Puget Systems debuts at SIGGRAPH 2024 its newest custom LLM AI Training and inference server with support for up to eight NVIDIA RTX Ada generation GPUs, NVIDIA L40S GPUs or NVIDIA H100 Tensor Core GPUs, and NVIDIA Networking Platforms. The company says that the whole system is optimized for maximum GPU performance in a 4U server in on-premises local data center deployments.
The newest Puget Systems AI Training and Inference Servers offer AMD’s EYPC line of processors, which are designed for demanding server workloads, offering up to 128 cores with support for 1.5TB DDR5 ECC RAM and 128 PCIe Gen 5 lanes. The servers use a pair of those CPUs to enable up to eight dual-width GPUs in a 4U rack mount chassis, and several hot-swap drive bays are easily accessible on the front, along with a pair of USB ports.
To demonstrate the power of its newest specialized AI Training and Inference servers, the team from Puget Systems will be on hand at SIGGRAPH 2024 in booth #536 to showcase a sophisticated, GPU-intensive, real-time AI image generation solution that uses a StreamDiffusion pipeline. The demonstration will be run locally on the Puget Systems server to showcase how these complex, hardware-intensive workflows – which require powerful GPUs, such as the NVIDIA RTX 6000 Ada Generation GPU, to generate high-quality imagery with smooth frame rates – are possible using Puget Systems’ custom, localized hardware.
Attendees at SIGGRAPH 2024 will see how the Puget Systems Generative AI and Machine Learning Server can pass a source image into a Stream Diffusion operator in TouchDesigner with a GUI to control the Stream Diffusion parameters, such as the Acceleration Library, Resolution, Steps, and Prompt. Additionally, various input methods, including webcams, visual noise, Adobe Photoshop (for live AI painting), and a mouse pointer for physical fluid simulation, can be selected in the GUI. The interpretation of the input image then creates an AI-generated output image in real time at high frame rates.
The Puget Systems AI Training and Inference Servers are rack-mount workstations specifically designed for GPU-intensive generative AI workflows, much like the StreamDiffusion demo shown in Puget Systems’ booth at SIGGRAPH 2024.
With support for up to eight NVIDIA GPUs with up to 752GB of total VRAM, this server is ideal for machine learning, AI, and rendering workloads. Additionally, for enterprise or data center customers, the Puget Systems AI Training and Inference Server also offers support for the NVIDIA L40S or NVIDIA H100 NVL Tensor Core GPU. The NVIDIA H100 NVL GPU is ideal for LLM inference with its high compute density, memory bandwidth, and energy efficiency, and NVIDIA NVLink architecture. It also delivers extraordinary acceleration to power the world’s highest-performing elastic data centers for AI, data analytics, and high-performance computing (HPC) applications. Select Puget Servers also support the NVIDIA BlueField networking platform, including the NVIDIA BlueField-3 DPU.
Puget Systems custom AI Training and Inference servers are available for configuration for a wide range of generative AI applications. Follow the link to learn more. If you’re in Canada, follow this other link to learn more about Puget Systems’ Canadian consulting and sales operations.