The Challenge: Procuring a compact, high-performance, DC-class mass storage solution suitable for AI Applications
AI servers are powerful computing platforms designed to host AI driven applications. Such applications require considerable computational power and memory resources to perform optimally. As such, AI servers typically employ large banks of GPUs (graphic processing units) to handle incoming data, such as high-definition video captured by AI and machine learning systems.
However, processing a particularly voluminous set of data, or analyzing a complex scenario or object, can potentially overwhelm available VRAM (video RAM), which would then slow the entire process to a crawl. To combat this risk, the majority of AI computing platforms utilize high-speed storage devices to supplement the VRAM.
However, as much of the interior space of an AI server or workstation platform is devoted to the GPUs, and motherboard hardware, little to no room is available for internal drives bays. This largely rules out SAS/SATA SSDs or U.2/U.3 NVMe media for most applications, both of which utilize the 2.5” drive form factor. As a result, M.2 AIC NVMe controllers are the most common choice. Fast, compact and widely available, M.2 SSDs can be installed directly into purpose-built PCIe NVMe storage controllers, such as our own SSD7500 series PCIe Gen4 M.2 NVMe RAID HBAs, which eliminate the need for an internal drive bays.
Unfortunately, M.2 media has its drawbacks. While M.2 SSDs deliver excellent random I/O and sequential performance, write endurance tends to suffer over the long haul, which is a typical setback for client-class NVMe SSDs. In addition, M.2 drives have short lifespans compared to enterprise class U.2/U.3 drives. Storage space is also a major concern; M.2 series suitable for AI applications, namely “DC” (datacenter) class SSDs, are capped to 4TB in capacity, which can limit their utility for workflows that must process large volumes of continuously incoming data.
All of this poses a serious challenge for solution providers; procuring a storage solution that successfully merges the form factor of an M.2 SSD with the balanced transfer characteristics and storage potential of U.2 media, and a maximum data throughput beyond what is capable with a single-AIC PCIe Gen4 solution. The answer? E1.S technology.
The Solution: E1.S NVMe RAID Storage
E1.S devices are a relatively new type of NVMe media designed to combine the form factor of an M.2 SSD, with the datacenter class reliability and mass storage capability of U.2/U.3 drives. They provide an excellent balance of random and sequential I/O performance, and are not prone to handicap transfer speeds during extended write sessions. They are similar in sized to 22110 form-factor M.2 SSDs; small enough to integrated directly into a PCIe AIC. They are also available with twice the storage capacity; up to 8TB currently, with larger models set to arrive in the near future. The vast majority of off-the-shelf E1.S SSDs are classified as DC class storage, and measure longevity in terms of DWPD (disk writes per day), rather than TBW (total bytes written), and are equipped with advanced features such as PLP (power loss protection).
But where to start? Most currently E1.S implementations are at the board level (a handful of E1.S ports integrated directly into a motherboard), and few include RAID capability. One option it to use PCIe carrier cards with up to 4x DC class 22110 M.2 SSDs. However, this is a temporary, stop-gap measure, and not a true solution. 22110 DC class media is capped at 4TB in capacity; simply too small for most AI applications.
Once again, HighPoint has the answer. The revolutionary SSD7749E E1.S NVMe RAID controller is up to the challenge.
Near Gen5 Performance in a Gen4 package
The ability of a single SSD7749E to generate 28Gb/s of transfer bandwidth is impressive enough. However, when deployed in Cross-Sync, dual-card configuration, SSD7749E HBAs can deliver a level of performance that is nearly on par with a dedicated PCIe Gen5 storage solution!
Many AI servers now utilize Gen5 capable motherboards; high-performance, 200Gbe+ network controllers are becoming increasingly common. There are currently no native PCIe Gen5 NVMe RAID controllers, which could potentially bottleneck some AI applications.
However, the SSD7749E can effectively bridge this gap. Cross-Sync RAID technology enables SSD7749E configurations to leverage up to x32 lanes of PCIe bandwidth, and deliver 55,000MB/s of throughput; a level of performance equivalent to the real-world capabilities of a native Gen5 solution, and capable of accommodating network speeds in excess of 400 GigE!
Administrators can fine-tune Cross-Sync configurations with HPT-Optimize, a performance optimization toolset capable of linking specific target applications with the resource pool assigned to the SSD7749E devices. Learn More…
HighPoint’s SSD7749E E1.S RAID Controller was built to tackle Demanding AI Server & Workstation Applications
HighPoint’s SSD7749E 8-Channel E1.S NVMe RAID controller was designed to address demanding data-dense, high-throughput, low-latency workflows associated with AI applications. The dual-width PCIe Gen4 HBA directly hosts up to 8 DC class E1.S NVMe SSDs. Think 28,000MB/s of transfer speed and 60+ TB of RAID storage capacity from a compact device that is roughly the same size and shape of a modern GPU! The SSD7749E’s zero-compromise dual-width HBA architecture eliminates the need for internal drives bays, cabling accessories or supplemental cooling apparatus; the intelligent hardware packaging enables the SSD7749E to be easily integrated into the industry standard rackmount systems employed by most AI Servers.
Dedicated Dual-Width Cooling System Eliminates Thermal Throttling
PCIe Gen4 NVMe media runs hot, especially under load, which is almost a 100% guaranteed condition for any storage solution employed by an AI application. Gen4 NVMe media, including E1.S SSDs, will scale back performance if a temperature threshold is breeched in an effort to avoid compromising the hardware. This feature is commonly referred to as “thermal throttling”. Thermal throttling is an ideal failsafe, but can cripple transfer performance when activated, especially for a RAID configuration.
The SSD7749E was designed to tackle the threat of thermal throttling head on. Its advanced cooling system takes full advantage of the HBA’s dual-width architecture, and combines a full-length anodized aluminum casing with integrated heat sinks and a pair of ultra-durable, low-decibel cooling fans that double as a secure locking mechanism. This robust, purpose-built solution fully encases, insulates and cools E1.S media without injecting excessive noise into the work environment.
The dual-fans inject cool air into the controller’s casing, which is then dispersed to the NVMe media and center mounted heat sink. Waste heat is then rapidly transferred away from critical componentry, and ejected to the outside environment via a ventilated bracket. The hardware cooling system was designed to work in conjunction with HighPoint’s SHI (Storage Health Inspector) software management interface, which allows administrators to instantly check the operating status and temperature of NVMe media in real-time via S.M.A.R.T. technology, and configure temperature thresholds on a per-drive basis, to ensure the SSD7749E is perfectly in sync with capabilities and official specifications of each individual drive.
Host Interface: PCIe 4.0 x16
Device Ports: 8x E1.S
Form Factor: FL-FH
The SSD7749E is a simple, cost-efficient PCIe Gen4 RAID storage upgrade solution for Intel and AMD x86, and ARM server platforms. The controller can directly host up to eight 9.5mm E1.S SSDs at speeds up to 28,000MB/s, yet is no larger than a modern dual-width PCIe graphics adapter, and can be easily integrated into systems with a free PCIe 4.0 x16 slot.