Run AI workloads where and when required with composable GPU acceleration for the Dell PowerEdge MX7000 Modular Chassis
Connect multiple high-performance GPUs to the PowerEdge MX7000 modular compute platform for high performance AI and ML applications
Powerful GPU accelerators to handle demanding and evolving workloads
Up to 30x 350W datacenter GPUs per MX7000 chassis for maximum performance on any workload
Dynamically allocate GPU resource when and where required for higher utilisation and efficiency
Combine standard Compute and GPU resource to build a system tailored to your needs
MX7000 with CoreModule XL and Liqid SmartStack MX is designed for deployment and installation in a range of environments to bring processing close to the data in order to avoid your valuable and business-critical information leaving your network. Run AI training and inference locally to keep your data secure and contained to your network, whilst enabling business functions to leverage modern AI models to drive growth. Supporting up to 35C ambient temperature, high performance GPU on MX7000 isn’t restricted to core datacenter, but can be deployed to remote locations where required.
Available to configure with a range of standard 350W PCIe dual-slot GPUs from NVIDIA, Intel and AMD, the solution can be configured appropriately for a wide range of workloads in AI, machine learning, rendering and simulation. By combining multiple GPUs in a linked configuration, up to 192GB of unified memory is available for running training and inference on large AI models in order to provide higher accuracy or get a project running with less time required for model optimisation and quantitzation.
Rather than silo GPU resources in a fixed configuration, Coremodule XL with Liqid Matrix allows the entire pool of GPU resource to be allocated to compute sleds on demand, based on the needs of the current workload. Using the configurable PCIe switch fabric, each GPU is connected directly to the target CPU as a native PCIe device without any virtualisation or bridging layer. This allows the higher level software stack to work as usual without any modification or need to know about the underlying connectivity, so no additional time is required to alter algorithms or models.
Based on standard components and standard architecture, CoreModule XL with Liqid SmartStack requires no modification to existing toolchains, libraries or APIs, so existing models, tools and processing stacks run the same as on any other platform using the same GPUs. This makes it easy to move workloads from an existing GPU platform and take advantage of the composible infrastructure.
Easy to scale from a single Compute sled connected to a single GPU, all the way up to eight dual-Xeon compute sleds with thirty 350W GPUs, CoreModule XL with Liqid SmartStack MX makes it easy to deploy PowerEdge modular infrastructure suited to any size of workload and task. The system can be upgraded with a single sled or single GPU at a time which join the pool of resources which can be dynamically assigned to work together or with other resources. GPUs which support the linking feature can be setup as dual or quad units to share their memory and work together on larger tasks and broken down again to single GPUs for re-use in a different use case.
| CoreModule XL with Liqid SmartStack | |
| Form factor | Dell MX7000 Modular Compute Chassis with Liqid EX-4410 PCIe expansion chassis |
| Maximum Configuration | Eight MX760C Dual Xeon 4th/5th Gen Compute Sleds 30x 350W Dual-slot GPU in three expansion chassis |
| Minimum configuration | 1x MX740C, MX750C or MX760C compute sled 1x 70W Low-profile GPU |
| Compute Sled connectivity | Compute Sled, PCIe Gen 4, Sixteen lane direct connection using Amulet Hotkey custom MX Mezzanine Card |
| Compute Sled Requirements | MX740C, MX750C or MX760C compute sled with Dual 4th or 5th Gen Intel Xeon CPUs |
| MX7000 Fabric | Fabric B1 and B2 required for CoreModule XL |
| Power and Cooling | Independent redundent power and cooling for each expansion chassis |
| Rack space | MX7000 modular chassis – 7U Liqid EX-4410 PCIe chassis – 4U (holds ten GPUs) Liqid Director – 1U PCIe 48-port switch – 1U |
MX7000 Modular Compute Chassis (Rear view shown) with CoreModule XL located in Fabric B1/B2, connected to all eight compute sleds inside the chassis.
CoreModule XL contains up to eight HBA expansion cards – one for each sled.
Direct connection from each compute sled via HBA card using high speed cable assemblies up to 2m (6ft) length.
The non-blocking PCIe Gen 4 packet switch provides connectivity from each compute sled to each of the GPU expansion chassis allowing each compute sled to access the GPU devices assigned to it.
Management server to control the assignment of GPU resources to each compute sled and dynamically re-assign resources as required. The Director provides a local webUI and a range of APIs for integration to existing system management, scripting tools and datacenter orchestration systems.
4U chassis with high speed PCIe Gen 4 x16 connectivity for up to ten PCIe double-wide GPUs or other PCIe devices. Up to three expansion chassis can be connected to an MX7000 chassis to provide a total of 30 GPUs in a resource pool.
Unless a compute system is dedicated to a single task for its entire operating lifetime, the use of that system, the workloads it needs to run and the needs of the business are usually better served by deploying resources in a flexible manner which can be adapted and reconfigured as needs require. With fixed hardware and GPU dedicated to specific servers, this reconfiguration is complex and time-consuming, so expensive processing resource may be left idle. By deploying those GPU resources in a composable architecture, the resources can be easily shared amongst multiple host servers and reconfigured many times per day to make best use of that investment.
It can be difficult to estimate how workload performance scales when running across multiple parallel GPUs. With a composable infrastructure driven from scripts or an automation platform, it’s easy to run the same test multiple times and get real data to show that performance improvement. This makes it easy to understand the resources required by a given compute job and where the allocating additional resource is a waste.
NVIDIA L40S 48GB
NVIDIA H100 80GB
Intel MAX 1100 48GB
AMD Instinct MI210 64GB
NVIDIA H100 NVL 188GB
Two Intel MAX 1100 48GB with XE-Link bridge
Two AMD Instinct MI210 64GB with Infinity Link bridge
Four Intel MAX 1100 48GB with XE-Link bridge
Four AMD Instinct MI210 64GB with Infinity Link bridge
For any product issues, please raise a support ticket with us. Support tickets let you track any issue you have with a product, and follow it through to a resolution.
Raise a ticketWe believe that partnerships play a key role in our products and solutions’ development, delivery and ongoing success. That is why we work hard to deliver an ecosystem of partners, resellers, and cloud service providers – working together to meet the needs of our customers around the globe.
Get in touchFind out how our solutions can benefit your organization
