Marvell Technology, a leader in data infrastructure semiconductor solutions, has launched the Marvell Structera product line of Compute Express Link (CXL) devices that enable cloud data centre operators to overcome memory performance and scaling challenges in general-purpose servers.
To address memory-intensive applications, data centre operators add extra servers to get higher memory bandwidth and higher memory capacity.
The compute capabilities from the added processors are typically not utilised for these applications, making the servers inefficient from cost and power perspectives. The CXL industry-standard addresses this challenge by enabling new architectures that can efficiently add memory to general-purpose servers.
The Structera product line comprises two CXL device families that are optimised for different use cases. The Structera A CXL near-memory accelerators are a new category of devices that integrate server-class processor cores and multiple memory channels with CXL to address high-bandwidth memory applications such as Deep Learning recommendation models (DLRM) and Machine Learning.
The Structera X CXL memory-expansion controllers enable terabytes of memory to be added to general-purpose servers and address high-capacity memory applications such as in-memory databases. The Structera CXL device families are the industry’s first to support four memory channels, integrate inline compression and use 5nm manufacturing processes.
In addition to the Structera A and X standard products, Marvell develops custom CXL silicon for cloud operators that are optimised for their unique architectures and workloads.
A new CXL use case: ML/AI acceleration
The Structera A accelerators integrate 16 Arm Neoverse V2 cores and are optimised to improve the performance of high memory-bandwidth applications such as DLRM. DLRM is characterised by both sparse and dense memory operations. Sparse operations may hit a memory wall, in which memory bandwidth is insufficient for the available compute. The Structera A 2504 accelerator, the first product in the family, supports up to 200 GB/sec memory bandwidth and 4TB of memory capacity.
For a DLRM server using a single 64-core processor, adding a single Structera A device would increase the number of compute cores by 25% (64 vs 80), aggregate memory bandwidth by 50% (400 GB/sec vs 600 GB/sec), total memory by up to 4TB and memory bandwidth per core by 25% (6.25 GB/sec vs. 7.5 GB/sec).
It would also improve memory bandwidth power efficiency from 1W to 0.83W per GB/sec.
Adding two Structera A devices would increase compute cores by 50%, double aggregate memory bandwidth, increase memory capacity by up to 8TB, increase memory bandwidth per core by 50% and improve memory bandwidth power efficiency to 0.75W per GB/sec.
The Structera X product family, by contrast, is designed to expand memory capacity per server to address high-capacity memory applications. The Structera X devices are the industry’s first to support four channels of DDR4 and DDR5 memory and support up to three DDR4 dual in-line memory modules (DIMMs) per channel (3DPC) for maximising memory capacity per controller and server node.
The devices are also the first CXL memory-expansion products that can simultaneously support two server CPUs to optimise power, space and memory utilisation.
The Structera X 2404 enables operators to recycle their DDR4 DIMMs from decommissioned general-purpose servers. Millions of functional DDR4 DIMMs are expected to become e-waste over the next few years as operators replace existing general-purpose servers with new ones utilising DDR5. Using ‘free’ DDR4 DIMMs to expand the capacity in these servers reduces CAPEX by thousands of dollars per general-purpose server. Repurposing decommissioned DDR4 DIMMs also addresses data centre sustainability goals.
