On-demand compute
Short-term GPU access for experiments, model evaluation, rendering or a time-limited project.
- Discuss the GPU and software environment
- Agree storage, duration and total cost
- Run your own workload
Percepta / Compute servicesGPU compute.
GPU compute.
Private AI.
GPU hosting for research, simulation and AI inference.
Rent capacity for your own workloads, reserve a server for ongoing work, or discuss hosting for your own language models. UK-hosted compute and on-premises AI systems for businesses and research teams.
Drawn live by your device.
GPU compute
GPU capacity for research, simulation and AI.
What are you working on?
01 / Research & engineeringResearch, simulation & modelling.Capacity for the next run, or the whole project.02 / LLM inferenceLLM hosting & private AI.Model selection, serving engines and compute.03 / GPU configurationsRTX PRO 6000 to H200.Single GPU, multi-GPU and high-memory options.
Direct access to PerceptaQuoted to your requirements
Established Vast.ai hostingUK data-centre deploymentOn-demand & longer-term access
Access & deployment
Compute services.
Run your own software on rented GPUs, or ask us to scope a managed deployment. Hardware access and application management are separate services, with clear responsibilities and pricing.
Reserved capacity
A longer-term arrangement for recurring jobs or a service that needs an agreed allocation of compute.
- Single- or multi-GPU configurations
- Defined term and resource allocation
- Support and maintenance agreed up front
Private AI systems
A local model in a UK-hosted environment or a system on your premises, designed around your business requirements.
- Model and hardware selection
- Deployment and data-flow planning
- Optional application integration by agreement
Already comfortable managing your own instance? Selected Percepta capacity is also offered through Vast.ai. Marketplace pricing, availability and terms apply; enquire directly if you need a particular configuration.
Businesses / Research / Engineering
What could you run?
You do not have to be an AI company to benefit from GPU compute. The important question is whether your software can use it effectively.
Research & experimentation
Model evaluation, parameter sweeps and reproducible training runs. Reserve capacity for a study or run independent experiments across GPUs.
Simulation & modelling
GPU-enabled numerical analysis and engineering solvers. The application, precision requirements and dataset determine whether RTX or H200 is appropriate.
LLM inference & adaptation
Run chat, coding and document-understanding models; evaluate alternatives or carry out supported adapter fine-tuning. Size the deployment for context length and concurrent requests.
Private business AI
Internal assistants and search over approved documents, using a hosted or on-premises model. Retrieval, access permissions and application integration are scoped together.
Computer vision
Train and evaluate detection models, test new datasets or run batch image analysis. Keep development compute separate from the systems deployed at your sites.
Rendering & batch processing
Render queues, visualisation and media-processing jobs. Match the GPU to the renderer, required features and software licence before reserving capacity.
For simulation and research: not every solver is GPU accelerated. Tell us the application, precision requirements and dataset size. We can discuss a representative trial before you commit to a longer allocation.
Model size / Concurrency / Precision
GPU configurations.
Choose around the workload: the model you need to serve, the number of simultaneous requests, or the solver you need to run. Exact hardware, availability and pricing are confirmed in your proposal.
CoolingActive double-flow-through
CoolingActive blower fan
CoolingPassive air-cooled; requires qualified ducted server airflow
Parts list
Workstation Edition- Form factor
- Extended-height, dual-slot
- Nominal dimensions (height x length)
- 5.4 x 12 inches; additional installation clearance required
- Cooling
- Active double-flow-through
Allow suitable chassis clearance, airflow and system power for this edition.
The extended-height, dual-slot cooler uses two flow-through airflow paths for the 600 W design.
- GPU architecture
- NVIDIA Blackwell / GB202
- CUDA cores
- 24,064
- Tensor cores
- 752 / fifth generation
- RT cores
- 188 / fourth generation
- Graphics memory
- 96 GB GDDR7 with ECC
- Memory interface
- 512-bit
- Memory bandwidth
- 1,792 GB/s
- Total board power
- 600 W
Choose the PSU for the complete system, not the GPU alone.
- Auxiliary power
- 1 x PCIe CEM5 16-pin
The GPU has a 600 W board-power rating and requires one PCIe CEM5 16-pin connection.
- Host interface
- PCIe 5.0 x16
- NVLink bridge
- Not supported
- Display connectors
- 4 x DisplayPort 2.1b
- Display operation
- Direct display output; enabling MIG compute mode disables physical display outputs
Parts list
Max-Q Workstation Edition- Form factor
- Full-height, dual-slot
- Nominal dimensions (height x length)
- 4.4 x 10.5 inches; additional installation clearance required
- Cooling
- Active blower fan
Allow additional room for the power cable and blower airflow.
The 300 W active blower design prioritises power and space efficiency. Check slot spacing, airflow and total system power with your system manufacturer.
- GPU architecture
- NVIDIA Blackwell / GB202
- CUDA cores
- 24,064
- Tensor cores
- 752 / fifth generation
- RT cores
- 188 / fourth generation
- Graphics memory
- 96 GB GDDR7 with ECC
- Memory interface
- 512-bit
- Memory bandwidth
- 1,792 GB/s
- Total board power
- 300 W
The 300 W rating is for the GPU. Check PSU capacity for the complete system.
