michelangelus - Fotolia

Fungible computational storage eyes composable use cases

Fungible's computational storage anchors a new offering for composable infrastructure. The Fungible Data Center license allows IT teams to consume resources on demand.

Computational storage startup Fungible is pitching turnkey composable infrastructure to hyperscale customers.

The venture-funded chipmaker on Tuesday introduced Fungible Data Center, a rack of bare-metal gear it describes as "hyper-converged disaggregated composable infrastructure." In layman's terms, the product combines the all-flash Fungible FS1600 storage system with commodity servers and switches, knit together by the Fungible True Fabric network stack and Data Center Composer virtualization software.

Fungible is among a handful of companies developing new storage processing technologies that run data operations close to the compute. The Fungible storage system is based on its F1 Data Processing Unit (DPU) programmable microprocessor. The dedicated offload card hosts the Fungible storage controller and associated data services. Block storage and compute resources scale independently for SaaS consumption.

Fungible Data Center supports multi-tenancy with secure isolation. Enterprise teams instantiate environments with one-click deployment templates. Fungible gained the software-based template technology last year via its acquisition of disaggregation vendor Cloudistics.

Pardha Pyla, Fungible's senior vice president of product solutions and designPardha Pyla

"What Fungible is delivering is a system with a software stack tightly integrated into its DPU chip. Delivering a full system is always a good thing, because it takes the guesswork out of integration. You get a fully optimized solution out of the box," said Ashish Nadkarni, an IDC group vice president of infrastructure systems, platforms and technologies.

Composable storage allows customers to acquire hardware resources as needed without buying a full rack. Users provision capacity from a pool of resources for as long as required. The resources are returned to the pool when the task is complete.

Designing the Fungible storage system to be composable lets IT teams dynamically configure resources based on the needs of an application, said Pardha Pyla, Fungible's vice president of product solutions and design.

"The gap between processing and networking has gone up dramatically over the last five or six years, and storage is seeing a concomitant gap. We have a chip at the junction of storage and networking that allows you to perform operations very close to where they originate," Pyla said.

Compose and compute

While there isn't yet broad adoption, computational storage may be ready to move beyond test benches into deployment, according to analysts. Technologies are emerging that allow data operations to be performed very close to the storage. Besides Fungible, other computational storage startups include Burlywood, Eideticom, Nebulon, NGD Systems, Pliops, SambaNova Systems and ScaleFlux.

Computational systems seek to address the limitations of hyper-converged infrastructure systems, in which users can only scale compute and storage by purchasing additional nodes. These newer technologies are designed to help organizations maximize the value of their compute and storage hardware spending, said Randy Kerns, a storage analyst at Evaluator Group, based in Boulder, Colo.

You're going to see some of these products implemented with apps that are data-intensive and that need some degree of data manipulation.
Randy KernsAnalyst, Evaluator Group

"You're going to see some of these products implemented with apps that are data-intensive and that need some degree of data manipulation" to condition data for AI and machine learning, Kerns said.

Nvidia Corp. has emerged as an influential player. Best known as a chipmaker for multimedia gaming, Nvidia has integrated its DGX supercomputer systems in storage systems via OEM deals. In separate transactions last year, Nvidia snagged computer networking company Mellanox Technologies and SwiftStack object storage, which is part of the Nvidia AI stack. (Nvidia's planned $40 billion takeover of chip designer Arm Holdings may be in jeopardy, however, with industry rivals urging federal regulators to review the deal.)

Nvidia's aggressive pursuit of data centers raised the antennae of Intel, which responded by acquiring AI chipmaker Habana Labs in 2019.

Fungible Data Center components

Fungible's projected users include large enterprises, research labs and service providers. That pits the startup against legacy vendors with composable infrastructure systems, including Hewlett Packard Enterprise's Synergy and Dell EMC's PowerEdge MX.

Pyla said Fungible Data Center was engineered to get the most of commodity bare metal, including standard top-of-rack Ethernet switches that connect to data center spines. The system comes with diskless servers that contain CPU, memory and Fungible's PCIe-based controller card, which offloads the x86 hypervisor data path to avoid wasted cycles.

The 2U Fungible FS1600 data plane accepts 24 U.2 NVMe SSDs. Compression, encryption and erasure coding are native data services. The out-of-band Fungible Composer virtualizes the bare-metal components into disaggregated services.

"In most places, we expect people will start with a half rack to one rack for high performance and then expand to run everything else in a way that is future-proof," Pyla said.

Fungible Data Center is available in small, medium and large configurations. Small and medium configurations include two FS1600 nodes with 92 TB of raw capacity, with the medium option having twice as many compute nodes (16). The largest configuration scales to seven storage nodes, 1.29 TB of storage and 32 compute nodes. Each option supports three composer nodes. Fungible Data Center pricing depends on the subscription, although the vendor did not provide specifics.

Next Steps

SambaNova funding round brings in $676 million

Dig Deeper on Flash memory and storage