NVIDIA

B300

SXM

B300 SXM — illustration of the card's form factor
VRAM
288GB
Architecture
Blackwell Ultra
Form Factor
SXM

Provider Marketplace

Cheapest
$3.75/hour
Starting from
Verda logo
VerdaVisit
Best Value
$6.94/hour
Starting from
RunPod logo
RunPodVisit
Enterprise Choice
$7.40/hour
Starting from
Hyperstack logo
HyperstackVisit

All Cloud Providers

6 Options available
Verda logo
VerdaCheapest
Spot · preemptible
$3.75/ hour
Estimated Cost
Provision
Fal logo
Fal
On-Demand
$4.49/ hour
Estimated Cost
Provision
Massed Compute logo
$6.60/ hour
Estimated Cost
Provision
RunPod logo
RunPod
On-Demand
$6.94/ hour
Estimated Cost
Provision
Modal logo
Modal
On-Demand
$7.10/ hour
Estimated Cost
Provision
Hyperstack logo
Hyperstack
On-Demand
$7.40/ hour
Estimated Cost
Provision

Estimates only — rates are collected automatically from public provider pages and may be out of date. Prices vary by region, commitment term, and availability, and typically exclude storage, egress, and tax. Confirm current pricing with the provider before purchasing. Last collected .

Architecture

MicroarchitectureBlackwell Ultra
Dynamic PrecisionSupported (FP4/FP8)

Memory & VRAM

Total Capacity288GB

Connectivity & Scaling

InterconnectNVLink
Generation5th generation NVIDIA NVLink
PCIe InterfacePCIe Gen6
TopologyNVLink Switch
Max GPUs/Node8
Scale-OutInfiniBand/Ethernet (ConnectX-8 up to 800 Gb/s; BlueField-3 up to 400 Gb/s)

Physical Design

Form FactorSXM
CoolingActive

Software Ecosystem

DockerDocker Engine; NVIDIA Container Toolkit
Kernel Optimoptimized Linux kernel

Server & Deployment

OEM AvailabilitySystems are shipping / Available now
PreconfiguredNVIDIA DGX SuperPOD; NVIDIA DGX BasePOD
Rack-ScaleDeployable in NVIDIA MGX racks; compatible with traditional enterprise racks
Ref ArchitecturesNVIDIA DGX BasePOD (reference architecture); NVIDIA DGX SuperPOD (turnkey, configurable with DGX systems)

System Compatibility

Required PCIePCIe Gen6
Rack PowerAC: 200-240 VAC, 12 AC power inlets; PSU: 3.3 kW @ 200-240 V, 16 A (12 PSUs, N+N redundancy; minimum 6 PSUs required; system boots only if at least 3 PSUs are operational); Busbar: 54 VDC, 300 A; Total power consumption: 14.5 kW.
OS CompatDGX OS 7 based on Ubuntu 24.04 LTS; additional support for Ubuntu and Red Hat Enterprise Linux 8 and 9, Rocky Linux

Multi-GPU Scalability

Scaling Characteristics

ParallelismEnables DeepSeek-R1 (671B MoE) inference on fewer GPUs with lower tensor parallelism overhead

Workload Readiness

LLM Training

supported

LLM Inference

supported

Real-Time Serving

supported

Market Authority

MLPerf Ranking

Highest throughput in MLPerf Inference v6.0 (April 2026)

Supercomputer Usage

Used in NVIDIA DGX SuperPOD deployments (example: Lilly’s DGX SuperPOD with DGX B300 systems)

Research Citations

Document references SemiAnalysis InferenceX benchmarks and MLPerf Inference v6.0

Community Benchmarks

SemiAnalysis InferenceX benchmarks and MLPerf Inference results cited for performance and cost claims

Enterprise Cases

Lilly — deployment of NVIDIA DGX SuperPOD with DGX B300 systems for drug discovery

Key Strengths

Limitations

Also in the Lineup

NVIDIA
A10
NVIDIA
A100 80GB SXM
NVIDIA
A100 80GB PCIe
NVIDIA
A10G
NVIDIA
A16
NVIDIA
A30
NVIDIA
A40
NVIDIA
B200 SXM

Expert Insight

The B300 represents a strategic leap in AI compute. When comparing cloud providers, consider not just the hourly rate, but also the interconnect bandwidth (InfiniBand/NVLink) and regional availability which can significantly impact total cost of ownership for large-scale training.

Glossary Terms

FP32 TFLOPS
VRAM
TDP
Cores
Information updated daily. Cloud pricing subject to vendor availability.