SLYD
Read first

The 5-step pipeline.

How energy turns into a deployed, financed, offtake-matched cluster, and what SLYD does at every step.

How SLYD works →
Hardware

New and recovered GPU systems.

NVIDIA and AMD systems through documented manufacturer and qualified channel supply, with financing and deployment coordinated on the same platform.

Explore GPU hardware →
Marketplace

Compute, hardware, and power in one book.

Browse available GPU capacity by accelerator, configuration, region, and price, or bring supply to qualified demand.

Open marketplace →
Pre-qualify

Start with an indicative structure.

Tell us deal size, structure, and offtake. Any range is preliminary and subject to underwriting, diligence, and documentation.

Open Configure →
From the blog

GPU market trends and deployment playbooks.

Infrastructure best practices, hardware comparisons, and industry analysis from the SLYD team.

Read the blog →

Library

NVIDIA rack-scale systems

NVIDIA Grace Blackwell systems

GB200 and GB300 are bought as racks, not as cards. This page separates the two platforms cleanly, keeps NVIDIA's sparsity basis attached to every number, and works through what the rack asks of the building.

What are NVIDIA Grace Blackwell systems?

NVIDIA Grace Blackwell systems combine NVIDIA Grace CPUs and Blackwell GPUs in tightly integrated AI platforms. The correct choice depends on the exact platform generation, rack or server configuration, workload, memory and interconnect requirements, site power and cooling capacity, software stack, and verified commercial availability.

Architecture

Why the CPU is part of the accelerator decision

In a conventional GPU server, CPUs and accelerators are separate subsystems joined over PCIe. Grace Blackwell integrates an Arm-based NVIDIA Grace CPU with Blackwell GPUs over a coherent link, and then joins many of those units into a single NVLink domain across the rack.

The practical consequence is that the scale-up domain becomes much larger than one server. A job that needs many accelerators to behave as one tightly coupled unit sees a rack rather than a node. That is the reason to consider the platform, and it is also the reason the rack becomes a facility project.

Building block

GB200 Grace Blackwell Superchip

NVIDIA publishes 1 Grace CPU with 2 Blackwell GPUs, 372 GB HBM3E at 16 TB/s, 3.6 TB/s of NVLink bandwidth, 72 Arm Neoverse V2 cores, and up to 480 GB of LPDDR5X CPU memory.

Rack system

NVL72 rack-scale systems

72 GPUs and 36 Grace CPUs joined in one NVLink domain with 130 TB/s of NVLink bandwidth, published by NVIDIA for both GB200 NVL72 and GB300 NVL72.

Smaller form

NVL4 and OEM variants

Manufacturers publish smaller Grace Blackwell configurations as well. Dell publishes PowerEdge XE8712 as a GB200 NVL4 system with 4 NVIDIA B200 GPUs and 2 Grace processors in a 1OU form factor.

Platform comparison

GB200 NVL72 and GB300 NVL72

These are two different platforms with two different memory profiles, and NVIDIA publishes their compute figures on different sparsity bases. The table keeps each on its own published basis rather than normalizing them into a single column.

Source: NVIDIA GB200 NVL72 and GB300 NVL72 product pages, reviewed 18 August 2026. NVIDIA footnotes GB200 NVL72 figures as sparse then dense, and states that all GB300 NVL72 Tensor Core specifications are with sparsity unless otherwise noted.
Specification GB200 NVL72 GB300 NVL72
Configuration 36 Grace CPUs, 72 Blackwell GPUs 36 Grace CPUs, 72 Blackwell Ultra GPUs
GPU memory and bandwidth 13.4 TB HBM3E, 576 TB/s 20 TB, up to 576 TB/s
Fast memory Not published as a separate figure 37 TB
CPU memory and bandwidth 17 TB LPDDR5X, 14 TB/s 17 TB LPDDR5X, 14 TB/s
CPU cores 2,592 Arm Neoverse V2 cores 2,592 Arm Neoverse V2 cores
NVLink bandwidth 130 TB/s 130 TB/s
FP4 Tensor Core 1,440 PFLOPS sparse, 720 PFLOPS dense 1,440 PFLOPS with sparsity, 1,080 PFLOPS without sparsity
FP8 or FP6 Tensor Core 720 PFLOPS sparse 720 PFLOPS with sparsity
FP16 or BF16 Tensor Core 360 PFLOPS sparse 360 PFLOPS with sparsity
Cooling Confirm with the specific configuration NVIDIA describes a fully liquid-cooled rack-scale architecture

SLYD does not publish a generational performance multiplier for these platforms. NVIDIA's own comparative claims come with workload, precision, software, and configuration context, and that context does not survive being compressed into a single number.

Selection

Which platform fits the requirement

Consider GB300 NVL72

When memory capacity is the binding constraint

Long-context inference, very large models, and workloads where the working set does not fit comfortably at the GB200 memory profile. GB300 publishes 20 TB of GPU memory and 37 TB of fast memory against 13.4 TB on GB200 NVL72.

Consider GB200 NVL72

When the NVLink domain is what you need

Training and inference that benefit from a 72-GPU scale-up domain without needing the larger memory profile. Both platforms publish the same 130 TB/s of NVLink bandwidth.