- Auxiliary power
- 1 x PCIe CEM5 16-pin
The card requires one PCIe CEM5 16-pin connection.
- Host interface
- PCIe 5.0 x16
- NVLink bridge
- Not supported
- Display connectors
- 4 x DisplayPort 2.1b
- Display operation
- Direct display output; enabling MIG compute mode disables physical display outputs
Parts list
Server Edition- Form factor
- Full-height full-length, dual-slot
- Nominal dimensions (height x length)
- 4.4 x 10.5 inches; additional installation clearance required
- Cooling
- Passive air-cooled; requires qualified ducted server airflow
Use a server with the forced airflow required for this passive card.
- GPU architecture
- NVIDIA Blackwell / GB202
- CUDA cores
- 24,064
- Tensor cores
- 752 / fifth generation
- RT cores
- 188 / fourth generation
- Graphics memory
- 96 GB GDDR7 with ECC
- Memory interface
- 512-bit
- Memory bandwidth
- 1,597 GB/s
- Total board power
- Up to 600 W; configurable, system and cable requirements apply
- Auxiliary power
- 1 x PCIe CEM5 16-pin
Board power is configurable up to 600 W; the supported setting depends on the board, firmware, cabling and host system.
- Host interface
- PCIe 5.0 x16
- NVLink bridge
- Not supported
- Display connectors
- 4 x DisplayPort 2.1b
- Display operation
- Disabled in factory-default display-off mode; display-on requires supported configuration
PERCEPTASolutions Ltd / Edinburgh
Sheet01 / 0302 / 0303 / 03
TitleRTX PRO 6000 Blackwell Workstation EditionRTX PRO 6000 Max-Q Workstation EditionRTX PRO 6000 Blackwell Server Edition
- 96 GB
- GDDR7 memory with ECC
- 1,792 GB/s
- Memory bandwidth
- 600 W
- Total board power
- PCIe 5.0
- x16 host interface
- 96 GB
- GDDR7 memory with ECC
- 1,792 GB/s
- Memory bandwidth
- 300 W
- Total board power
- PCIe 5.0
- x16 host interface
- 96 GB
- GDDR7 memory with ECC
- 1,597 GB/s
- Memory bandwidth
- Up to 600 W
- Configurable board power
- PCIe 5.0
- x16 host interface
ViewExploded axonometric
RTX PRO 6000
96GBper Blackwell RTX PRO 6000 GPU / GDDR7 ECC
LLM hosting, a development environment or a single-GPU job.
- LLM serving
- Model weights, context cache and runtime must fit together. Quantisation, context length and concurrent requests determine capacity.
- Other workloads
- Model evaluation, supported adapter fine-tuning, computer vision training and GPU rendering.
- Before booking
- Validate the model, precision and software stack against your latency or job-time requirement.
RTX PRO 6000 servers
96GB / GPUSeparate memory on each GPU
More independent work, or a model distributed across GPUs.
- More concurrent requests
- Run multiple model replicas when each copy fits on one GPU. Route requests between workers.
- Larger models & training
- Use tensor or pipeline parallelism, or distributed training, where the model and framework support it.
- Research & batch work
- Run separate experiments, simulation jobs or render workers. Check data transfer and server topology before scaling one job.
H200 servers
141GBper NVIDIA H200 GPU / HBM3e
Memory-heavy inference and supported scientific workloads.
- LLM serving
- More memory per GPU for weights and context cache. HBM3e bandwidth is relevant to memory-bound inference.
- Simulation & numerical work
- Double-precision (FP64) capability for compatible scientific software; choose by solver requirements, not AI performance figures.
- Before booking
- Confirm the H200 variant, GPU count, interconnect and application support. Test the intended workload.
VRAM is not a model-size guarantee. Allow for model weights, the KV cache used by active requests and runtime overhead. Multiple GPUs need explicit software support to share a model; their memory is not automatically pooled.
Manufacturer specifications: RTX PRO 6000 / H200. These are configuration options, not a live availability list. CPU, system RAM, storage and networking are sized alongside the GPUs.
Language models / Private deploymentsLLM inference
LLM inference
& model hosting.
Run open-weight models on UK-hosted GPUs or a system on your premises. Choose the model for your application, then size the compute for the context, response time and number of users it needs to support.
GPU memoryA container of GPU memory holding the model weights, the runtime overhead and a context and KV cache for each simultaneous request; the space left above them is the operating headroom.
Simultaneous requests
Context length
A model loading successfully is not the same as serving it well under load. Longer conversations and more simultaneous requests change the memory budget and response time.
Model selection
Models for your application.
Models evolve quickly. We help you choose from the latest available options based on your tasks, data, privacy requirements and budget.
A smaller model may meet your needs with less hardware. We test quality, response time and memory use on your workload, then confirm the model, licence and configuration before deployment.
Chat, analysis & internal assistants
Internal chat, summarisation and analysis, with reasoning suited to the task. Evaluate quality, languages and response time on your own tasks.
