Summary

With the recent announcement of 3rd Gen Intel® Xeon® Scalable processors, Dell has announced 2 different models of the R650 and 3 different models of the R750 to meet emerging customer demands. This paper is intended to highlight the engineering elements of each design and to describe the reason for the expansion of the portfolio.

These 3 classes of systems are designed to optimize for differing workloads.

• The R750xa design is optimized for heavy compute environments and to support this, it has been designed for maximum cooling with front mount slots for the use of GPU’s and support for CPU Thermal Design Points (TDP’s) up to 270W and 40 cores.
• The R650 and R750 systems are also designed for heavy compute environments but are designed for maximum flexibility. This includes enhanced drive counts to deliver optimal storage capacity, and CPU’s with TDP’s up to 270W and 40 cores.
• The R650xs and R750xs systems are designed for traditional virtualized environments and support CPU’s with TDP’s up to 220W and 32 cores. performance in a Software Defined Storage environment like vSAN. Both models support a maximum of 1TB of memory using 64GB DIMM’s.

Introduction

Optimizing between cost, performance and scalability is a difficult balancing act when designing a Server. Mainstream environments like virtualization have established design points that focus on cores, memory capacity and storage density to achieve the ideal configuration. The advent of new technologies like Persistent Memory places additional demands on the design and emerging applications like Artificial Intelligence (AI) and Machine Learning (ML) stretch these designs even further.

The challenge for server design teams is to strike an effective balance that delivers maximum performance for each workload/environment but doesn’t overly burden the customer with unnecessary cost for features they might not use. To illustrate this, consider that a server designed for maximum performance with an in-memory database will require higher memory density while a server designed for AI/ML might benefit from enhanced GPU support and a server designed for virtualization with software defined storage might benefit from enhanced disk counts as shown in the chart below. All of these technologies could take advantage of a new processor design and all need access to memory, but each requires a unique approach to deliver optimization.

While it may be technically possible to build a single system that could achieve all of this, the end result would be much more expensive to purchase and could be potentially larger. For example, a system capable of powering and cooling multiple 400W GPU’s needs to have bigger power supplies, stronger fans, additional space (particularly for double wide GPU’s) and high core count CPU’s. Conversely, a system designed as a virtualization node might require none of these optimizations. Trying to optimize for all often results in unacceptable trade-offs for each.

To achieve truly optimized systems, Dell Technologies is launching 3 classes of its industry leading PowerEdge Rack Servers. The “xa” model, the “standard” models and the “xs” models. The “xa” model is designed for optimization in AI/ML environments and to support that, delivers optimized power, cooling and enhanced GPU support. The “standard” models are flexible enough to deliver an enhanced virtualization or Database environment with the addition of storage capacity and extra memory expansion using DRAM or Persistent Memory (PMEM) and the “xs” models are designed for mainstream virtualization with large disk capacities, CPU support for up to 32 cores and cost effective memory capacities of up to 1TB. 

Design Optimizations

As noted above, the “xa” model is optimized for GPU, the “standard” models are optimized for high performance compute and the “xs” models are optimized for virtualized environments. Below is an overview of the key feature differences:

¹Based on current 3rd Gen Intel® Xeon® Scalable processor family

²DW=Double Wide GPU

³SW=Single Wide GPU

While key specifications differ between models, much remains the same. It is important to note that all models support key features such as:

• iDRAC 9 and OpenManage
• OCP3.0 Networking options
• PCIe 4.0 slots – up to 3 in the R650xx series and up to 8 in the R750xx series
• PERC 11 RAID including optional support for NVMe RAID
• 3200MT/s Memory
• Installation in standard depth racks

