Marvell Technology has announced its intent to acquire compute express link (CXL) startup Tanzanite in an all-cash deal it says will accelerate its composable infrastructure aspirations.
Founded in 2020, Tanzanite was an early player in the emerging CXL marketplace specializing in memory expansion, tiered memory, and memory pooling technologies. At the heart of these developments was Tanzanite's smart-logic interface connector, which enables memory to be pooled across compute servers at low latencies using CXL.
For those unfamiliar with the term, CXL is an open-standard interconnect introduced in 2019 that piggybacks PCIe to provide a consistent interface between host CPU processors, memory, accelerators, and other peripherals. Since its inception, the technology has garnered support from more than 190 vendors, including Intel, AMD, IBM, and Nvidia.
Marvell envisions a future in which CPU/GPU compute, memory, and storage resources will be disaggregated into individual boxes, then composed on the fly. It believes Tanzanite and CXL will play a pivotal role in achieving this goal.
Neither party discussed the terms of the transaction, which is expected to close during Marvell's second fiscal quarter of 2022.
"Fully composable, higher-bandwidth, and near-memory compute is what everyone is trying to get to, and this is a pathway of intellectual property that's going to help Marvell get there," Daniel Newman, principal analyst and founder of Futurum Research, told The Register on Monday.
Marvell's interest in Tanzanite centers primarily around memory pooling, in which a microprocessor in one server can access memory resources in another across a PCIe fabric.
Today, memory-intensive workloads – like AI/ML, analytics, and search – often require large multi-socket servers to achieve the desired memory density. "If you don't need all the additional compute, it's an ineffective and really expensive way to expand your memory footprint," Thad Omura, VP of marketing for Marvell's flash business unit, told The Register.
While it's been possible to compose CPU compute and storage for years, "memory is kind of the last frontier," he explained. "Now, for the first time, CXL is going to drive not just memory expansion, but pooling of the technology, which is then going to drive a fully composable datacenter."
Tanzanite first demoed this capability as a proof of concept in March, when it enabled multiple processors to access a pool of 80TB of memory at latencies it claimed were lower than could be achieved in a dual-socket server.
"This is really the first step in being able to make the datacenter fully composable," Omura said.
CXL's broader implications
Marvell believes CXL will underpin a shift to disaggregated compute architectures in the datacenter. The Tanzanite acquisition is a step toward realizing this goal, and the technology's implications aren't limited to memory.
Marvell envisions CXL accelerators being integrated into a wide swath of its product portfolio – including its data processing units and storage systems, according to Omura.
"If you think about all of those different product types that CXL is going to touch, we believe Marvell is in the best position to do that," he explained.
A long way to go
While Intel's long-delayed Sapphire Rapids Xeon Scalable and AMD's upcoming Genoa and Bergamo Epyc gen-4 processors will mark CXL's arrival on mainstream platforms, the technology is still very much in its infancy.
One of the biggest challenges facing CXL is extending its benefits beyond the server chassis – a must for composable infrastructure.
Traditionally, server-to-server communications have been done over Ethernet, though this isn't practical at memory speeds, according to Omura.
"Conversion of CXL to Ethernet and then back into memory may, from our perspective, impose too much latency," he explained. "When you start looking at pooling of memory, it'll be important that you keep latency end-to-end to a minimum."
One way Marvell is looking to get around this could involve a blade-server style backplane.
Futurum Research's Daniel Newman opined, "If they can solve this, they are going to add a whole lot of value by bringing memory and compute almost seamlessly together without latency."
Realizing Marvell's composable infrastructure ambitions is going to take time, Omura admitted. "This is not an overnight type of thing, but the industry has identified a standard way to go after solving this complete infrastructure composability issue."Get our Tech Resources