GB300 NVL72 is the Grace Blackwell Ultra rack-scale platform. Plan it as an integrated deployment spanning compute, fabric, data movement and operations.
| System | 72 × Blackwell Ultra GPU + 36 × Grace CPU |
|---|---|
| Aggregate GPU memory | ~20 TB HBM3e |
| NVLink bandwidth | 130 TB/s aggregate (5th-gen NVLink) |
| FP4 Tensor Core (sparse) | 1,440 PFLOPS (sparse, published) |
Register interest — no payment or commitment required. Arvica confirms configuration, price and availability before you order.
Grace Blackwell Ultra
Exact GPU variant, CPU, RAM and interconnect depend on the confirmed instance. Memory values are hardware references, not allocated resources.
Manufacturer specifications ↗Not offered on demand. Pricing and delivery are set in an Arvica proposal.
Request capacity proposal →Request reserved capacity through an Arvica proposal. Location, start time and access environment are confirmed before provisioning.
Not offered on demand. Pricing and delivery are set in an Arvica proposal.
Term, volume and dedicated-capacity prices are set in an accepted Arvica proposal. Configuration, region and delivery are confirmed in your proposal.
Availability confirmed before order acceptance. Taxes and additional services are confirmed before payment.
Evaluate rack-scale capacity for sustained large-model inference or distributed training with demanding communication requirements. Supply the parallelism plan and service-level objectives with your request.
Compare GB200 for a Grace Blackwell deployment, or B300 nodes when you need smaller increments. Rack-scale hardware does not automatically improve a workload that is limited by software or data ingestion.
NVL72 denotes a 72-GPU system. Confirm usable GPU memory, allocation boundaries and CPU memory independently; aggregate memory should not be treated as one ordinary GPU buffer.
Document the NVLink domain and the network between racks. Include checkpoint storage, ingress and egress paths, power and cooling requirements in the capacity agreement.
Validate Arm-compatible host software and Blackwell Ultra GPU kernels together. Agree on a reproducible model benchmark and deployment acceptance procedure.
NVIDIA hardware reference ↗ · Selection guidance is Arvica's assessment; validate it against your workload
Specify the OS, framework and container image version. Arvica confirms the driver and GPU compatibility before activation. A template name is not a software licence or a guaranteed preinstalled image.
Save an SSH public key during configuration. Keep the private key on your own device. The verified deployment record provides the actual host and login user once access is ready.
Separate boot disk, datasets, model weights and checkpoints in your capacity plan. Confirm persistence, backup ownership, attachment rules and deletion behaviour before relying on a volume.
Confirm ingress, egress, public IP access and private networking where required. For regulated or location-sensitive data, specify the required region in advanced settings before submitting.
Once Arvica supplies verified access, connect using your private key and the exact host/user shown in your instance. Then inspect the allocated GPUs and topology.
ssh -i ~/.ssh/your_key USER@HOST nvidia-smi nvidia-smi topo -m
Replace USER and HOST with your own deployment details. These commands inspect an existing machine; they do not provision one.
Parameter count × bytes per weight estimates weights only. Add KV cache, activations, temporary tensors and framework overhead. Training additionally needs gradients and optimiser state.
Record model version, precision, input/output length, concurrency, latency and error rate together. Compare the cost of a completed job or successful request under the same conditions.
Include all allocated GPUs, runtime, storage, networking, taxes and support. Monthly projections use 730 hours for comparison; your contract defines the billable period and minimum term.
Define your GPU allocation, interconnect, storage, location and commitment. Arvica prepares a capacity proposal around the workload.
Request capacity ↗Model fit and throughput depend on precision, software, batch size, context and system topology. Compare workload estimates rather than an unsupported tokens-per-second promise.
Training cost calculator →Evaluate rack-scale capacity for sustained large-model inference or distributed training with demanding communication requirements. Supply the parallelism plan and service-level objectives with your request.
NVL72 denotes a 72-GPU system. Confirm usable GPU memory, allocation boundaries and CPU memory independently; aggregate memory should not be treated as one ordinary GPU buffer.
Validate Arm-compatible host software and Blackwell Ultra GPU kernels together. Agree on a reproducible model benchmark and deployment acceptance procedure.
Document the NVLink domain and the network between racks. Include checkpoint storage, ingress and egress paths, power and cooling requirements in the capacity agreement.
Do not assume so. Resource allocation, storage and reserved commitments may continue billing. Check the contract and request the appropriate lifecycle action through support.
Retention depends on the agreed storage configuration. Export data and confirm backup and volume retention before requesting termination.
Current configurations enter Arvica capacity review. Running status is shown only after infrastructure is verified.
It is not offered on demand. Price, currency, delivery and terms are set in your Arvica proposal; storage, network, tax and service charges are itemised before you commit.
Use the enterprise capacity path for multi-node networking and reserved allocations.
| GPU | Memory reference | Best-fit decision |
|---|---|---|
| GB300 NVL72 | 72-GPU system | Evaluate rack-scale capacity for sustained large-model inference or distributed training with demanding communication requirements. Supply the parallelism plan and service-level objectives with your request. |
| B200 | 180–192 GB HBM3e | Evaluate B200 for workloads with a Blackwell-compatible software stack and a clear need for low-precision acceleration. Measure quality as well as throughput when changing precision. |
| B300 | 288 GB HBM3e | Evaluate B300 when model state, activations or KV cache constrain your current deployment. Use a representative model and concurrency target to decide whether fewer larger GPUs improve total cost. |
| GB200 NVL72 | 72-GPU system | Consider this platform for distributed training or inference that benefits from a large NVLink domain. Start with the model parallelism strategy, dataset and expected utilisation. |
| Configuration | Practical difference | Confirm before ordering |
|---|---|---|
| NVL72 rack · Bare Metal · 72-GPU system | Dedicated physical host configuration; host access follows the confirmed service | Exact card, GPU allocation, peer-to-peer connectivity, CPU and disk limits |
Weights alone are approximately parameters × bytes per parameter. A 7B model at BF16 needs about 14 GB just for weights; a 70B model about 140 GB. Add KV cache, activations, framework buffers and training state. These are arithmetic estimates, not fit guarantees.
Compare the same model revision, precision, sequence length and concurrency. Record time to first token, p95 latency, output tokens/s, peak memory and full instance cost. Training comparisons should include step time and convergence at the same quality target.
Manufacturer specifications are linked below. No Arvica application-throughput benchmark is published for this configuration yet. Request a workload-specific acceptance test before making performance commitments.
NVIDIA ↗Configure dedicated PCIe / SXM, virtual-machine and bare-metal capacity with Arvica. Submit your requirements for a deployment proposal.
1 configurations available for enquiry
