Buying an NVIDIA DGX B300 starts with a facility question, not a benchmark: can the intended rack deliver and remove roughly 14.5 kW for one 10U node? If the answer is uncertain, the GPU specification is almost secondary. DGX B300 is a complete eight-GPU Blackwell Ultra appliance, and it belongs in a design that treats compute, network, storage, power and support as one system.
This is a source-backed technical review based on NVIDIA's current product page and DGX B300 system documentation. It is not a hands-on laboratory test. Performance figures below are vendor specifications, with sparse and dense precision conventions stated where they matter.
Verdict
DGX B300 is a sensible choice for organisations that need one tightly connected eight-GPU domain, can keep it busy, and prefer an NVIDIA-owned support boundary over a mix of OEM firmware, adapters and software contracts. Its strongest arguments are 2.3 TB of aggregate HBM3e, eight 800 Gb/s GPU-fabric ports, a standardised DGX software environment and a supported route from a single node into a SuperPOD deployment.
It is not a sensible default for every AI team. A lightly used DGX B300 is an expensive space heater. Development groups running models that fit comfortably in one or two PCIe GPUs will usually get better utilisation from smaller servers or shared cloud capacity. Buyers who already operate a strong OEM support relationship may also prefer an HGX B300 system with different cooling, CPU, storage or rack-density choices.
Configure the NVIDIA DGX B300 on GPUMachines.
NVIDIA DGX B300 specifications at a glance
| Area | NVIDIA specification | |---|---| | GPU complex | 8 x NVIDIA B300 Blackwell Ultra SXM GPUs | | GPU memory | 8 x 288 GB HBM3e, 2.3 TB aggregate | | AI compute | 72 PFLOPS sparse FP8 training; 144 PFLOPS sparse FP4 inference | | NVLink fabric | 2 x fifth-generation NVLink switches; 14.4 TB/s aggregate bandwidth | | Host CPUs | 2 x Intel Xeon Platinum 6776P | | Host memory | 2 TB standard; up to 4 TB supported | | Cache storage | 8 x 3.84 TB E1.S NVMe self-encrypting drives | | Boot storage | 2 x 1.92 TB M.2 NVMe drives | | Cluster network | 8 x single-port ConnectX-8 VPI, up to 800 Gb/s each | | Storage and management network | 2 x dual-port BlueField-3 DPUs, up to 400 Gb/s per port | | Rack size | 10U | | Power | 14.5 kW stated consumption; 15 kW system maximum | | AC power supplies | 12 x 3.2 kW, N+N redundant | | Operating environment | DGX OS 7 based on Ubuntu 24.04 LTS, with supported enterprise Linux alternatives |
NVIDIA's marketing page rounds total GPU memory to 2.1 TB and power to about 14 kW, while the system guide reports 8 x 288 GB, or 2.3 TB aggregate, and 14.5 kW consumption. For rack planning, the installation document is the more useful reference. It also gives a 15 kW maximum.
Eight B300 GPUs as one compute domain
The core of DGX B300 is its eight B300 Blackwell Ultra SXM GPUs. Each carries 288 GB of HBM3e, so the node has enough aggregate accelerator memory to keep very large models and substantial KV caches within one NVLink-connected system. Aggregate capacity does not turn eight devices into one flat block of memory; software still has to partition model weights, activations and communication correctly. The practical benefit is that tensor and pipeline parallel jobs can remain inside a high-bandwidth eight-GPU domain before they cross the external cluster fabric.
Two fifth-generation NVLink switches provide 14.4 TB/s of aggregate NVLink bandwidth. That internal topology is the main architectural difference between DGX B300 and a server populated with eight independent PCIe cards. PCIe GPUs can be excellent for many independent inference replicas, rendering jobs or smaller training runs. DGX becomes more attractive when a single job needs sustained, low-latency communication across all eight accelerators.
NVIDIA rates DGX B300 at 72 PFLOPS of sparse FP8 training performance and 144 PFLOPS of sparse FP4 inference performance. Those are peak tensor figures, not application throughput guarantees. The useful question for a buyer is whether the intended framework, numerical format, model architecture and batch or latency target can use those modes. Dense performance is half the sparse number where NVIDIA marks its convention. No procurement model should turn a peak PFLOPS value directly into tokens per second without a workload benchmark.
Host CPUs and system memory
Two Intel Xeon Platinum 6776P processors handle host-side work. The CPU layer prepares input, runs storage and network services, coordinates the accelerators and supports parts of the software stack that do not execute on GPU. It is not the headline component, but poor data preparation or underfed input pipelines can still leave expensive GPUs waiting.
DGX B300 ships with 2 TB of system memory and supports an upgrade to 4 TB using NVIDIA-qualified 128 GB DIMMs. Teams should size host memory around dataset transforms, CPU preprocessing, checkpoint handling, container density and orchestration overhead rather than using a fixed ratio to GPU memory. An inference platform serving many models may put very different pressure on host RAM from a training node reading large sequential shards.