“xa” Design

As noted above, the R750xa is optimized for enhanced GPU support. This support is accomplished by moving 2 of the rear PCIe cages to the front as highlighted in the graphic below. Each of these cages can support up to 2 Double Width GPU’s and in the case of the NVidia A100, each pair can be linked together with NVLink bridges. Additional PCIe slots are available in the rear of the system. GPU workloads typically require less internal storage than mainstream workloads so with this change, internal storage has been located in middle of the front of the server and provide up to 8 SAS/SATA, NVMe or a mix of drive types. All of these configurations are available with optional support for RAID using the new PERC11 based H755 (SAS/SATA) or H755n (NVMe). These RAID controllers are located directly behind the drive cage to save space and are connected directly to the Motherboard of the system to ensure PCIe 4.0 speeds. To accommodate these new technologies, the depth of the chassis has been extended by 101.2mm (compared to the R750 “standard”) but will still fit within a standard depth rack. To ensure the highest levels of performance, this model ships with optional support for the 2nd Generation of Intel® OptaneTM Memory, up to 32 DIMM slots and Processors with up to 40 cores.

“Standard” Design

The R650/R750 “standard” models have been designed to accommodate the flexibility necessary to address a wide variety of workloads. With support for large numbers of hard drives (up to 12 in the R650 and up to 28 in the R750), these models also offer optional performance and reliability features with the new PERC 11 RAID controller using the PERC H755 (SAS/SATA) or H755n (NVMe) including a “Dual PERC” option with multiple controllers. These RAID controllers are located directly behind the drive cage to save space and are connected directly to the Motherboard of the system to ensure PCIe 4.0 speeds. To ensure the highest levels of performance, these model ship with optional support for the 2nd Generation of Intel® Optane^TM Memory, up to 32 DIMM slots and Processors with up to 40 cores. In addition, both models support GPU but to a lesser extent than the “xa” series.

“xs” Design

When designing for virtualization, a number of key factors emerge. Storage requirements often serve software defined storage schemas (like vSAN) while the ability of a hypervisor to segment memory and cores creates a need to balance between the two. To meet these demands, the new “xs” designs include support for up to 16 DIMM’s, which translates to 1TB of DRAM when using 64GB DIMM’s, CPU’s with up to 32 cores and internal storage of up to 24 drives (16 SAS/SATA+8 NVMe – R750xs) or 10 drives (SAS/SATA or NVMe – R650xs). These designs assign 1 DIMM socket per channel allowing customers to scale out with balanced configurations. These models were also optimized to provide a lower acquisition cost. While the cost of a DIMM socket might appear insignificant, the impact of reducing the number of DIMM sockets is large. The most obvious is power and cooling. Any design needs to reserve enough “headroom” for a full configuration and by cutting the number of DIMM sockets in half, an “xs” power budget can be reduced. This in turn reduces the amount of cooling required which allows the use of more cost effective fans and potentially reduced cost by limiting baffles and other hardware used to direct air flow. This also helps explains why an “xs” system can operate on a power supply as small as 600W while a “standard” system requires a minimum of 800W power supplies to operate. Another impact to cost is the fact that increasing the number of DIMM sockets in a system increases the complexity of the design. A DDR4 DIMM has 288 pins and by removing 16 sockets from the design, 4,608 electrical traces were also removed. Reducing the number of electrical traces by this scale allows the motherboard to be built with fewer “layers” which translates directly into a lower cost. Recent pricing trends for memory have created an opportunity to achieve excellent performance, scalability and balance with smaller numbers of DIMM’s. Specifically, the $/GB ratio of a 64GB DIMM is evolving to be similar to the ratio of a 32GB DIMM. This means that customers can achieve the same balance that was achieved with previous generations with fewer DIMM sockets.

Conclusion

With the launch of the new 3rd Gen Intel® Xeon® Scalable processors, Dell Technologies is able to deliver a range of new technologies to meet customer requirements. From the “xa” model and its ability to deliver high GPU density to the “standard” models that deliver a robust platform for a wide range of workloads through to the “xs” series that delivers compelling price:performance, customers can now achieve a level of optimization not previously available.