Accelerating AI Innovation and Sustainability: The High-Density, High-Performance Dell PowerScale F910
Mon, 03 Jun 2024 16:31:19 -0000
|Read Time: 0 minutes
Accelerating AI Innovation and Sustainability: The High-Density, High-Performance Dell PowerScale F910
In the era of rapid technological advancement, enterprises face an unprecedented challenge: accelerating AI innovation while minimizing environmental impact. As the demand for AI processing power skyrockets, so does the associated energy consumption, thus leading to a significant increase in carbon footprint. Power consumption has become a recurring topic with the introduction of AI. The New York Times reports, A.I. Could Soon Need as Much Electricity as an Entire Country (nytimes.com). These challenges call for a modern solution bridging the gap between performance and sustainability.
Enter Dell’s PowerScale F910 platform, a high-density, high-performance node designed to accelerate AI innovation while dramatically reducing an enterprise’s carbon footprint. A cutting-edge platform, transforming the way organizations approach AI, reaching ambitious performance goals without compromising our environmental responsibility.
How much more density are we talking about? The F910 offers 20% more density per rack unit than the F710 which was released merely three months ago with extraordinary performance. Moreover, the F910 delivers density with performance, as it’s up to 2.2x faster to AI insights.
In this blog post, we’ll explore how our high-density, high-performance platform redefines the landscape of AI innovation and sustainability. Discover how your enterprise can leverage this groundbreaking technology to accelerate AI initiatives, drive business growth, and make a positive impact on the environment.
AI innovation and sustainability challenges
Many challenges arise as enterprises increasingly rely on AI to drive innovation and gain a competitive edge. One of the most significant challenges is the exponentially increased demand for AI processing power. As AI models become more sophisticated and data volumes explode, enterprises require hardware solutions to keep pace with these evolving needs.
However, with great processing power comes great energy consumption. The energy-intensive nature of AI workloads has led to a concerning rise in enterprises' carbon footprints. Data centers housing AI infrastructure consume vast amounts of electricity, often generated from non-renewable resources, contributing significantly to greenhouse gas emissions. The impact here is not only environment but it is also a risk to businesses as consumers, investors, and regulators who increasingly prioritize sustainability.
Moreover, AI hardware's high energy consumption translates into substantial operational costs for enterprises. The electricity required to power and cool AI systems can quickly eat into an organization’s bottom line, making it challenging to justify the ROI of AI initiatives. In fact, Forbes reports Generative AI Breaks The Data Center: Data Center Infrastructure And Operating Costs Projected To Increase To Over $76 Billion By 2028 (forbes.com).
To address these challenges, enterprises urgently need hardware platforms that deliver high performance while prioritizing energy efficiency. They require platforms that can handle the demanding workloads of AI innovation without compromising on sustainability goals. The industry is calling for a paradigm shift in AI hardware design that places equal emphasis on processing power and environmental responsibility.
Fortunately, as the AI factory, Dell Technologies has answered the call to balance AI innovation with sustainability, by developing the new PowerScale platform that tackles these challenges head-on. In leveraging cutting-edge technology and innovative design principles, our solution enables enterprises to accelerate AI innovation while significantly reducing their carbon footprint. In the next section, we’ll examine how our hardware platform revolutionizes the AI landscape and paves the way for a more sustainable future.
Introducing the PowerScale F910
At Dell, we understand the pressing demand for a hardware solution that can bridge the gap between AI innovation and sustainability. From this need, we’ve developed a cutting-edge hardware platform designed to address the challenges enterprises face in the AI landscape.
Our high-density, high-performance hardware platform is a testament to our commitment to push the boundaries of AI technology while prioritizing environmental responsibility. By leveraging state-of-the-art hardware and the technical innovations of OneFS, the F910 provides unparalleled processing power in a compact, energy-efficient package.
Overview
The F910's front panel has a bezel protecting the 24 NVMe SSD drives, as displayed in the image below.