The fixed DGX configuration reduces CPU and memory choice compared with some OEM HGX systems. That is deliberate. NVIDIA controls the validated platform, firmware and support matrix. A customer who needs a different host CPU count, a specific memory speed or a more unusual expansion topology should compare OEM HGX B300 offerings before ordering.
Local storage: fast cache, not the whole data platform
Eight 3.84 TB E1.S self-encrypting NVMe drives provide the local cache tier. Two 1.92 TB M.2 NVMe devices hold the operating system and support software encryption. The switch from U.2 drives used in earlier DGX generations to E1.S keeps the local cache serviceable in a dense front-access layout.
Local NVMe is useful for hot dataset shards, checkpoints, container layers, compiled artefacts and temporary output. It can lower repeated reads from shared storage and shorten restart time after a job lands on the same node. It is not a substitute for the cluster storage design. A multi-node training job can consume data far faster than a single shared file service can supply it, especially when all eight GPUs per node enter the same training phase together.
A serious deployment should model at least four storage flows: initial dataset ingest, steady-state training reads, checkpoint writes and recovery after interruption. Capacity alone is not enough. The plan needs aggregate throughput, metadata behaviour, redundancy, network paths and the time required to refill local cache after maintenance. The GPUMachines product record keeps the eight E1.S cache devices separate from the two M.2 boot devices so these roles are visible in the quote.
The network is part of the machine
DGX B300 presents eight OSFP ports from eight single-port ConnectX-8 VPI adapters. Each port supports up to 800 Gb/s InfiniBand or Ethernet. NVIDIA connects these adapters directly to the GPU tray over PCIe Gen6, giving the system one high-speed external rail per GPU. That layout is built for scale-out collectives, not merely for conventional server traffic.
Two dual-port BlueField-3 DPUs add four QSFP112 ports, each supporting up to 400 Gb/s InfiniBand or Ethernet. NVIDIA describes these as storage and management network connections. Separate 1 GbE RJ45 interfaces serve host management and the BMC.
Those port counts have immediate commercial consequences. One DGX B300 can require eight 800 Gb/s fabric switch ports, plus the selected BlueField-3 connections, optics or copper assemblies, and management cabling. A cluster quote that lists only the servers is incomplete. The bill of materials should identify switch models, occupied ports, cable media, lengths, breakout behaviour, spare-port policy and the mapping from each GPU rail to the fabric.
Ethernet and InfiniBand versions also create different operating requirements. Spectrum-X Ethernet may suit organisations standardising on Ethernet and RoCE, while Quantum InfiniBand remains familiar in tightly coupled HPC and AI fabrics. Either can work, but they should not be mixed casually. Congestion control, subnet management, telemetry and cable qualification all need owners.
Rack power and cooling reality
NVIDIA states 14.5 kW power consumption for the system and a 15 kW maximum. The AC version uses twelve power inlets at 200 to 240 V and twelve front-loaded power supplies in an N+N configuration. A 54 V DC busbar version is available for compatible racks, but the power shelves that energise the busbar are separate infrastructure.
The AC system weighs about 168 kg before packaging. NVIDIA cites airflow of 1,500 CFM at the stated fan setting and a maximum heat output of roughly 49,476 BTU per hour. These figures change the rack conversation. Four 10U nodes fit geometrically in 40U, yet four systems could demand around 58 kW in operation and roughly 60 kW at maximum. Many conventional enterprise racks cannot supply or cool that load.
Power design needs A and B feeds, breaker and PDU headroom, connector compatibility, maintenance states and the effect of one failed feed on the remaining path. Cooling design needs inlet temperature, front-to-rear airflow, containment, pressure and a credible heat-rejection figure. The system supports power capping, which can help a site stay within an electrical envelope, but a cap may reduce application performance. It is a control mechanism, not a substitute for proper facility capacity.
NVIDIA requires installation by partner-network personnel or its field service engineers. Ignoring that requirement can void the hardware warranty. Floor loading, lifting equipment, rail compatibility, shipping access and service clearance should therefore be settled before delivery day.
Software and support boundary
DGX OS 7 is based on Ubuntu 24.04 LTS. NVIDIA also documents support for Ubuntu, Red Hat Enterprise Linux 8 and 9, and Rocky Linux. The product is positioned with NVIDIA AI Enterprise software and Mission Control for AI data-centre operations and orchestration, including Run:ai technology.
The value of DGX is partly organisational. Hardware, firmware, drivers, libraries and the reference software stack sit inside a support model designed around NVIDIA's system. That can shorten fault isolation when a problem crosses several layers. It can also reduce local freedom: firmware cadence, approved DIMMs, operating-system combinations and service procedures follow the DGX support matrix.
Check the quote for software entitlement, support level, support duration, installation, remote monitoring, replacement terms and renewal pricing. Do not assume a line that says NVIDIA AI Enterprise explains every licence right. The commercial schedule should state what is included, for how long, and how the entitlement changes if the server is moved to another site.
