You Can Have Your GPU and Cool It Too

Solve problems faster, more sustainably, using Dell PowerEdge servers with Intel GPU Max Series

As the need for accelerated compute grows, the importance of embracing the full spectrum of the technology ecosystem, specifically in regard to GPU diversity for infrastructure in HPC and AI systems, grows too. At the same time, sustainability and energy efficiency have become top priorities for many HPC data center procurement plans, according to a global study conducted in April 2023 by HPC analyst Hyperion Research.

These criteria top even price as imperatives; only performance is more important — and HPC users are starting to think more in terms of performance per watt, with wattage incorporating multiple uses of energy including powering the equipment, cooling the HPC site, and more.

A recent benchmarking test by the Strategic Technology Analysis Center (STAC), which sets the industry standard for HPC benchmarking for financial services institutions, along with the recent announcement from Dell Technologies, Intel and Cambridge University about Dawn Phase 1, the fastest GPU-accelerated supercomputer deployed in the UK today,¹ suggest that HPC data centers may no longer have to prioritize performance over efficiency – because Dell Technologies offers technology that provides both, and can deliver it quickly.

Record-breaking Performance and Efficiency

The STAC-A2^TM Benchmark is the industry standard for testing technology stacks used for compute-intensive analytic workloads involved in pricing and risk management. It is such a well-known and well-understood benchmark in financial services that many now use it as a proxy for judging how well a technology stack would do with other HPC workloads.

In October 2023, STAC performed the first STAC-A2 Benchmark tests on liquid-cooled Dell PowerEdge XE9640 servers with 4^th Gen Intel Xeon scalable CPUs and Max Series 1550 GPUs. This is also the first liquid-cooled system with publicly disclosed STAC-A2 audit results. The baseline for comparison was against all other systems tested to date, with 40 systems tested in the last 10 years, as recently as August 2023.

Compared to all publicly reported solutions to date, this solution based on the Dell PowerEdge XE9640 servers with Intel^® Data Center GPU Max 1550 (SUT ID INTC230927) set numerous performance and efficiency records, including (but not limited to):

• The fastest warm^[1] (0.405 s) and cold^[2] (1.09 s) times in large problem size benchmarks
• A space efficiency^[3] (238 options / hour / cu. in.) 2.3x better than the previous best result
• The best energy efficiency^[4] (314,493 options / kWh), 1.0% better than the previous record

This system set records in performance efficiency by leveraging the Dell PowerEdge XE9640’s density and direct liquid cooling (DLC)—which allows for efficient use of rack space and improves performance. This is the first water-cooled system under test audited by STAC. Direct liquid cooling allowed the PowerEdge XE9640 to cool > 3kW in a 2U form-factor, which smashed the previous space-efficiency record (throughput / volume) by 2.3X.

Compared to a system from HPE tested in August 2023 (SUT ID NVDA230721), this solution delivered:

• 78% of the throughput^[5]
• 98% of the speed in warm runs in the baseline problem size benchmark^[6]
• 1.7x the speed in cold runs of the large problem size benchmark²
• 1.2x the speed in warm runs of the large problem size benchmark¹
• 4.3x the space efficiency³

Compared to a report on the August 2023 test by HPE, the Dell server with only a four-GPU system had 78% of the throughput, essentially identical performance in the baseline benchmarks (warm), and significantly better performance in the large problem size. The PowerEdge XE9640 performed 1% better in energy efficiency with only four GPUs vs. the HPE server that was able to amortize the base server power over the throughput of eight GPUs.

What We Tested

The stack featured a Dell PowerEdge XE9640 server with 4 x Intel^® Data Center GPU Max 1550 accelerators and 2 x Intel^® Xeon^® Platinum 8468 processors at 2.1 GHz, with 32 GiB of memory and running Ubuntu Linux 22.04.3 LTS.

• About the Dell PowerEdge XE9640 server: This direct liquid cooled 2RU server handles the most demanding AI and simulation workloads while optimizing data center cooling efficiency and maximizing GPU core density per rack.

• About the Max Series GPU: This is Intel’s highest density processor, packing over 100 billion transistors into a 47-tile package with up to 128 gigabytes (GB) of high bandwidth memory. The oneAPI open software ecosystem provides a single programming environment for both new processors. Intel’s 2023 oneAPI and AI tools will deliver capabilities to enable the Intel Max Series products’ advanced features.

“These STAC results and the Dawn Phase 1 supercomputer announcement are both powered by Dell PowerEdge XE9640 servers with Intel Max Series GPUs,” stated Ogi Brkic, Vice President and General Manager, Data Center AI Solutions Category. “The advances represent the advantage that Dell Technologies and Intel offer the industry by providing the GPU-starved market with an alternative.”

Read the report from STAC to view details about the system under test.

Connect with Dell

“STAC” and all STAC names are trademarks or registered trademarks of the Strategic Technology Analysis Center, LLC.

^[1] STAC-A2.β2.GREEKS.10-100k-1260.TIME.WARM

^[2] STAC-A2.β2.GREEKS.10-100k-1260.TIME.COLD

^[3] STAC-A2.β2.HPORTFOLIO.SPACE_EFF

^[4] STAC-A2.β2.HPORTFOLIO.ENERG_EFF

^[5] STAC-A2.β2.HPORTFOLIO.SPEED

^[6] STAC-A2.β2.GREEKS.TIME.WARM

¹ based on Cambridge Open Zettascale Lab’s own performance analysis