The front panel has an LCD that offers a range of information and status updates. It also has an option to add a node to an existing PowerScale cluster. The LCD display is also used to view the node’s iDRAC IP address, MAC address, cluster name, asset tag, power output, and temperature information. Furthermore, the front panel has an LED for the status on the left. For example, a failed drive illuminates an amber LED.
Moving on to the rear of the F910, we can see all the connections.
The power supplies are split across the backplane, allowing maximum airflow through the center of the chassis. The front-end and back-end network interfaces are on opposite sides, offering Ethernet connectivity. The other interfaces on the rear include iDRAC, serial, and management NICs.
The F910 nodes use NVMe SSDs. In a 6 RU rack configuration of 3 nodes, the F910 raw capacity spans a minimum of 276.5 TB to a maximum of 2.16 PB. The available drive capacities for the F910 are listed in the following table.
Non-SED Drive Capacities | SED-FIPS Drive Capacities | SED-Non-FIPS Drive Capacities |
3.84 TB | 3.84 TB | 15.36 TB |
7.68 TB | 7.68 TB | 30.72 TB QLC |
15.36 TB QLC | 15.36 TB QLC* | |
30.72 TB QLC | 30.72 TB QLC* |
*Future availability
For a new cluster deployment, a minimum of 3 F910 nodes is required to form a cluster. For existing cluster deployments, the F910 is node pool compatible with the F900, allowing the F910 to be added in a multiple of 1. If an existing cluster does not have any F900s, a minimum of 3 nodes is required to form a new node pool.
High density
In addition to being a high-performance platform, the F910 is the highest-density all-flash PowerScale node. We’ve engineered our system to pack not only an exceptional amount of computing power but also drive density, maximizing AI capabilities without the need for extensive physical infrastructure. This reduces the spatial footprint of AI hardware and minimizes the energy required for cooling and maintenance. See the table below to compare how the F910 compares to our other all-flash platforms.
Platform | Cluster Density per Rack Unit |
PowerScale F200 | 30.72 TB |
PowerScale F210 | 61 TB |
PowerScale F600 | 245 TB |
PowerScale F710 | 307 TB |
Isilon F810 | 231 TB |
PowerScale F910 | 360 TB |
The F910 offers 20% more density compared to the F710. This number is further magnified if we compare the F910 to the F810, where the F910 offers a 55% gain! So now, let's characterize this into. What would look like in a data center? Let’s take a scenario where a data center currently has 10 racks, and each rack is filled to its current maximum capacity. With a 55% gain in density per rack unit, we can now fit 55% more computing and storage resources in each rack compared to the previous setup. If we take a simple scenario where a current data center has 10 racks, with 55% density gain, that's 10 ÷ 155% ≈ 6.45 racks.
In this scenario, the impact of a 55% gain in density per rack unit is still significant, as it allows the data center to reduce its rack consumption by 35%. Reducing physical space requirements can lead to cost savings, improved efficiency, and greater flexibility for future growth. Now, let’s visualize this in the image below.
Data reduction is enabled by default out of the box, further increasing the F910’s high density and capacity envelope. The inline data reduction process incorporates both compression and deduplication. When these elements are combined, they significantly boost the overall density of a cluster. As the density per Rack Unit (RU) increases, it decreases the Total Cost of Ownership (TCO) for the solution, reducing the carbon footprint.
High performance
The F910 achieves the ultimate performance envelope by taking advantage of hardware and software updates. PowerScale OneFS 9.7 and 9.8 introduced several performance-oriented updates.
Note: The PowerScale F910 requires OneFS 9.8 at minimum. The mention of OneFS 9.7 here is to understand the performance leap from previous OneFS releases.
OneFS 9.7 introduced a significant leap in performance by enhancing the following:
- Implementing a round-robin distribution strategy across thread groups has significantly reduced thread lock contention, increasing OneFS's overall efficiency and performance.
- Contention on turnstile locks has been reduced by increasing the value of Read-Write (RW) Lock retries, optimizing system performance.
- In the context of NVMe storage nodes, writing operations are strategically executed around the journal for newly allocated blocks, therefore maintaining high performance and preventing data processing and access delays.
OneFS 9.8 further pushes the pure software performance envelope, optimizing the OneFS 9.7 updates to further build on them. Additionally, OneFS 9.8 introduces enhancements to thread handling and lock management. Finally, general code updates have brought about a significant performance leap.
On the hardware front, the F910 leverages the PowerEdge platform for extreme performance. Powered by a dual-socket Intel® Xeon® Gold 6442Y Processor, it delivers higher core counts, faster memory speeds, and improved security features. The F910 features PCIe 5.0 technology, which doubles the bandwidth and reduces the latency of the previous generation, thus enabling faster data transfers and more efficient use of accelerators. Furthermore, the F910 takes advantage of the DDR5 RAM, offering greater speed and bandwidth. The table below summarizes the F910’s hardware specifications:
Attribute | PowerScale F910 Specification |
CPU | Dual Socket – Intel Sapphire Rapids 6442Y (2.6G/24C) |
Memory | Dual Rank DDR5 RDIMMs 512 GB (16 x 32 GB) |
Front-end networking | 2 x 100 GbE or 25 GbE |
Infrastructure networking | 2 x 100 GbE |
NVMe SSD drives | 24 |
The combination of hardware and software updates allows the F910 to tackle even the most challenging workloads, minimizing time for AI insights. Overall, the F910 delivers AI insights 2.2x faster than previous generations. Let’s take that into context for a minute. If learning an AI model takes 10 hours to complete, being able to do it 2.2 times faster means it would be finished in approximately 4.55 hours. That’s a significant improvement in efficiency and productivity. You’re saving approximately 5.45 hours that can be used for other AI models.
NVIDIA DGX SuperPOD certification
Dell PowerScale is the world’s first Ethernet-based storage solution certified on NVIDIA DGX SuperPOD. The collaboration between Dell and NVIDIA is designed to help customers achieve faster and more efficient AI storage. Dell PowerScale exceeds the performance benchmark requirements for DGX SuperPOD. Integrating PowerScale and DGX SuperPOD allows for handling vast amounts of data at unprecedented speeds, thereby accelerating the process of training AI models.
The PowerScale F910 expands the family of the already NVIDIA DGX SuperPOD-certified storage solution, while accelerating training times and balancing sustainability. For more on the PowerScale NVIDIA DGX SuperPOD certification, see h19971-powerscale-ethernet-superpod-certification.pdf (delltechnologies.com).
Services
Accelerate AI outcomes with help at every stage from Dell Professional services. Trusted experts work alongside you to align a winning strategy, validate data sets, implement, train and support GenAI models and close skills gaps to help you maintain secure and optimized F910 operations now and into the future. Furthermore, Dell services embeds sustainability throughout our services portfolio to proactively help customers approach the most pressing environmental challenges from sustainability to reducing waste.
Customer feedback
During our F910 beta program, partners and customers tested and validated the F910's performance. We wanted to confirm our performance and density gains in a real-world environment. More importantly, we wanted to know what the gains would be on an existing workload.
At the onset, after the first batch of tests, all of the initial feedback was consistent. To paraphrase,
“I know you all claimed a lot more performance, but I didn’t think it would be this good.”
We were ecstatic to hear that feedback. As the tests rolled on, more glowing reviews continued to come in.
In the end, John Lochausen, Technical Solutions Architect at World Wide Technology, summed up the sentiment best:
“We're hyper-focused on AI innovation in our AI Proving Ground, and the all-flash PowerScale F910 has exceeded our expectations. It doubles performance, reducing the power and energy costs required for the same workload, further advancing our customers' sustainability goals.”
Conclusion
In conclusion, what the F910 proves is that when it comes to AI innovation and sustainability, you can have your cake and eat it, too. Organizations can now accelerate AI innovation while accelerating sustainability. To summarize, the F910 checks all the modern AI workload requirements: High-Performance 🗹 High-Density 🗹 Power-Efficient 🗹 NVIDIA-Certified 🗹
For more on the PowerScale F910, see PowerScale All-Flash F210, F710, and F910 | Dell Technologies Info Hub
Author: Aqib Kazi, Senior Principal Engineering Technologist