Serial memory controller boosts memory bandwidth -

Serial memory controller boosts memory bandwidth

TORONTO — Microchip chose to enter the memory infrastructure market at the Flash Memory Summit with the introduction its SMC 1000, a serial memory controller designed to alleviate the bottleneck between the CPU and memory.

Microchip’s SMC 1000 8x25G enables CPUs and other compute-centric SoCs to use four times the memory channels of parallel attached DDR4 DRAM within the same package footprint, according to product marketing manager Jay Bennett in a telephone interview with EE Times . From a CPU point of view, the number of embedded cores has been steadily increasing, but the memory bandwidth capability of that CPU has not been keeping pace. “Individual cores within the CPU are each individually experiencing a decrease in aggregate bandwidth and also an aggregate increase in the latency for each of their individual transactions,” he said.

The problem is being compounded by changing and increasingly demanding workloads, said Bennett, and the key objective of the SMC 1000 is to increase available memory bandwidth. It interfaces to the CPU via 8-bit Open Memory Interface (OMI)-compliant 25 Gbps lanes and bridges to memory via a 72-bit DDR4 3200 interface. This reduces the number of host CPU or SoC pins per DDR4 memory channel, which allows for more memory channels and increasing the memory bandwidth available.

Acting AG Whitaker

Citing Dell-EMC research, Microchip said current processor DRAM buses are limited in number and have limited performance scaling and multi-core processor devices are experiencing increased latency per core. (Source: Microchip)

Emerging workload demands means it’s no longer feasible to use legacy interfaces such as PCIe, SAS, or SATA, according to Bennett. “Those interfaces are too inefficient. That's why these new interconnects are required.” Microchip has opted to use the OMI standard — championed by IBM, AMD, and Google, among others — on one side of the device, while the other side of the device is a 72-bit DDR 4 interface. “It is a broad specification, has a lot of coherent commands, and is a very rich protocol.”

Using OMI will enable customers to take advantage of a broader ecosystem and not get locked into a single solution, said Bennett, which allows for other memories to be used, whether it’s DDR5 when it hits the market, or other storage-class memories. “The door is open for additional innovation with storage class media of whatever flavor,” he added. Because OMI is significantly more pin efficient and the IP is available on a royalty-free basis at no cost, he said, the result is reduced silicon die, IP, and packaging costs.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.