Gen-Z emerging technology for memory intensive applications

Library Content Type:
Publish Date: 
Saturday, December 5, 2020
Event Name: 
Event Track:
Focus Areas:

In modern day technologies like Artificial Intelligence, Machine Learning algorithms and in-memory database application, the biggest challenge is how to increase the memory available to the CPU.

What is the need of new protocols like Gen-Z, CXL, CCIX ?
As Real estate has already saturated on most of the CPU motherboards. It is difficult to add more high-speed signals on PCB. This not only increase cost of manufacturing, but also adds engineering challenges for silicon and package manufacturers. So adding more DIMM slots on the motherboard is not a scalable solution.

There is a trend to use serial protocols for memory as well, but it comes with its challenges of latency and throughput.
How Gen-Z solves the problem ?
(1) Gen-Z re-defines the Data-Link layers and Transport layers of serial interface, but reuses same Physical layers from PCIe or IEEE 802.3 standard. This enables OEM and hardware manufacturers to use same physical connectors or network cables for Gen-Z, as for other serial protocol like PCIe or high-speed Ethernet and does not require re-design of the board.
(2) Contrary to other fabric protocols like RDMA or NVMe-over-Fabric, Gen-Z defines its own packet formats, command structures, and handshake mechanism. Thus, it removes all the overhead of network stack like TCP/IP or UDP framework and improves the media utilization for load/store memory semantics.
*Benefits and use-cases*
Gen-Z is much more efficient in transferring small size data at cache line granularity, and has low latency than other fabric protocols.

Most High Performance Compute (HPC) workloads whether is research about space, or finding the cure for COVID with right protein combination requires multiple CPU to run simulation using same data-sets stored in large pool of RAM. This is where Gen-Z is helpful.

Gen-Z is helpful in virtualized environment as it allows administrators to re-allocate the memory pool based on end-user requirements without adding more memory to each server node. This increases the utilization of your memory resources, and give you best return on your investment.

Watch video: