Today, we’re unboxing the Supermicro AS-8125GS-TNMR2, a high-performance GPU server designed for cutting-edge workloads like Artificial Intelligence (AI), Machine Learning (ML) and High-Performance Computing (HPC). But there's a twist! This system is a pure AMD thoroughbred equipped with the latest AMD Genoa CPUs and AMD Instinct MI300X accelerators – state-of-the-art chips designed to take your computational capabilities to the next level.
Let's open the box and see what this system is about.
Step 1: Opening the Box
The Supermicro AS-8125GS-TNMR2 arrives in robust, industrial-grade packaging on a pallet, designed to protect the hardware during transit. Before opening, it’s important to check for damage and report any issues.
As we lift open the box, the 8U rackmount chassis makes an immediate bold impression with a minimalist design emphasising air intake surface area—perfectly engineered for high-density GPU workloads.
In the accessory box, we find power cables, miscellaneous brackets and screws.
Step 2: A Closer Look at the AS-8125GS-TNMR2
The front of the chassis is divided into two trays. The GPU tray, indicated by the sleek hexagonal open mesh on the top half, houses the accelerators. The lower half is the compute tray, home to the storage, mainboard and processors.
The rear of the system is dominated by fans and power supplies, which are required to keep this behemoth of compute power cool. In total, there are 6 x hot-swap 3000W 96% titanium PSUs, which double up duty by both providing power and pulling air through the system. These are similar to those found in Supermicro’s SuperBlade product line.
The middle of the system contains another removable I/O tray. By default, the system provides 8 x PCIe 5.0 x16 low-profile slots and 2 x PCIe 5.0 x16 full-height, full-length slots. Primarily, these will be utilised for networking, with the low-profile slots used for single-port high-bandwidth connections in a 1:1 ratio with the GPUs to maximise bandwidth when used in a cluster. The remaining slots can have more local networking; for example, ours is equipped with a simple RJ45 10GbE adaptor.
In the bottom left, you can also see our dedicated BMC/IPMI port, 2x USB ports and a standard VGA output.
Step 3: The Star of the Show—AMD Instinct MI300X
The GPU tray is removed by undoing the 2x captive screws on either side, which unlock the handles. The handles can be pulled down, and the tray slides out. There are catches on either side of the rails for the tray to release it all the way out. This tray is heavy, and a second person is needed to safely remove it. Once the tray is out, we get an understanding of the weight—eight tower heatsinks, each dedicated to cooling an MI300X chip with 192GB of HBM3 memory and a peak power of 750W each! It’s safe to say a significant portion of the weight is allocated to the heatsinking needed to dissipate this heat, together with 4U of uninterrupted rack surface area to access fresh air. This is approaching the limits of traditional air cooling without switching to liquid cooling technologies.
Each of the MI300X accelerators is installed in an OAM socket that directly links to each other via AMD’s Infinity Fabric. Four connect to each CPU via PLX/PCIe switches, which also facilitate direct network access, bypassing the CPU. For those wanting some juicy numbers, we’ve got:
- Over 155,000 stream processors peaking at 2.1GHz
- 40TB/s of potential memory bandwidth
- A total aggregate HBM3 memory of 1,536GB
Here, you can see a shot inside the chassis showing the massive numbers of connections to the CPU board. You can see the outermost connections on the left and right with the large pins to provide power and the high-density connectors in between for the PCIe lanes. Gently slide it back! Despite it looking cool to slam that tray home, you do not want to damage those connectors.
Step 4: The Trusty Backbone—Mainboard and CPU
The compute tray on the front features 5x hot-swap fans pulling in air to keep the standard system components cool, below which is an array of U.2 NVMe drives. This is cabled with 12 drives supported by default, but it’s possible to add support for an additional 4x NVMe and even a couple of SATA if required.
Removing the compute tray is the same as the GPU tray above, with two captive screws on either side, levers to pull down, and it then slides out in the same manner with retention levers to press down on the rails when it’s halfway out. This is substantially lighter than the GPU board, though it’s still recommended to bring a friend, especially if mounted high in a rack.
Here, we have the brains of the operation: two AMD SP5 sockets with a 400W TDP and Turin support. Ours is fitted out with two AMD Genoa 9654 96-core/192-thread CPUs, accompanied by 24x 64GB DDR5-4800 memory modules, totalling 1,536GB of system RAM. This is cooled by air from the chassis fans directed by a sturdy Perspex air shroud, helping to concentrate the air through the large heatsinks and the accompanying memory.
Hiding under the PCIe ribbon cable in the middle of the board between the memory banks of each CPU is where you find our double-stacked 1.92TB Samsung NVMe M.2 drives we’ll be using as our boot drives. To the right, you can see 4x small heatsinks covering the PLX/PCIe switch chips driving the connections between CPU and GPUs. Below these in the image is yet another PLX bridge to provide additional lanes to the PCIe slots at the rear and the NVMe at the front. If you notice the empty gap on the other side, there is space for a second of these to enable another two x16 expansion slots and the remaining 4 NVMe drives.
Step 5: Conclusions
Unboxing the Supermicro AS-8125GS-TNMR2 with AMD Instinct MI300X accelerators reveals a true powerhouse in AI, ML and HPC systems. With its combination of dual AMD EPYC processors, high-performance AMD Instinct MI300X accelerators and massive memory bandwidth, this system is engineered to tackle the most demanding computational tasks. Whether you're working on large-scale AI model training, big data analytics or complex scientific simulations, this server offers the flexibility, power and performance needed to stay ahead of the curve. The system remaining air-cooled is another feather in its cap for those integrating into existing environments without access to DC water cooling.
For those seeking a future-proof system designed to handle the next generation of workloads, the AS-8125GS-TNMR2 is a top-tier choice. With its seamless integration of AMD's powerful hardware, this server provides the scalability and efficiency needed for any modern enterprise or research lab.
This AS-8125GS-TNMR2 with AMD Instinct MI300X accelerators is now available for customer testing in our Boston Labs facility. Boston Labs is all about enabling our customers to make informed decisions in selecting the right hardware, software and overall solution for their specific challenges. If you’d like to request a test drive of the MI300X accelerators or other technology, please get in touch by emailing [email protected], or call us on 01727 867100, and one of our experienced sales engineers will gladly guide you through building the perfect solution just for you.
Author
Thomas Mitchell
Junior Field Application Engineer
Boston Limited