1 min read
Add as a preferred source on Google

Liqid Announces Managed AI Service as It Pursues Memory Pooling

Liqid as a Service is a fully managed, single-tenant offering. The company is also moving toward preconfigured appliances.

A memory expansion unit connects to GPU servers in a data center aisle / TokenPost.ai
A memory expansion unit connects to GPU servers in a data center aisle / TokenPost.ai

Liqid announced a fully managed, single-tenant offering called Liqid as a Service at SC26, as the company pursues a rack-level approach to assigning GPU and memory resources to AI servers.

Liqid co-founder and Chief Technical Officer Sumit Puri described the company’s vision as pooling GPUs and memory in racks, then assigning those resources to servers based on the model or workload they run. The announcement did not specify that the new service itself includes that resource-allocation capability.

Puri said Liqid evaluates performance using tokens per second, tokens per watt and tokens per dollar, with the last measure carrying the most weight. He also said memory is a major supply constraint and that delivering each application the amount it needs could help address supply-chain pressure.

Liqid is also moving toward preconfigured appliances. Puri predicted that businesses would want dedicated capacity to consume AI tokens, comparing the idea to unlimited-use plans for mobile minutes.

The company’s approach shares a broad theme with SK hynix’s proposal for shared memory racks for AI inference.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…