Skip to content
NVIDIA HGX B300 eight-GPU reference baseboard
NVIDIA HGX B300 baseboard reference with eight GPU cooling assemblies. This is an integrated SXM platform, not a standalone PCIe graphics card. Image source ↗

NVIDIA · Integrated GPU platform / module

NVIDIA HGX B300

US$454,250.00 Per unit International supply by quotation
Request a formal quotation
For our best available priceCall salesWhatsApp
NewExact priced item

Priced item: NVIDIA HGX B300 NVL8 eight-GPU baseboard, eight B300 SXM 270GB GPUs, NVLink Switch and ConnectX-8; 935-26287-0070-000; one complete baseboard

Seller SKU GW-NVIDIA-HGX-B300

Supplied unit: One HGX B300 NVL8 baseboard with eight B300 SXM 270GB GPUs, NVLink Switch and ConnectX-8. Baseboard assembly only.

Your written quotation confirms the USD price, shipping, any destination taxes or duties, warranty and payment terms.

Product overview

NVIDIA HGX B300 NVL8 eight-GPU baseboard, eight B300 SXM 270GB GPUs, NVLink Switch and ConnectX-8; 935-26287-0070-000; one complete baseboard. Supplied unit: One HGX B300 NVL8 baseboard with eight B300 SXM 270GB GPUs, NVLink Switch and ConnectX-8. Baseboard assembly only. NVIDIA model-level reference specifications — reference memory / capacity: 2.1 TB GPU memory aggregate, manufacturer-rounded figure; interface: Fifth-generation NVLink / NVLink Switch; form factor: Eight-GPU SXM baseboard platform. Confirm the exact OEM configuration and its host, power, cooling and interconnect requirements before ordering.

Configuration overview

This is an integrated GPU module or system platform with host-specific requirements. Confirm the exact module or complete-system scope, interconnect, power, cooling and supported software in your quotation.

Specifications & compatibility

NVIDIA REFERENCE SPECIFICATIONS

Technical specifications

Specifications describe the identified NVIDIA model or reference configuration. Board-partner designs, OEM firmware, laptop power limits and complete-system configurations can differ.

Product type
Eight-GPU baseboard platform; complete server configuration is separate
GPU configuration
8 x Blackwell Ultra SXM GPUs
Total GPU memory
2.1 TB, NVIDIA rounded platform specification
FP4 Tensor peak, platform
144 PFLOPS sparse / 108 PFLOPS dense
FP8 / FP6 Tensor peak, platform, sparse
72 PFLOPS
FP16 / BF16 Tensor peak, platform, sparse
36 PFLOPS
TF32 Tensor peak, platform, sparse
18 PFLOPS
FP32 peak, platform
600 TFLOPS
FP64 / FP64 Tensor peak, platform
10 TFLOPS
NVLink generation
Fifth generation; NVLink 5 Switch
GPU-to-GPU NVLink bandwidth
1.8 TB/s per GPU
Total NVLink bandwidth
14.4 TB/s
Networking bandwidth, reference platform
1.6 TB/s
GPU memory, latest reference architecture
288 GB HBM3e per GPU; 2,304 GB aggregate across eight GPUs
Memory bandwidth per GPU
Up to 8 TB/s
PCIe host connections, baseboard
8 x Gen5 x16 links and 1 x Gen4 x2 link
Reference GPU networking
8 x ConnectX-8 SuperNICs, up to 800 Gb/s per adapter

HGX is a baseboard/platform, not a standalone PCIe graphics card. Complete server CPUs, system RAM, storage, power and cooling are OEM-specific. Sparse Tensor figures require supported structured sparsity.

Published NVIDIA references use different GPU-memory figures: the product overview specifies 2.1 TB, while the latest HGX AI Factory reference architecture specifies eight 288 GB GPUs and 2,304 GB aggregate. Confirm exact installed GPU memory and usable capacity in the OEM quotation.

Capacity / memory
2.1 TB GPU memory aggregate, manufacturer-rounded figure
Interface
Fifth-generation NVLink / NVLink Switch
Form factor
Eight-GPU SXM baseboard platform

An integrated platform or module. Confirm its supported host, interconnect, power and cooling configuration.

Complete/integrated system enquiry, not an ordinary standalone PCIe board. Do not treat HGX, DGX, Grace Blackwell and individual GPU modules as interchangeable.

Official platform identity; quote complete supported configuration

Manufacturer reference ↗ (opens in a new tab)
Browse categories & filters
Filters & sort
Ordering the right configurationCompatibility, equipment condition and delivery details

Match the part to its host.

NVIDIA model names identify GPUs and product editions; they do not establish one universal board-vendor part number. Laptop GPUs are supplied within complete laptops. SXM modules, HGX/DGX systems and integrated CPU–GPU platforms require their supported host and interconnect.

Memory type, rank, transfer rate and population rules must match the server and CPU. A storage capacity does not establish NVMe, SATA or SAS compatibility. SSD and mechanical HDD media, firmware, sector format, carriers and security variants need separate checks.

For a precise quotation, send:

  1. Complete manufacturer/OEM part number and label, including suffix.
  2. Quantity, acceptable condition and delivery city/country.
  3. Server, motherboard, CPU or storage-backplane model.
  4. Application, required configuration and compatibility requirements.

Where a price is not displayed, call or WhatsApp for the current USD price. Confirm equipment condition, quantity, warranty, shipping and dispatch before ordering.

CONFIGURATION & QUOTATIONSNeed help choosing your hardware?Share your model, part number or project requirements.Request a quote

Tell us what you need

Our team will confirm the exact configuration, condition, delivery and warranty terms in your quotation.

Add purchase, rental or project details optional

We use these details to respond to your enquiry. Privacy policy. Fields marked * are required.

Prefer the detailed enquiry form? ↗

Call GPU WORLD WhatsApp