The World’s First 32 GPU Single-node AI Supercomputer for Next-Gen AI and Accelerated Computing
SuperNODE In the News
AI and Accelerated Computing Challenges
- Large Language Model (LLMs) and Generative AI need large numbers of GPUs
- Standard server configurations restrict GPUs to what fits inside the server sheet metal
- Fixed server architectures result in lower utilization rates of expensive and hard to come by accelerators
- Networking fixed configuration GPU servers over legacy networks increases latency thus reducing performance
Want to try your own code on SuperNODE for free?Request access to our lab.
GigaIO SuperNODE™ with FabreX™ Dynamic Memory Fabric
- Connects up to 32 AMD Instinct™ MI210 GPUs or 24 NVIDIA A100s and up to 1PB storage to a single off-the-shelf server
- Enables lower power, adaptive GPU supercomputing
- Delivers the ability to train large models with tools like PyTorch or TensorFlow, scaling via peer-to-peer communication on a single node, instead of MPI over several nodes
- Accelerators can be split over several servers, and scale using node-to-node communication for larger data sets
Benefits of GigaIO SuperNODE
- Shorten time-to-results with single-node code getting vastly more compute power
- Keep code simple: use your existing software without any changes – “It just works”
- Secure the ultimate flexibility for any workload: unprecedented power in “BEAST mode,” flexible configurations in “SWARM mode,” or shared resources in “FREESTYLE mode
- A single node solution reduces network overhead, cost, latency, and server administration
- Save on power consumption (7KW per 32-GPU deployment)
- Save on rack space (30% per 32-GPU deployment)
“This is an incredible platform for HPC and ML/AI. It is really wild to see 32 GPUs appear on ROCm SMI!”
AMD Fellow, HPC
“As AI workloads become more broadly adopted, systems that offer the ability to harness the compute power of multiple GPUs and better manage data saturation at ultra-low latency are essential. And as large language model applications drive demand for more GPU performance, technologies that work to minimize node-to-accelerator traffic are better positioned to provide the necessary performance for a robust AI infrastructure.”
“AMD collaborates with startup innovators like GigaIO in order to bring unique solutions to the evolving workload demands of AI and HPC. The SuperNODE system created by GigaIO and powered by AMD Instinct accelerators offers compelling TCO for both traditional HPC and generative AI workloads.”
Corporate Vice President and General Manager,
Data Center and Accelerated Processing
A New Era of Disaggregated Computing
Technologies that reduce the number of required node-to-accelerator data communications are crucial to providing the stripped-down horsepower necessary for a robust AI infrastructure.
The GigaIO SuperNODE can connect up to 32 AMD Instinct™ MI210 GPUs or 24 NVIDIA A100s to a single node at the same latency and performance as if they were physically located inside the server box. The power of all these accelerators, seamlessly connected by GigaIO’s transformative PCIe memory fabric, FabreX, can now be harnessed to drastically speed up time to results.
The SuperNODE is a simplified system capable of scaling multiple accelerator technologies such as GPUs and FPGAs without the latency, cost, and power overhead required for multi-CPU systems.
Made Possible by FabreX
The GigaIO SuperNODE is powered by FabreX, GigaIO’s transformative high-performance PCIe memory fabric. In addition to enabling unprecedented device-to-node configurations, FabreX is also unique in making possible node-to-node and device-to-device communication across the same high-performance PCIe memory fabric. FabreX can span multiple servers and multiple racks to scale up single-server systems and scale out multi-server systems, all unified via the FabreX software.
Resources normally located inside of a server — including accelerators such as GPUs and FPGAs, storage, and even memory — can now be pooled in accelerator or storage enclosures, where they are available to all of the servers in the system. These resources and servers continue to communicate over the FabreX native PCIe memory fabric for the lowest possible latency and highest possible bandwidth performance, just as they would if they were still plugged into the server motherboard.
Unprecedented Compute Capability Available Now
Available today for emerging AI and accelerated computing workloads, the SuperNODE engineered solution, part of the GigaPod family, offers both unprecedented accelerated computing power when you need it, and the ultimate in flexibility and in accelerator utilization when your workloads only require a few GPUs.