Workloads that justify DGX B300
Large model training is the obvious case. The eight-GPU NVLink domain reduces the amount of communication that has to leave the node, while FP8 support and large HBM capacity suit modern mixed-precision training. Fine-tuning jobs with long context, larger batches or memory-hungry optimiser states can also benefit, even when they do not need a large multi-node cluster.
Reasoning inference is the second major use. FP4 tensor performance and 2.3 TB of aggregate HBM3e can support large models, substantial KV cache and multiple concurrent workloads. Actual latency and throughput depend on model structure, quantisation, serving framework, scheduler and user concurrency. Published peak compute is not a substitute for a representative prompt and output-length test.
Scientific computing, multimodal training and simulation workloads may fit when their communication patterns map well to the NVLink and external fabric. The node can also consolidate several smaller jobs, but shared scheduling and isolation have to be planned. Buying an eight-GPU appliance to run eight unrelated low-duty tasks can work technically; it may still be a poor financial use of the platform.
When not to buy it
Do not buy DGX B300 simply because it is the current flagship. A workstation, a two- or four-GPU PCIe server, DGX Spark-class development system or rented cloud capacity may be better when workloads are intermittent, models fit on one GPU, staff are still evaluating frameworks, or the facility cannot accept a 15 kW node.
An OEM HGX B300 system deserves comparison when the buyer needs liquid cooling, higher rack density, different host CPUs, additional local storage or an existing service contract with Dell, HPE, Lenovo, Supermicro, GIGABYTE, ASUS, ASRock Rack or MSI. The silicon may be similar, but the chassis, cooling, management and support experience are not interchangeable.
DGX B200 remains relevant where it is already qualified, available at a meaningful discount or sufficient for the model roadmap. A forced refresh can cost more in revalidation and downtime than it returns in throughput. Teams should compare measured job completion time and useful tokens per watt, not generation names alone.
Price and total cost of ownership
NVIDIA does not publish a list price for DGX B300; partner quotes determine the transaction. Public US reporting places retail close to $400,000, while configuration, support, region and supply can move the number. GPUMachines currently uses $399,000 as a planning starting price, not as a binding offer.
The server price is only the first line. Add fabric switches and ports, optics or cables, management switching, rack PDUs or busbar shelves, installation, support, shared storage, scheduler software and facility power. At a sustained 14.5 kW, one system consumes about 127 MWh over a full year before cooling overhead. Real utilisation will vary, but even a partial-duty estimate should include the site's electricity rate and PUE.
Idle time matters more than a small purchase-price discount. A cheaper system used at 20 per cent can cost more per completed job than a well-scheduled node bought at a premium. Before ordering, build a utilisation plan covering named workloads, expected queue depth, model owners, maintenance windows and overflow to cloud.
Procurement checklist
- Confirm that representative models need the eight-GPU NVLink domain and 2.3 TB aggregate HBM3e.
- Benchmark the intended training or serving stack, including precision, context length and concurrency.
- Choose Ethernet or InfiniBand and produce a port-by-port fabric bill of materials.
- Size shared storage for aggregate node throughput, checkpoint traffic and recovery.
- Verify 200-240 V AC distribution or the complete 54 V DC busbar design.
- Reserve 15 kW per node for maximum-power planning and check heat rejection.
- Confirm rack depth, rail compatibility, floor loading, lifting and service access.
- Record installation responsibility so warranty requirements are met.
- Read the software and support schedule, including renewal and replacement terms.
- Check export-control, delivery-country and end-use requirements before allocation.
Frequently asked questions
Is DGX B300 the same as an HGX B300 server?
No. HGX B300 is the eight-GPU accelerator platform that OEMs integrate into their own servers. DGX B300 is NVIDIA's complete appliance with specified CPUs, memory, storage, networking, chassis, software and support. The accelerator generation is related, but the products and service boundaries are different.
How much GPU memory does DGX B300 have?
The system guide specifies eight GPUs with 288 GB of HBM3e each, or 2.3 TB aggregate. NVIDIA's product page displays a rounded 2.1 TB figure. Memory remains distributed across eight GPUs and is coordinated through the application and NVLink fabric.
What power should a rack planner use?
Use 14.5 kW as the stated operating figure and 15 kW as the maximum system budget. A design should also include cooling and distribution overhead rather than treating 14.5 kW as the complete facility load.
Are the B300 GPUs selectable in the configurator?
No. DGX B300 ships with a fixed eight-GPU SXM complex. The GPUMachines configurator records the board as pre-installed and includes its cost in the starting system price. Individual PCIe GPU choices would misrepresent the product.
Does the system include the network switches and cables?
No. The server includes ConnectX-8 and BlueField-3 adapters, but external switches, transceivers and cables depend on the selected topology and are quoted separately.
Sources and evidence boundary
Primary specifications come from the NVIDIA DGX B300 product page and the NVIDIA DGX B300 System User Guide. The user guide is also the source for physical dimensions, weight, power, airflow, storage topology and installation details. NVIDIA's quick-start guidance states that partner or NVIDIA field installation is required for warranty coverage.
All performance numbers in this review are NVIDIA peak specifications, not GPUMachines benchmark results. Prices are planning figures subject to a formal quote.