English
The MSI XpertStation WS300 leverages NVIDIA’s latest GB300 Grace Blackwell Ultra DGX Station architecture, delivering a desk-side supercomputing node with 20 petaFLOPS of FP4 sparse compute. It pairs a 72-core Grace Arm CPU and Blackwell Ultra B300 GPU via a 900GB/s bidirectional NVLink‑C2C interconnect, sharing a unified 748GB coherent memory pool: 252GB high‑bandwidth HBM3e GPU memory and 496GB LPDDR5X system memory, sufficient to host trillion‑parameter AI models locally. Integrated ConnectX‑8 SuperNIC hardware delivers dual 400GbE ports for 800Gb/s cluster fabric bandwidth, bringing true data center‑grade AI performance to office and desktop environments.
Core Specifications
Architecture: NVIDIA GB300 Grace Blackwell Ultra
CPU: 72‑core Arm Neoverse V2 Grace processor
GPU: Blackwell Ultra B300 (5th‑gen Tensor Cores, 208B transistors)
Interconnect: NVLink‑C2C (900GB/s bidirectional)
Unified Memory: 748GB coherent pool (252GB HBM3e @7.1TB/s; 496GB LPDDR5X @396GB/s)
Storage: 2×2TB PCIe 5.0 SSDs (RAID 1 boot), 2×open PCIe 6.0 M.2 expansion slots (Micron 4600 Gen5 SSD validated)
Networking: ConnectX‑8 SuperNIC (800Gb/s), 10GbE host port, 1GbE BMC management port
Expansion: 1×PCIe 5.0 x16, 2×PCIe 5.0 x8 slots (supports RTX PRO Blackwell series GPUs)
Cooling: Custom liquid cooling (1400W CPU+GPU rating), dual 360mm radiators
Power: 1600W 80 PLUS Platinum PSU, dedicated 20A circuit required for full output
Management & Security: AST2600 BMC (IPMI/Redfish), TPM 2.0, hardware root of trust, chassis intrusion detection
Design & Platform Architecture
MSI builds the complete desk-side workstation around NVIDIA’s fixed GB300 core platform, handling chassis design, liquid cooling, power delivery, I/O layout, and serviceability while NVIDIA provides the Grace Blackwell baseboard and enterprise software stack. The system features a professional compact form factor, with internal engineering optimized to sustain peak performance of the advanced GB300 superchip.
The flagship B300 GPU uses TSMC 4NP dual‑die design linked via 10TB/s NV‑HBI inter‑die interface, integrating 160 streaming multiprocessors and 640 fifth‑gen Tensor Cores. As a data center‑grade accelerator, it lacks consumer display outputs and NVENC encoders but includes seven NVDEC/NvJPEG decoders and supports MIG partitioning into seven isolated GPU instances for multi‑workload parallel development. Its multi‑precision compute peaks at 20 PFLOPS (NVFP4 sparse), 10 PFLOPS (FP8), 330 TOPS (INT8), and 80 TFLOPS (FP32).
Complementing the GPU, the 72‑core Grace Arm CPU features Armv9 architecture, 128‑bit SVE2 vector units, 64KB L1 cache per core, 1MB private L2 cache, and 114MB shared L3 cache. Its 3.2TB/s coherent mesh fabric unifies CPU cores, memory, I/O, and NVLink‑C2C, eliminating traditional CPU‑GPU data transfer bottlenecks and handling data preprocessing, tokenization, and model orchestration to fully feed the B300 accelerator.
Upgradeable SOCAMM LPDDR5X modules deliver power efficiency and easy field replacement. The NVLink‑C2C coherent memory architecture is the platform’s key advantage: unlike discrete workstations with separate CPU/GPU memory spaces, Grace and B300 share a single address space. HBM3e stores high‑priority model weights and tensors, while capacious LPDDR5X acts as extended memory for oversized models, avoiding multi‑GPU segmentation for large workloads.
The I/O topology centers on the Grace CPU, with direct PCIe 5.0 boot storage and full‑length expansion slots. A pre‑wired 12VHPWR connector and anti‑sag bracket support heavy RTX PRO add‑in GPUs. The ConnectX‑8 SuperNIC functions as a secondary I/O hub, controlling PCIe 6.0 storage expansion, Wi‑Fi 7/Bluetooth, and 10GbE networking. Front/rear I/O covers high‑speed USB, audio, BMC management, and high‑bandwidth networking for versatile connectivity.
