We’re excited to announce that Supermicro will be attending Future of Memory and Storage (FMS) 2026 and has been nominated for the best of show award in the Server and Networking category for the Supermicro ARS-22L-FE324R NVIDIA® CMX™ server. FMS will be held at the Santa Clara Convention Center & Hyatt Regency in Santa Clara, CA from August 4 – 6.
This year’s event brings together innovators in advanced memory and storage technologies. Stop by our booth to see our latest solutions and attend our three speaking sessions.
Visit Us at Booth #946
The Supermicro ARS-22L-FE324R NVIDIA® CMX™ server, based on NVIDIA STX, is the world’s first air-cooled AI native storage server designed specifically for KV cache/context memory storage methods used in Agentic AI / reasoning applications. Features include built-in NVIDIA® Bluefield®-4, Spectrum-X™ networking and Vera CPUs with NVIDIA BMC management. Both Time-to-First-Token and power are dramatically reduced.
Supermicro ARS-22L-FE324R NVIDIA® CMX™ context memory server is based on NVIDIA’s reference design. Using Supermicro’s expertise in thermal management, it uniquely uses standard air-cooling and 240V power to allow it to be deployed in many environments including development labs, traditional data centers and colocation centers.
As enterprises adopt AI Factory infrastructure and AI inference workloads scale, the inference efficiency is a key factor in delivering cost-effective and power efficient AI. Key to solving this challenge is the implementation of KV (key-value) cache for storing prior inference results and reusing this data with longer context AI queries. NVIDIA has a new tier of memory focused on KV cache which is enabled by the Supermicro ARS-22L-FE324R NVIDIA® CMX™ context memory storage.
Supermicro, together with NVIDIA, is at the forefront of this new domain and is developing new platforms to address large-scale AI inference. The Supermicro ARS-22L-FE324R NVIDIA® CMX™ is a new class of storage infrastructure designed for fast, efficient inference at a giga-scale. As power demands in data centers increase, improving power efficiency in networking and storage frees more power for the brains of AI, the GPUs.
The transformer model with the self-attention mechanism fundamentally changed generative large language models by enabling parallel processing of users’ query inputs and the ability to weigh the importance of different words in the query based on matching words, word position, and other factors. Processing the entire sequence of input query words in parallel, rather than one at a time, left-to-right, dramatically increased the model's performance.
The next change in inference efficiency is the implementation of the KV cache. As inference becomes more sophisticated, it has evolved from a one-shot question-and-answer paradigm to a conversational multi-sequence process in which, as in human conversations, the meaning of the current query depends on the previous sequence of queries. This conversional approach requires storing the sequence of previous queries, extending the context window from the current query to previous queries, even days prior.
Instead of calculating the importance of each word (token) in the input query each time the word is encountered, the KV cache stores the data associated with the word (value) and an associated lookup index (key) so that this data can be reused without recomputing these values the next time the word is encountered. Storing the key-value pairs in a disaggregated inference infrastructure requires a new type of storage infrastructure, which NVIDIA calls CMX™. NVIDIA CMX is built on NVIDIA STX, the modular foundation for AI-native storage, to help optimize the entire data lifecycle from data analytics and model training to full agentic workflows at scale. While the conversation context could be stored in the GPU’s High Bandwidth Memory (HBM), it is far too expensive and not large enough to store all the context for all of the queries being processed. The CMX™ solves this by introducing a new tier of storage that enables scaling the KV caches for large inference deployments.
Supermicro ARS-22L-FE324R NVIDIA® CMX™ enables efficient storage of critical data and accelerates AI-native key-value (KV) cache access. This feature enables very fast data sharing across nodes and delivers significantly improved power efficiency for this task. Supermicro and NVIDIA are working together to expand the set of high-performance enterprise AI solutions that incorporate these new technologies.
With BlueField-4, KV cache data can be stored on fast SSDs, accelerating its distribution to other GPUs. As part of a software-defined infrastructure, the NVIDIA BlueField-4 increases performance and security while freeing GPUs to perform their intended tasks.
However, KV cache does not require the same durability, redundancy, or data protection features used for long-lived enterprise data. In the context of AI inference, it can be recomputed if lost.
For data centers, using the Supermicro ARS-22L-FE324R NVIDIA® CMX™ context memory system, the following results can be expected:
Supermicro ARS-22L-FE324R NVIDIA® CMX™ system is air-cooled with standard power drops which allows it to be easily used in both development environments and conventional air-cooled data centers.
Designed for large-scale service providers to increase storage density with the same power and physical footprint of traditional disk-based storage enclosures. In a 4U 19” form-factor with 2600W air-cooled enclosure using a single processor server-node, the system supports 160 U.2 SSDs with up to 40PB using 245TB SSDs with one-tenth the power per TB of 30TB HDDs.
