New Server Hits the Machine-Learning Track

The new Dell EMC DSS 8440 server accelerates machine learning and other compute-intensive workloads with the power of up to 10 GPUs and high-speed I/O with local storage.

As high-performance computing, data analytics and artificial intelligence converge, the trend toward GPU-accelerated computing is shifting into high gear. In a sign of this momentum, the TOP500 organization notes that new GPU-accelerated supercomputers are changing the balance of power on the TOP500 list. This observation came in 2018 when a periodic update to the list found that most of the new flops came from GPUs instead of CPUs.[1]

This shift to GPU-accelerated computing is having a major impact on the HPC market. IDC projects that the accelerated server infrastructure market will grow to more than $25 billion by 2022, with the accelerator portion accounting for more than half of that volume.[2]

“With AI manifesting itself in the datacenter and the cloud at a phenomenal rate and with traditional high-performance computing increasingly looking for performance beyond the CPU, the quest for acceleration is heating up, as is the competition among vendors that offer acceleration products,” an IDC research manager notes.

Driving accelerated computing forward

At Dell EMC, the Extreme Scale Infrastructure (ESI) group is helping organizations catch the accelerated-computing wave with a new accelerator-optimized server designed specifically for machine learning applications and other demanding workloads that require the highest levels of computing performance.

This new 2-socket, 4U server, the Dell EMC DSS 8440 server, has 10 full-height PCIe slots in front, plus 6 half-height PCIe slots in back to create the right balance of accelerators, launching with 4, 8 or 10 NVIDIA® Tesla® V100 GPUs. It also incorporates extensive I/O options with up to 10 drives of local storage (NVMe and SAS/SATA) to provide increased performance for compute-intensive workloads, such as modeling, simulation and predictive analysis in scientific and engineering environments.

The new design enables accelerators, storage and interconnect on the same switch for maximum performance, while providing the capacity and thermals to accommodate future technologies. Offering efficient performance for common frameworks, the DSS 8440 server is ideal for machine learning training applications, reducing the time it takes to train machine learning models and time-to-insights. It allows organizations to easily scale acceleration and resources at the pace of their business demands.

The rise of a new machine

The DSS 8440 was developed in response to customer demand for even higher levels of acceleration than were previously offered by Dell EMC, according to Paul Steeves, a product manager for the new server.

As our customers push further ahead with machine learning solutions, it has become obvious that there was a need for increased amounts of accelerated raw horsepower,” Steeves says. “While accelerated servers exist from our competitors, many of our customers want open solutions, with choice not just now, but also over time as technology advances.

In addition, Dell EMC designed the DSS 8440 server specifically with machine learning training in mind, Steeves notes. For example, the system includes 10 high-performance local drives and extensive I/O options to deliver a more targeted solution for today’s growing number of machine learning workloads.

Key takeaways

  • The DSS 8440 server offers extremely high levels of acceleration with up to 10 NVIDIA V100 GPUs in an open PCIe fabric architecture that allows other open-standard components to be easily added in future versions.
  • The DSS 8440 server delivers the raw compute performance that HPC-driven organizations need today, coupled with the flexibility to adopt new machine learning technologies as they emerge.

Putting the system to work

The DSS 8440 server is designed for the challenges of the complex workloads involved in the process of training machine learning models, including those for image recognition, facial recognition and natural language translation.

“It is particularly effective for the training of image recognition and object-detection models, where it performs within a few percentage points of the leading numbers — but with a power efficiency premium,” Steeves notes.

Another strength of the DSS 8440 server is its ability to enable significant multi-tenant capabilities.

“With 10 full-height PCIe slots available, customers can assign machine learning or other compute-intensive tasks to several different instances within a single box,” Steeves says. “This allows them to readily distribute compute among departments or projects.”

The bottom line

As organizations move more deeply into machine learning, deep learning applications and other data- and compute-intensive workloads, they need the power of accelerators under the server hood. The new Dell EMC DSS 8440 server meets this need with a versatile balance of accelerators, high-speed I/O and local storage.

To learn more

[1] TOP500, “New GPU-Accelerated Supercomputers Change the Balance of Power on the TOP500,” June 26, 2018.

[2] IDC, “Worldwide Accelerated Server Infrastructure Forecast, 2018–2022,” March 2018.