Development tools & code assistance
Code generation, explanation and repair. Test against your languages and repository context; IDE integration and tool execution are separate application work.
Document understanding
Interpret text, tables and charts in document images. Validate extraction on representative pages, with document preparation and access controls included in the scope.
Embeddings & reranking
Generate vectors for semantic search, then rank retrieved documents for relevance. These complement a language model; ingestion, indexing and permissions need their own setup.
Serving engines
vLLM & SGLang.
Both support batching and prefix caching. We compare them on the exact model, GPU and request pattern, rather than assume one is faster.
vLLM
Continuous batching and paged attention help schedule requests and manage KV-cache memory. Model support covers generation and retrieval workloads.
SGLang
RadixAttention supports reuse of shared prompt prefixes. A candidate for repeated-context serving, with model-specific distributed inference options.
Validate before deployment. We agree model/version and licence, context and concurrency limits, then test response quality, time to first token and output speed on the chosen hardware. Deployment assistance and ongoing management are scoped separately from GPU rental.
Discuss LLM hostingPrivate AI / Deployment optionsUK hosted.
UK hosted.
Or on your premises.
For businesses that do not want confidential prompts or documents routinely sent to a public AI service. We can discuss local-model hosting and on-premises systems, with the controls and responsibilities agreed before deployment.
| UK hosted | Your premises | |
|---|---|---|
| 01Where the model runs | UK colocation facilityNo server room required Run a selected model on allocated compute at our UK colocation facility. This is remotely hosted computing, not an on-premises system. | Your on-site systemKeep inference on site A local model can process prompts and documents within your premises. |
| 02Where your data goes | Agreed secure connection The AI model, GPU compute and approved business data sit inside a data boundary. | Local network External tools, updates and remote support need separate consideration. |
| 03Who is responsible | Define the data boundary Agree access, tenant isolation, storage, logging, backups and administrator permissions before moving confidential information. | Plan the physical setup We scope hardware, power, cooling, networking and handover with you. Installation and ongoing management are agreed separately. |
Deployment design is agreed per project. A private model alone does not guarantee confidentiality or regulatory compliance. | ||
Deployment scope
Data access & integration.
For an internal assistant, agree which documents it can retrieve and whose permissions it follows. For a hosted model, agree access controls, logging and retention. We scope these alongside the model, not as an assumption about where the server sits.
Operating experience / UK colocationOur hosting
Our hosting
infrastructure.
We already host customer GPU workloads through Vast.ai and operate company-owned compute in UK colocation. You deal directly with Percepta about the hardware, deployment and support you need.
01
Location & data handling
We confirm the hosting location and data-handling requirements for your deployment. Storage, backups and any external services are included in that review.
02
Operating responsibilities
Agree who manages the operating system, models, updates, backups and monitoring. Renting compute does not automatically include application support or backup protection.
03
Availability by agreement
Capacity, maintenance windows, response times and any service-level commitment are defined in your contract. Marketplace hosting history is not a guarantee for every application.
Cost & commitment
How we quote.
For research budgets and business projects, the useful number is the cost of getting the work done. We scope the resources and term before you commit.
01
Compute
GPU model and count, CPU allocation and system memory.
02
Data
Working storage, retention, backups and any transfer costs.
03
Time
Project hours or a longer-term reserved allocation.
04
Support
Self-managed compute, deployment assistance or an agreed managed scope.
Have an existing cloud estimate? Bring it with your workload details. We can compare equivalent resources and scope; savings depend on utilisation, software, support and data transfer.
Before you book
A few practical details.
Can I rent for one project rather than a long contract?
Yes. Discuss short-term access with us or browse available instances on Vast.ai. Direct bookings are quoted and confirmed before the allocation starts; availability and minimum terms depend on the configuration.
Will my simulation or model run faster on a GPU?
Only if the software can use the GPU effectively. We need the application and version, GPU support, licence conditions, precision requirements and a representative workload. Some simulations are CPU-bound or require a different class of accelerator.
Does private AI mean nobody else can access my data?
No. A local model changes where inference runs, but access permissions, administrator access, logs, external tools, backups and support still matter. We agree the deployment boundary and data-handling arrangements with you; hosted and on-premises options have different responsibilities.
Can you build and install a system for our premises?
We can scope a local AI workstation or server around your use case. Hardware, software licensing, installation, network integration, training and ongoing support are confirmed separately in the proposal.
Are software licences, backups and managed support included?
Do not assume they are included in a GPU rental. Bring details of any commercial software or restricted datasets. We confirm permitted use, customer responsibilities and any additional storage, backup or support services before the work starts.
Talk directly to PerceptaTell us what
Tell us what
you need to run.
A first conversation does not need a perfect specification. The application, deadline and budget are a useful starting point.
- Your software or model and intended use
- GPU, memory and storage needs, if known
- Project duration, deadline and budget
- Hosted or on-premises; any data restrictions
Please describe the requirement without sending confidential datasets, credentials or personal records.
Send a secure enquiry
All fields marked * are required.