Nvidia GB300 Specifications (including memory bandwidth and LLM benchmarks) based on 2026 systems

The market entry this past March of Nvidia’s GB300 based DGX Station offers enterprise customers an opportunity to engage with previously unavailable levels of AI performance directly within their own infrastructure. 

Nvidia GB300 systems like the ASUS ExpertCenter Pro ET900N G3 (which I will be testing in a coming article) bring data-center class specifications into tenable desktop formats, enabling technologists and teams to test existing models without external cloud dependencies or associated subscriptions and with minimal integration effort. 

Beyond performance benchmarking, maintaining sensitive data, and intellectual property entirely on-site, enterprises can evaluate secure deployment strategies that align with their firm’s data sovereignty goals. 

When it comes to deployment, the DGX Station is especially appealing when compared to other solutions due to a lower operational complexity and professional enterprise stack that provides many data center features without the overhead of data center infrastructure.

The models up for pre-order now are powered by the “Nvidia GB300 Grace Blackwell Ultra Desktop Superchip”. This combines a 72-core ARM-based Grace CPU with a Blackwell Ultra GPU right inside a (typically large-ish) desktop tower format. A standard rig can provide up to ~20 FP4 PFLOPS (not accounting for additional GPUs, thermal constraints)

GB300 Hardware Specifications

The core of the aforementioned Nvidia GB300 package is kitted with 748 GB coherent memory pool split between the GPU/CPU cores (see below). Currently this remains identical across all models and vendors. However, original equipment manufacturers (OEMs) differentiate their SKUs through integration. Ex: OS, case acoustics/cooling, networking ports, and additional GPUs.

• Processor Architecture: GB300 Grace Blackwell Ultra Desktop Superchip
• Maximum Model Size (Inference): Up to 1 trillion parameters (using FP4 or quantized formats)
• Maximum Model Size (Fine-Tuning): 200 to 400 billion parameters (using LoRA/QLoRA)
• Total Coherent Memory Pool: 748 GB (Integrated CPU and GPU memory space)
• GPU Memory (VRAM): 252 GB HBM3e
• GPU Memory Bandwidth: 7.1 TB/s
• CPU Memory (System RAM): 496 GB LPDDR5X
• Interconnect Technology: NVLink-C2C (Chip-to-Chip)
• Interconnect Bandwidth: 900 GB/s bidirectional
• Form Factor: Desktop workstation

GB300 Performance Matrix @ 1600 Watts

Given the integration constraints, the theoretical peak performance for all OEM devices should be similar.

FP4 Tensor Core: 20 PFLOPS (sparse) / 15 PFLOPS (dense)
INT8 Tensor Core: 330 TOPS
FP8 / FP6 Tensor Core: 10 PFLOPS
FP16 / BF16 Tensor Core: 5 PFLOPS
TF32 Tensor Core: 2.5 PFLOPS
FP32: 80 TFLOPS
FP64 / FP64 Tensor Core: 1.3 TFLOPS

GB300 Benchmarks on big LLMs

Model / Workload	Tok/Sec
Kimi 2.5, 1.1T	40-50 (station total)
Nemotron Ultra, 550B	~35 / single request. Scales to 4-5.
GLM-5.2-REAP 504B	~60

GB300 Memory Architecture Explained

Perhaps the most defining GB300 feature is the “dual-tier unified memory architecture” that can run enormous models, but create a performance caveat depending on model size (weights + context). On the primary tier, the system features HBM3e GPU memory having 7.1 TB/s of bandwidth, which is high performance silicon for embedding, inference, etc. tasks.

Beyond this, it uses an NVLink-C2C coherent bridge to access standard/slower LPDDR5X memory at 900 GB/s, creating an overflow pool for large workloads. When a quantized model, such as a 100B model at FP4 or FP8 precision, fits entirely within the GPU pool, it accesses the maximum 7.1 TB/s bandwidth to generate tokens at a fantastic clip.

However, when a model’s size exceeds this threshold (eg: a 100B unquantized FP32 or Llama 3.1 at 405B), the system will access the overflow data across the coherent bridge; this can negatively influence the token generation for those weights subject to the 900 GB/s limit, resulting in a drop in (Ex: inference) speeds for parameters stored outside the GPU-only RAM. On the plus side, with proper configuration, tuning MoE models can allow for running previously unattainable workloads with minimal penalties when active and inactive experts (potentially with different quantizations) fit within their respective (HBM3e/LPDDR5X) RAM boundaries.

DGX Station Products by OEM

At the time of writing, these are the currently launched instantiations:

• ASUS – ExpertCenter Pro ET900N G3
• Dell – Dell Pro Max with GB300
• Exxact – Valence / TensorEX stations
• Gigabyte – W775-V10-L01
• HP – ZGX Fury
• MSI – XpertStation WS300
• Supermicro – Super AI Station

---