NVIDIA B300: what it is and how to run it in the UK
NVIDIA B300 is a Blackwell Ultra data centre GPU generation for AI training and inference. For buyers, its significance is practical: memory, precision and interconnect choices affect how large models are served and how work spans accelerators. Start with your model, required precision, context and concurrency, then ask for the exact offered configuration and test it. A B300 name on a product page is not enough to predict throughput or prove that a service fits your workload.
B300 GPU: what the name tells you, and what it does not
NVIDIA places B300 within its Blackwell Ultra family, which it designs for reasoning, training and high-throughput inference. The company's technical material emphasises high-capacity HBM3E, low-precision tensor compute and GPU-to-GPU communication.
Those are meaningful design considerations, but they are not a universal hosted-service specification. The product behind an offer can differ, and the resources exposed to your workload depend on the allocation. Ask for the complete product name and whether you are evaluating an accelerator allocation, a server or a managed service.
The same caution applies to performance figures. Dense and sparse figures are not interchangeable, and a peak tensor-compute number does not tell you the tokens per second your model will produce. Keep the precision, workload and measurement conditions attached to every comparison.
Start with memory, not a headline compute number
A model's weights are only one consumer of accelerator memory. An inference service also needs room for runtime overhead and its key-value cache. That cache depends on the model, context length and active requests. Training has different working-memory requirements, and distributed work adds communication considerations.
For that reason, do not ask only whether a model fits. Ask what fits at your actual precision, concurrency and context, with reasonable operating headroom. A test that fills memory exactly as your production service will is more useful than a demonstration with a short prompt and one request.
NVIDIA's own documentation distinguishes high-level accelerator capability from the specifications of particular systems. Treat any system memory total carefully. Confirm whether a figure refers to one accelerator, an allocation or a complete server, and whether that full capacity is exposed to your workload.
DGX B300 is a system, not a blanket service specification
DGX B300 is NVIDIA's integrated system built with Blackwell Ultra GPUs. NVIDIA's datasheet describes an air-cooled chassis and lists 2.1 TB of GPU memory space for the system. It also describes NVIDIA software for workload management, telemetry and recovery.
The DGX B300 name therefore answers a different question from B300 GPU. One identifies a complete system and its supporting stack, while the other identifies the accelerator generation. A rental offer should tell you precisely which product and resource boundary you are buying.
Do not carry a complete DGX B300 system specification into an unrelated rental quote. If a provider gives you a system figure, ask what portion is allocated to your workload and what sits outside the service boundary. This is especially important when the provider, server supplier and accelerator maker are different organisations.
What B300 generation capacity means for your workload
Use this matrix before you compare offers. It separates what the manufacturer documents from what a service provider still has to prove.
| Question | Why it matters | Evidence to request |
|---|---|---|
| Which exact B300 product is offered? | B300 GPU and DGX B300 are not interchangeable descriptions | Complete configuration, with system and allocation named separately |
| How much memory can you use? | Weights, runtime overhead and cache compete for capacity | Accessible memory per allocation, stated with its units |
| What precision will you run? | Peak figures depend on the numeric mode and test conditions | The supported mode that matches your model and accuracy target |
| How does work communicate? | Distributed jobs depend on the connections between resources | The interconnect relevant to your allocation |
| What is the tenancy? | An allocation may not give you a complete accelerator or system | Whether accelerators, hosts and storage are shared |
| Which software is supported? | Driver and serving-stack choices affect deployment work | Supported software versions and responsibility for changes |
| How will it be measured? | Manufacturer figures cannot predict every model or serving pattern | Your reproducible acceptance test and its settings |
A practical acceptance test for NVIDIA B300 capacity
Freeze the parts of the test that a provider comparison should not change. Record the model revision, weights format, serving engine, precision, maximum context, concurrency and the distribution of requests. Use production-like prompts or batches. If the application has a response target, state it before the test starts.
For inference, capture time to first token, inter-token latency and throughput at each tested load. Keep cold starts separate from a warmed service. Watch memory consumption throughout the run, not only at startup, because cache growth can change the result.
For training or fine-tuning, test the proposed data path and checkpoint process alongside the compute step. A fast kernel is of little help if the rest of the pipeline stalls. If the job crosses resource boundaries, test that full boundary rather than a smaller, easier subset.
Save the test scripts and the conditions with the results. If an offer changes, you then have a fair basis for deciding whether it still meets the same requirement.
Running B300 generation capacity in the UK
Location, ownership and operating responsibility belong in the same evaluation as the accelerator. Confirm where your workload will run, who owns the equipment, whether any capacity is resold and who is responsible for each part of the service. Read those answers alongside the allocation and software details.
LinkPool has B300 generation GPU capacity on its own hardware in the UK. It is UK owned and operated, and we do not resell another provider's GPU capacity. Those are the facts we put behind our offer; the next step is a workload-specific discussion, not an assumption that one configuration suits every buyer.
If you are also comparing GPU server hosting and GPU rental in the UK, the same buyer checklist applies. For a closer look at the offering, see LinkPool's UK owned and operated B300 generation GPU capacity.
Sources
- NVIDIA, "NVIDIA Blackwell Ultra", datasheet, technical specifications and technology explanations: https://dam-cdn.nvd.orangelogic.com/AssetLink/1k0p832eq8r5ca0u5383ie5o4tp3bst1.pdf
- NVIDIA, "NVIDIA DGX B300", datasheet, system positioning and memory-space specification: https://resources.nvidia.com/en-us-dgx-systems/dgx-b300-datasheet
- NVIDIA, "Metrics", NIM benchmarking documentation, definitions of time to first token and inter-token latency: https://docs.nvidia.com/nim/benchmarking/llm/metrics
- NVIDIA, "Advanced performance tuning", Dynamo documentation, guidance on representative workloads and benchmark conditions: https://docs.nvidia.com/dynamo/kubernetes/operations/performance-tuning
Frequently asked questions
What is NVIDIA B300?
NVIDIA B300 refers to a generation of Blackwell Ultra data centre GPU used in NVIDIA systems for AI training and inference. Treat the name as the start of a configuration check, because the exact product, accessible memory and service allocation still matter.
What should I check in a B300 GPU offer?
Check the exact product name, accessible memory per allocation, tenancy, interconnect, host resources, supported software and operating boundary. Test your own model at the required precision, context length and concurrency.
Is DGX B300 the same as a B300 GPU?
No. DGX B300 is an NVIDIA system built with Blackwell Ultra GPUs. A B300 GPU is the accelerator generation. A system datasheet and a provider's service allocation are different specifications and should not be treated as interchangeable.
Who owns LinkPool's B300 generation GPU capacity?
LinkPool owns the hardware in the UK. The capacity is UK owned and operated, and LinkPool does not resell another provider's GPU capacity.