Micron has launched reminiscence enlargement modules that assist the two.0 era of Compute Specific Hyperlink (CXL) and include as much as 256GB of DRAM operating over a PCIe x8 interface.
CXL is an open interconnect normal with broad trade assist that’s meant to be a connection between machines permitting for the direct sharing of contents of reminiscence. It’s constructed on high of PCI Specific for coherent reminiscence entry between a CPU and a tool, comparable to a {hardware} accelerator, or a CPU and reminiscence.
PCIe is often utilized in point-to-point communications, comparable to SSD to reminiscence, whereas CXL will ultimately assist one-to-many communication. To this point, CXL is able to easy point-to-point communication solely.
Growth on the CXL normal started in early 2019, but it surely has solely just lately come to market as a result of it required a quicker PCIe bus in addition to native assist from CPU distributors Intel and AMD. Solely their most up-to-date CPUs assist it.
For now, the preliminary functions revolve round attaching DRAM to a PCIe interface. That is what Micron is providing with its CZ120 reminiscence enlargement modules. The modules can be found in 128GB and 256GB capacities, which is fairly huge for a reminiscence module. They use a particular dual-channel reminiscence structure able to delivering a most of 36GB/s of bandwidth.
Ryan Baxter, senior director of the info heart phase at Micron, mentioned security measures have been paramount on this launch. “There are plenty of security measures in 2.0 that do not exist or usually are not supported in 1.1,” Baxter mentioned.
Safety is necessary when you’ve gotten servers speaking to one another. The CXL 2.0 normal now helps any-to-any communication encryption by way of using {hardware} acceleration constructed into the CXL controllers. Which means silicon suppliers shouldn’t have to construct encryption safety into their very own {hardware}.
Baxter mentioned {that a} lack of safety is why prospects testing and deploying CXL 1.1 tended to solely experiment and use decrease capability reminiscence. Some prospects may need deployed CXL 1.1 in some inner workloads, however many averted doing something actually bold whereas they waited for two.0, he mentioned.
CXL 2.0 will even assist persistent reminiscence, which shops information and reminiscence like NAND flash however is far quicker, nearly as quick as DRAM. CXL 2.0 permits distinct PMEM assist as a part of a sequence of pooled assets.
Micron sees two key major use circumstances for CXL 2.0: including reminiscence to a system to supply additional reminiscence to a CPU beneath heavy workload; and supporting bandwidth intensive workloads, for the reason that PCIe spec is definitely quicker than reminiscence slots.
So, guess which workloads they take into account? “We’re seeing [interest] with AI coaching and inference, the place these use circumstances are driving a a lot larger reminiscence footprint across the CPU,” mentioned Baxter. He additionally cited extra conventional makes use of, comparable to in reminiscence databases, as benefiting from CXL 2.0 reminiscence capability.
CXL has plenty of deliberate obsolescence in it. The two.0 model just isn’t backwards suitable with 1.1, and the three.0 model won’t be suitable with 2.0, since 3.0 makes use of the subsequent era of PCIe. Baxter doesn’t count on to see vital 2.0 use and product availability till a minimum of subsequent yr if not 2025.
Copyright © 2023 IDG Communications, Inc.