Power & Thermal Management
The single 1600W Platinum PSU powers all system components, with a shared power budget for the GB300 chip and optional RTX PRO GPU. NVIDIA’s vsloshd dynamic power service redistributes power headroom in real time, prioritizing add‑in RTX cards and throttling B300 performance when power is constrained. A hardware power brake protects system stability under degraded power delivery, throttling output instead of shutting down entirely.
MSI’s custom liquid cooling loop cools the B300 GPU, Grace CPU, SOCAMM memory, and ConnectX‑8 hardware, supported by dual 360mm radiators and a chassis system fan. Rated for 1400W combined CPU/GPU thermal load, the system maintained stable temperatures (71°C GPU peak, 65°C CPU peak) during full‑power 1292W stress testing.
Key Workflow Capabilities
The WS300 transforms enterprise AI development by bringing cluster‑grade workflows to local workstations, eliminating shared resource scheduling delays. Its MIG partitioning enables isolated multi‑instance GPU testing for multi‑GPU workload validation and tool development. It unifies physical AI development loops, supporting robot demonstration capture, Isaac Sim synthetic data generation via optional RTX PRO GPUs, and GR00T model fine‑tuning on the B300 in one device.
Its greatest strength is Blackwell Ultra kernel optimization: developers can profile, tune, and iterate CUDA/Triton kernels directly on target B300 hardware locally, reserving large cluster systems only for final scaled validation. This drastically improves development efficiency for AI framework and inference engine teams.
Performance Benchmarks
MAMF Compute Performance
Against Blackwell RTX PRO 6000 and DGX Spark systems, the WS300’s GB300 delivers dominant multi‑precision compute: 1967 TFLOPS (BF16), 3909 TFLOPS (FP8), and 6134 TFLOPS (NVFP4). It achieves a 4–5x performance lead over RTX PRO 6000 and 17–19x over DGX Spark across all precisions.
Memory & NVLink Bandwidth
B300 local HBM3e hits 6875GB/s read and 6009GB/s write throughput, near its 7.1TB/s rated bandwidth. Cross‑NVLink transfers to Grace LPDDR5X top out at 390GB/s, limited by system memory bandwidth rather than the NVLink‑C2C interface. Bidirectional concurrent transfers reach 253GB/s, defining the practical cross‑memory data movement ceiling.
LLM Inference Performance
The 748GB coherent memory pool enables seamless inference across model sizes, from HBM‑resident lightweight models to large architectures requiring cross‑memory offloading. Compact models like DeepSeek v4 Flash and MiniMax M2.7 deliver extreme throughput, scaling to 3532–9602 tokens/second at high concurrency. Borderline HBM‑fit MiniMax M3 offloads minor weights to Grace memory with steady throughput scaling.
Oversized models including 433GB GLM‑5.2 and 550B Nemotron‑3‑Ultra rely heavily on LPDDR5X offloading, operating stably though limited by cross‑NVLink data migration. In head‑to‑head testing on shared models (GPT‑OSS, Llama 3.1, Mistral Small, Qwen3 Coder), the WS300 delivers 2.3–3x higher throughput than RTX PRO 6000 and 14–19x over DGX Spark at full concurrency, with a 6x advantage for long‑context workloads that stress KV cache capacity.
GDS Storage Performance
Paired with Micron 4600 Gen5 SSDs, the system’s GPU Direct Storage bypasses CPU memory staging, streamlining model loading and checkpoint I/O. Sequential read throughput peaks at 13.1GiB/s and writes at 12.0GiB/s, scaling efficiently across thread counts and block sizes to eliminate storage bottlenecks for AI dataset streaming and model checkpointing.
Target Users & Limitations
The WS300 is purpose-built for AI labs, kernel/compiler developers, robotics teams, and enterprise AI groups needing exclusive local Blackwell Ultra access to avoid cluster scheduling bottlenecks. Its integrated MIG, remote BMC management, liquid cooling, and full enterprise software stack deliver turnkey data center functionality for desk-side development.
Key limitations include Arm‑based host compatibility restrictions for some x86 toolchains, mandatory 20A circuit installation, no native display output, and shared power budget constraints when using add‑in RTX PRO GPUs. Economically, 1–2 units per team deliver optimal value; larger deployments shift cost-effectiveness to full rack B300 cluster setups.
Conclusion
The MSI XpertStation WS300 is a class‑leading desk-side supercomputer, enabling single‑user development on frontier trillion‑parameter models previously only accessible via data center clusters. MSI’s robust OEM implementation delivers stable thermal and power management, reliable expandability, and enterprise serviceability on NVIDIA’s flagship GB300 platform. It fills a critical niche for iterative Blackwell Ultra AI development, complementing large-scale cluster infrastructure and earning placement in top local AI and agentic hardware rankings.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!