MSI XpertStation WS300 Block Diagram and Topology
The claim to fame of the DGX Station systems is NVIDIA’s GB300 superchip. Combining an NVIDIA Grace CPU, a Blackwell Ultra GPU, and a ConnectX-8 SuperNIC, the GB300 is an extremely powerful processor configuration. This is essentially a server configuration that just barely fits inside a desktop workstation.
Quickly running down the specs here for anyone not familiar with GB300 or DGX Station, the Grace CPU is a 72-core chip based on Arm’s Neoverse-V2 CPU cores. NVIDIA introduced it as part of the Grace Hopper hardware generation in 2022. With NVIDIA’s 4-year CPU development cycle, it has remained Blackwell’s designated CPU partner. Vera intersects the Rubin GPU generation. NVIDIA uses a full Grace configuration, with all 72 CPU cores enabled.
For memory, Grace offers a 512-bit LPDDR5X memory bus. For GB300 systems, this is implemented as four 128-bit SOCAMM modules. The slim memory modules allow removable/swappable LPDDR5X, though as we noted earlier, DGX Station systems lock this down as a non-serviceable part. The WS300 and other systems come with 496GB of memory (with NVIDIA reserving the last 16GB), providing 396GB/s of memory bandwidth. We tested running LLMs from LPDDR5X because it offers more memory bandwidth than the DGX Spark/GB10 platforms.
For I/O, Grace has no peripheral I/O of its own. Instead, it keeps things generic with 64 PCIe Gen5 lanes. As a result, all USB connectivity comes through discrete USB controllers.
As for the GPU, DGX Station systems feature a slightly detuned B300 Blackwell Ultra GPU. This Blackwell Ultra configuration ships with 7 stacks of HBM3E memory enabled, giving the GPU access to 252GB of memory overall. In terms of bandwidth, this totals 7.1TB/second. Otherwise, in terms of compute throughput (the ultimate reason the DGX Station even exists), a single system can deliver 20 PFLOPS of spare FP4 throughput, 10 PFLOPS of FP8, or 330 TOPS of INT8.
The final piece of the silicon puzzle is NVIDIA’s ConnectX-8 SuperNIC. A staple of GB300 servers (and replacing the CX7 in GB200 servers), CX8 is both a high-speed 48-lane PCIe Gen6 switch and a high-throughput NIC. On servers, CX8 Ethernet ports form the basis of GB300 scale-out networking. On DGX Station systems, they fulfill a similar role.
With NVIDIA’s standardized DGX Station motherboard, DGX Station systems share an equally standardized topology. As MSI’s diagram shows, all devices connect to the GB300 superchip, including the PCIe slots, two M.2 slots, and a USB controller. Most of Grace’s PCIe lanes are accounted for here, with 40 going to expansion slots in one form or another.
Meanwhile, the CX8 SuperNIC has its own tree of child devices. This includes two 400Gb Ethernet ports, two PCIe Gen6 x4 M.2 slots, the Marvell 10GbE controller, and the USB controller powering the sole 10Gbps USB-C port on top of the system.
As we discovered last month, both the CPU and the GPU connect to the CX8 as host devices. NVIDIA is taking full advantage of the chip’s advanced PCIe switching and multi-host capabilities, giving it direct access to both the CPU and the GPU even on DGX Station systems.
From the CPU, we found bandwidth figures consistent with a PCIe Gen5 x8 link. This allows for just 252Gbps of bandwidth on paper, and in practice we measured 228.9Gbps.
Meanwhile, the GPU has the fastest link of all to the CX8: a PCIe Gen6 x16 link. This fully leverages the PCIe Gen6 capabilities of both the Blackwell Ultra GPU and CX8 NIC to deliver a 968Gbps connection. This is vastly faster than the CPU connection, and rivals the bandwidth of the NVLink-C2C connection between the CPU and GPU.
An important topology feature is that it gives the GPU a faster, lower-latency connection to the NIC, as well as to all the devices hanging off it. That means the GPU can route traffic directly to the NIC, with the PCIe Gen6 x16 connection providing enough throughput to saturate both 400GbE links. Furthermore, it makes the two PCIe Gen6 x4 M.2 slots on the CX8 more useful than they first appear, because it lets the GPU read from them at the full 30GB/second bandwidth they offer, similar to how you would set up a GB300 server.
In this respect, Grace is almost a third wheel in the GB300 topology. With the Blackwell Ultra GPU as the critical component, the CX8 NIC is, in many ways, the GPU’s most important helper.
Next, let us cover out-of-band management.
MSI XpertStation WS300 IPMI with OpenBMC Management
Because DGX Station systems like MSI’s WS300 are built around a server accelerator, they don’t have native graphics support. You can add this with an RTX Pro video card, but it is not required. Instead, the company designed the systems to run headless (if desired), even building the platform with out-of-band management capabilities and a BMC, like a server.
Interestingly enough, MSI and NVIDIA are using OpenBMC here, so it will look either very familiar or moderately familiar based on which server brands you use and their BMC implementations.
OpenBMC is gaining popularity, and MSI’s solution feels like an industry-standard default implementation, but skinned for MSI. This feels different than other DGX Station systems we have tested.
OpenBMC also provides a lot of information on the different parts of the system that you can view from an inventory perspective directly in the BIOS. You probably rarely use this if you have one workstation or server, but if you have a fleet, it becomes useful and saves trips to the machine.
On the sensor readings side, you can see the various sensor values. I would say it feels like NVIDIA worked with all its DGX Station vendors to make sure this is available and that the machines can run 24×7 without issue. Still, given the price, I can imagine most buyers will be companies, so having IT manage these systems like servers and troubleshoot potential issues remotely using sensor data is useful.
In the event of an unwelcome power loss, the BMC IPMI lets you configure the power restore policy, whether to always boot the system, always keep it off, or restore its last power state.
It also offers a KVM-over-IP solution with an HTML5 iKVM viewer. It also includes a Java iKVM and remote SOL, but that is usually what people want to see.
We have been testing three of these DGX Station systems simultaneously, and we even put an NVIDIA RTX Pro 2000 Blackwell in one of them. If you have one of these, you might use it as a desktop. Many who have one of these systems will manage it like a server via the CLI and the BMC. If your company is buying several DGX Station systems, you are more likely to manage them as servers, so a solid OpenBMC platform helps when you start scaling out. It is also useful if you need to do an OS recovery, which we did on this system.
Next, let us get to the performance.
