized so that the iteration loop stays at the desk. Enough accelerator memory to fine-tune without a cluster booking, fast local NVMe for the dataset working set, and a container image identical to the one the cluster runs so code moves without surprises.
Specified for viewport responsiveness rather than batch throughput, which is a different machine. Certified driver branches for the studio's tool set, OpenUSD-based pipelines tested end to end, and colour management verified against the facility's reference display.
Memory bandwidth and single-thread clock chosen against the solver in use, since many engineering codes scale with neither core count nor GPU. Licence-bound applications are profiled for frequency sensitivity before we recommend a socket.
Burn-in & Acceptance
Infant mortality in accelerators, memory and power supplies shows up under sustained load, not in a boot test. We run every system hot in our own facility first, and the log comes with it.
72h
Sustained synthetic load at target ambient before any system is released for shipment. Extended to 168 hours for liquid-cooled racks.
100%
Every unit runs the agreed acceptance benchmarks on its final configuration, with the raw logs supplied rather than a summary.
2.7%
Share of components rejected or replaced during burn-in across the last twelve months. Those are failures your team never has to diagnose.
A boot test proves the machine turns on. Three days at full clock in a warm room proves it works. Those are different claims and only one of them is worth signing.
Specification to Commissioning
Lead times for accelerators move constantly, so we do not publish a duration we cannot hold. What we do commit to is the sequence, the artefact produced at each step, and who signs it.
We instrument the real job and produce a profile: memory footprint, bandwidth saturation, interconnect traffic, storage access pattern and the frequency sensitivity of the hot path. Artefact — a profile report with the bottleneck named.
Measured, not assumed. Circuit capacity and phase balance, breaker sizes, floor loading, supply air temperature, containment state, chilled-water flow and return temperature, and the physical access route including the narrowest door. Artefact — a site constraints register.
Sustained and peak draw per node aggregated to the rack, PDU and busway sizing with derating, redundancy topology, and the heat-rejection path traced to the chiller. Artefact — the model, with the density decision and its headroom stated.
Component selection against the profile and the thermal envelope, with firmware levels, driver branch and BIOS settings written into the build specification. The acceptance benchmarks and pass criteria are agreed here, before the order. Artefact — build sheet and signed acceptance test.
Assembly in our own facility. Cable management planned for airflow and for service access, firmware levelled across the fleet, BIOS and BMC configuration applied from a template so every unit is identical. Artefact — as-built record per serial number.
Sustained concurrent load at elevated ambient, with thermal, clock, power and memory-error telemetry logged. Anything marginal is replaced and the unit restarts its burn-in. Artefact — the burn-in log, supplied in full.
The agreed benchmarks run on the exact shipping configuration, documented with method, ambient conditions, driver and firmware versions, and raw output. If a result misses the criteria we tell you before dispatch, not after. Artefact — the benchmark report.
Rack, cable, power on, BMC and network onboarding, firmware verified against the as-built record, acceptance benchmarks repeated in your facility, and a thermal check under sustained load before handover. Artefact — commissioning report, updated asset register and an operations handover session.
Support & End of Life
Accelerator lead times are long, spares policy decides your real availability, and disposal is where a well-run estate most often leaks data. All three are specified at purchase, not improvised in year four.
On-site means an engineer attends, not that a part is couriered to your dock. Response targets are written against business hours or 24/7 as you require, and they are measured. Where a vendor's own warranty is the faster path we say so and manage the case rather than duplicating the cover.
Specified at purchase · measured monthly
Accelerator lead times make a global spares pool a poor availability strategy. For dense estates we recommend an on-site kit — power supplies, fans, one accelerator, cables and optics — plus a cold spare node where the workload cannot absorb the loss of one. Consumed spares are replenished on use.
Sized to your availability target, not ours
One firmware and driver baseline per fleet, tested on a staging unit against your own workload before rollout, with a documented rollback. Refresh planning starts at year two so the replacement is budgeted rather than urgent, and so the residual value is realised while it still exists.
Baseline under version control
Every storage device is accounted for by serial number. Self-encrypting NVMe is cryptographically erased and verified by read-back; where classification or policy requires it the media is physically destroyed instead. Chassis and accelerators go to an accredited recycler with the downstream evidence retained.
Certificate per serial · register reconciled
Let's Talk
Send the workload, the rack you intend to put it in, and the supply air temperature if you know it. You get a power and thermal model, a configuration sized to the profile, and a straight answer about whether the facility can sustain it before anyone writes a purchase order.
Built and burned in locally. Commissioned on site by the team that assembled it.