Consider a smaller form

When a full rack is more than the workload needs

NVL4-class and other OEM Grace Blackwell configurations bring the architecture into a conventional server footprint, which is often the practical option for sites without rack-scale liquid cooling.

Consider an 8-GPU server

When the CPU coupling is not the point

If the design does not depend on the integrated Grace CPU or the rack-scale NVLink domain, an HGX B200 or B300 server is usually the simpler deployment.

Manufacturer implementations

Who builds Grace Blackwell systems

The examples below are systems the manufacturers publish in their own current documentation, reviewed 18 August 2026. They are named to show where the architecture appears, not to indicate that any particular system is available.

Manufacturer-published Grace Blackwell implementations. Confirm the current configuration with the manufacturer before relying on any detail.
Manufacturer Published system As described by the manufacturer
Dell PowerEdge XE9712 GB300 NVL72 system, 48U, 36 Arm-based NVIDIA Grace processors and 72 NVIDIA B300 GPUs
Dell PowerEdge XE8712 GB200 NVL4 system, 1OU, 2 Arm-based NVIDIA Grace processors and 4 NVIDIA B200 GPUs
Lenovo Lenovo NVIDIA GB300 NVL72 Rack-scale AI system published in Lenovo's AI server documentation
GIGABYTE NVIDIA GB300 NVL72 GIGAPOD Rack-scale solution published with direct liquid cooling
Supermicro GB200 NVL72 and GB300 NVL72 solutions Published within Supermicro's NVIDIA Blackwell portfolio
Buyer diligence

Confirm before configuring

  • Whether the quote describes a superchip, a node, or a full rack
  • Exact platform generation and GPU variant
  • Published memory profile for that exact configuration
  • Cooling method and facility water requirements
  • Rack power requirement and connector design
  • Included networking, optics, and cabling
  • Software stack, licensing, and support scope
  • Installation, commissioning, and acceptance responsibilities
  • Seller authority and governing transaction documents
  • Export, end-use, and geographic restrictions
FAQ

Grace Blackwell questions

What is NVIDIA Grace Blackwell?

Grace Blackwell is NVIDIA's architecture pairing Arm-based NVIDIA Grace CPUs with Blackwell-generation GPUs over a coherent link, delivered mainly as rack-scale systems rather than as individual cards. The rack, not the accelerator, is what gets specified, powered, cooled, and installed.

What is the difference between GB200 NVL72 and GB300 NVL72?

Both are 72-GPU, 36-CPU rack-scale systems with 130 TB/s of NVLink bandwidth. GB200 NVL72 uses Blackwell GPUs and publishes 13.4 TB of HBM3E GPU memory. GB300 NVL72 uses Blackwell Ultra GPUs and publishes 20 TB of GPU memory and 37 TB of fast memory. The larger memory capacity on GB300 is the difference that most often decides between them.

What is the difference between a GB200 Superchip and a GB200 NVL72 rack?

The GB200 Grace Blackwell Superchip is one Grace CPU with two Blackwell GPUs, publishing 372 GB of HBM3E and 3.6 TB/s of NVLink bandwidth. GB200 NVL72 is the rack-scale system built from those building blocks, publishing 72 Blackwell GPUs, 36 Grace CPUs, and 130 TB/s of NVLink bandwidth. Quotes and specifications must state which one they describe.

Do Grace Blackwell racks require liquid cooling?

NVIDIA describes GB300 NVL72 as a fully liquid-cooled, rack-scale architecture. Facility water, heat rejection, water quality, CDU placement, redundancy, and serviceability therefore have to exist before the rack is delivered, and the specific requirements come from the exact OEM or NVIDIA configuration being purchased.

Can I compare the FP4 numbers for GB200 and GB300 directly?

Only if the sparsity basis is matched first. NVIDIA publishes GB200 NVL72 NVFP4 as 1,440 PFLOPS sparse and 720 PFLOPS dense, and GB300 NVL72 FP4 as 1,440 PFLOPS with sparsity and 1,080 PFLOPS without sparsity. Comparing a sparse figure to a dense figure produces a result that means nothing.

Which manufacturers build Grace Blackwell systems?

Server manufacturers publish their own Grace Blackwell implementations. Dell publishes PowerEdge XE9712 as a GB300 NVL72 system and PowerEdge XE8712 as a GB200 NVL4 system, Lenovo publishes a Lenovo NVIDIA GB300 NVL72 rack-scale system, and GIGABYTE publishes an NVIDIA GB300 NVL72 GIGAPOD direct-liquid-cooling rack. Configuration, service, and commercial terms differ by manufacturer.

Does SLYD have Grace Blackwell systems in inventory?

Product names on this page are descriptive and do not indicate availability. Whether a specific configuration can be obtained, in what quantity, for what geography, and on what terms is confirmed at the transaction level against current records, with the confirmation date and scope stated.

Sources for the specifications on this page

Specifications belong to the manufacturers named and are reproduced with their own caveats. They are not SLYD guarantees of performance, configuration, or availability.

Scope a Grace Blackwell deployment

Share the workload, target scale, site, and timeline. Rack-scale platforms are decided by facility capacity as much as by the compute requirement, so both are worked through together.

Page updated: August 18, 2026

Reconnecting to the server...

Please wait while we restore your connection

An unhandled error has occurred. Reload 🗙