Most data centers have a limited power capacity provided by the utility. This limits the amount of CPU and GPU servers, which in turn, limits the revenue of the data center. By minimizing other sources of power usage, particularly from storage, more of the limited available power can be allocated to the profit-making CPU and GPU computing.
This is the catalyst for the Supermicro ASG-4116S-NU160R, Ultra High-Capacity All-Flash Server which replaces disk-drive based servers with all-flash servers. High-capacity flash using QLC media now is available in capacities up to 245TB per SSD as compared to HDDs which are available in capacities of 30TB or slightly more. One of the key design strategies was to keep the same power envelope for the 3.5” HDD and the U.2 SSD which is 9W. 9W is lower than typical for high-capacity SSDs, so the development team worked closely with key SSD suppliers to utilize the low power modes supported by the SSD to limit the power consumption to 9W without significant impact to performance.
By keeping the system enclosure design the same as Supermicro’s 4U disk-array server especially the physical dimensions and the total power, the Ultra-High Capacity All-Flash Server is a drop-in replacement for existing large disk-array servers. A single All-Flash system using 245TB SSDs can replace 10 racks of HDD storage servers. The design also maintained the air-cooled design of the HDD design to allow universal data center combability. With 160 U.2 SSDs, ensuring adequate cooling using air flow was a challenge which was solved by innovative ducting and thermal simulations.
This solution allows large service providers to replace large-scale racks of HDD Storage with only 10% of the rack footprint and the same power envelope per rack. With 40PB capacity per 4U chassis using 245TB SSDs, this allows 400PB per 42U rack, replacing 10 racks of disk-based servers and saving 90% of the power. For power capped data centers, this saved power can be utilized for compute and GPU systems allowing more GPUs in the same data center power limit.
The modular front I/O system incorporates several design innovations including the sliding inner-drawer design perfected on Supermicro’s Top loading servers. The design allows all the SSDs to be serviceable without removing the server from the rack. Another innovation is co-designing the system with SSD suppliers to utilize the low power states supported by the SSD to minimize total power consumption.
See these solutions up close at our booth and talk to our experts about these solutions.
Title: Reinventing Storage for Long-Context LLMs: Tiered KV Cache from HBM to NVMe
Topic: AI & ML Applications
Speakers:
Date & Time: August 4, 2026 at 9:05 AM Pacific Time
Location: Mission City Ballroom
Description: Explosive AI growth requires us to reinvent the rules of storage. As context windows and concurrent sessions grow, LLM inference is quietly hitting a wall where KV cache, not FLOPs, becomes the real performance bottleneck; and the traditional options (more GPUs, more HBM, shorter prompts) are all painfully expensive. In this session, Supermicro and Graid Technology present a tiered KV cache design that turns dense NVMe-backed GPU servers into a high-performance KV cache tier that lets you scale context, concurrency, and sessions per node without blowing up your GPU budget. Using Supermicro NVMe-dense GPU platforms with Graid SupremeRAID™, the architecture turns SSD into a high-throughput, resilient KV cache tier with full enterprise RAID protection (0/1/5/6/10). We will also discuss the 5 tiers of KV cache storage and how a large scale disaggregated inference workflow partitions the KV cache data.1. HBM on GPUs2. CPU DRAM on the storage server3. Local SSD on the storage server4. KV cache storage using DPUs5. Network storage which can be File or Object.
Title: Panel Discussion – Storage as a Service: Neo-Cloud Economics vs Hyperscale Control
Topic: Data Center Storage & Memory
Speakers:
Date & Time: August 4, 2026 at 9:45 AM Pacific Time
Location: Ballroom A
Description: AI NeoClouds are racing to differentiate but storage margins often flow upstream. Hyperscalers increasingly bundle Storage-as-a-Service with GPU compute, capturing performance premiums and tightening ecosystem control. Meanwhile, scale-out storage platforms (Weka, VAST Data, DDN, Hammerspace) power NeoCloud clusters, abstracting flash volatility while competing on throughput and latency guarantees. Beneath them sit controller, NVMe, and flash suppliers whose economics depend on density, endurance, and bandwidth scaling. Open-source stacks and bare-metal storage clusters offer an alternative path but with operational trade-offs. This panel maps the full stack: hyperscale services, NeoCloud offerings, software, defined storage layers, component vendors, and flash technology providers, asking who owns margin, who owns performance, and how AI-era storage models evolve.
Beer, Pizza and Chat with the Experts
Topic: Multi-Tiered AI Storage
Date & Time: Wednesday, August 5. 7:15 PM – 8:30 PM Pacific Time
Speakers:
Description: A relaxing social setting at FMS. Enjoy food and drinks with leading experts. Join discussion tables covering a wide variety of topics.
Explore our products, attend one of our sessions, or ask any questions. We look forward to seeing you there!