FAQ

Liqid uses a PCIe fabric for composable infrastructure and this enables extremely flexible usage of different types of resource (compute, storage, networking, graphics processing) with the lowest possible latency.

Liqid Fabric introduces approximately 150 ns latency.

Liqid's flexible fabric is suited to using high performance and high cost items across multiple physical machines. When a use case drives toward massive infrastructure agility or “infrastructure as code,” the practicality of Liqid Composable solution and the Fabric becomes higher. The flexibility that a composable fabric can add to existing technology stacks can further extend the life of equipment and reduce the procurement cycles and associated economic burdens. Typical use cases for this technology are the following: AI, machine learning, Genomics research and high performance modeling, research and development.

Liqid does not certify devices, but the fabric allows for any standard PCIe device to be added to the fabric and used across machines created on the fabric. For further confidence, Liqid can work with your engineering team to extensively test devices as needed. Liqid internal qualification testing starts with GPUs from major vendors and networking devices from Mellanox and Intel as well as NVMe from Samsung, Toshiba, and other vendors (including Liqid’s own products).

Liqid has several partnerships for expansion chassis. Each of these chassis has capabilities that make them valuable. As an example, the Inspur GX4 is ideally suited to AI/machine learning devices, where some the OSS chassis are more uniquely suited to visualization and storage expansion use cases. It is important to note details like external GPU/video port requirements and needs surrounding networking and power/cooling when selecting chassis.

The basic answer is unlimited device and hosts depending on topology. However, while a fabric can scale to hundreds of devices and hosts with fan-out and fabric topologies, there are practical limits to consider. A typical Liqid Composable topology will have several hosts and many devices on a single fabric. PCIe enumeration is the first limit that affects devices mapped to a host. Typically, device counts on a single host do not exceed 256 devices. Complex GPUs are limited to about 16 per host but can stretch a little further with low bandwidth applications like crypto-mining. SSDs used for capacity can extend this total closer to theoretical limits, but storage performance will become bottlenecked beyond about 16-24 modern, high-performance NVMe SSDs devices per host. Note: A standalone Liqid Grid Switch can accommodate 96 lanes of PCIe. The balancing act of performance and cost with topology must be considered.

Liqid currently works with many partners providing hosts that work seamlessly with the Liqid Grid and Composable products. Please call your dedicated Liqid sales person for further details if making a purchase decision. Liqid also uses well-tested server products with its Dev Kits. While Liqid does not certify products, it does recommend products with the minimum qualification criteria for use with the grid fabric and its software:

• IPMI 1.2 compatibility
• EFI shell boot capability
• 1 or more available PCIe Gen 3.0 x16 slot(s), HHHL slot size or greater.
• Redundant power/cooling

Liqid has physical redundancy with redundant power in each switch. Furthermore, multi-switch topologies can replicate configuration and device information to other switches to enable quick recovery from switch-level failures. In addition to this, customers can use multi-fabric topologies to fault isolate devices and enable hosts to accommodate switch failure. The software platform that manages the stack has automated resiliency services to monitor devices and subordinate services and automatically recover from failures.

Yes, Dev Kits come from the factory with pre-engineered components but can be reconfigured within limits in the field. Your dedicated Liqid Support team can help you determine practical paths for scale and device substitutions as needed.