The CXL consortium has had a daily presence at FMS (which rechristened itself from ‘Flash Reminiscence Summit’ to the ‘Way forward for Reminiscence and Storage’ this yr). Again at FMS 2022, the corporate had introduced v3.0 of the CXL specs. This was adopted by CXL 3.1’s introduction at Supercomputing 2023. Having began off as a bunch to machine interconnect commonplace, it had slowly subsumed different competing requirements equivalent to OpenCAPI and Gen-Z. Consequently, the specs began to embody all kinds of use-cases by constructing a protocol on high of the the ever present PCIe growth bus. The CXL consortium contains of heavyweights equivalent to AMD and Intel, in addition to a lot of startup firms making an attempt to play in numerous segments on the machine facet. At FMS 2024, CXL had a first-rate place within the sales space demos of many distributors.
The migration of server platforms from DDR4 to DDR5, together with the rise of workloads demanding massive RAM capability (however not significantly delicate to both reminiscence bandwidth or latency), has opened up reminiscence growth modules as one of many first set of extensively accessible CXL gadgets. Over the past couple of years, we’ve had product bulletins from Samsung and Micron on this space.
SK hynix CMM-DDR5 CXL Reminiscence Module and HMSDK
At FMS 2024, SK hynix was displaying off their DDR5-based CMM-DDR5 CXL reminiscence module with a 128 GB capability. The corporate was additionally detailing their related Heterogeneous Reminiscence Software program Improvement Equipment (HMSDK) – a set of libraries and instruments at each the kernel and person ranges aimed toward rising the benefit of use of CXL reminiscence. That is achieved partially by contemplating the reminiscence pyramid / hierarchy and relocating the info between the server’s important reminiscence (DRAM) and the CXL machine primarily based on utilization frequency.
The CMM-DDR5 CXL reminiscence module comes within the SDFF form-factor (E3.S 2T) with a PCIe 3.0 x8 host interface. The inner reminiscence is predicated on 1α know-how DRAM, and the machine guarantees DDR5-class bandwidth and latency inside a single NUMA hop. As these reminiscence modules are meant for use in datacenters and enterprises, the firmware contains options for RAS (reliability, availability, and serviceability) together with safe boot and different administration options.
SK hynix was additionally demonstrating Niagara 2.0 – a {hardware} resolution (at present primarily based on FPGAs) to allow reminiscence pooling and sharing – i.e, connecting a number of CXL reminiscences to permit totally different hosts (CPUs and GPUs) to optimally share their capability. The earlier model solely allowed capability sharing, however the newest model permits sharing of information additionally. SK hynix had offered these options on the CXL DevCon 2024 earlier this yr, however some progress appears to have been made in finalizing the specs of the CMM-DDR5 at FMS 2024.
Microchip and Micron Show CZ120 CXL Reminiscence Enlargement Module
Micron had unveiled the CZ120 CXL Reminiscence Enlargement Module final yr primarily based on the Microchip SMC 2000 sequence CXL reminiscence controller. At FMS 2024, Micron and Microchip had an indication of the module on a Granite Rapids server.
Further insights into the SMC 2000 controller have been additionally supplied.
The CXL reminiscence controller additionally incorporates DRAM die failure dealing with, and Microchip additionally offers diagnostics and debug instruments to investigate failed modules. The reminiscence controller additionally helps ECC, which varieties a part of the enterprise class RAS characteristic set of the SMC 2000 sequence. Its flexibility ensures that SMC 2000-based CXL reminiscence modules utilizing DDR4 can complement the primary DDR5 DRAM in servers that assist solely the latter.
Marvell Pronounces Structera CXL Product Line
A couple of days previous to the beginning of FMS 2024, Marvell had introduced a brand new CXL product line below the Structera tag. At FMS 2024, we had an opportunity to debate this new line with Marvell and collect some further insights.
In contrast to different CXL machine options specializing in reminiscence pooling and growth, the Structera product line additionally incorporates a compute accelerator half along with a memory-expansion controller. All of those are constructed on TSMC’s 5nm know-how.
The compute accelerator half, the Structera A 2504 (A for Accelerator) is a PCIe 5.0 x16 CXL 2.0 machine with 16 built-in Arm Neoverse V2 (Demeter) cores at 3.2 GHz. It incorporates 4 DDR5-6400 channels with assist for as much as two DIMMs per channel together with in-line compression and decompression. The mixing of highly effective server-class ARM CPU cores implies that the CXL reminiscence growth half scales the reminiscence bandwidth accessible per core, whereas additionally scaling the compute capabilities.
Purposes equivalent to Deep-Studying Suggestion Fashions (DLRM) can profit from the compute functionality accessible within the CXL machine. The scaling within the bandwidth availability can also be accompanied by diminished vitality consumption for the workload. The strategy additionally contributed in the direction of disaggregation throughout the server for a greater thermal design as a complete.
The Structera X 2404 (X for eXpander) can be accessible both as a PCIe 5.0 (single x16 or two x8) machine with 4 DDR4-3200 channels (as much as 3 DIMMs per channel). Options equivalent to in-line (de)compression, encryption / decryption, and safe boot with {hardware} assist are current within the Structera X 2404 as effectively. In comparison with the 100 W TDP of the Structera X 2404, Marvell expects this half to eat round 30 W. The first goal of this half is to allow hyperscalers to recycle DDR4 DIMMs (as much as 6 TB per expander) whereas rising server reminiscence capability.
Marvell additionally has a Structera X 2504 half that helps 4 DDR5-6400 channels (with two DIMMs per channel for as much as 4 TB per expander). Different points stay the identical as that of the DDR4-recycling half.
The corporate burdened upon some distinctive points of the Structera product line – the inline compression optimizes accessible DRAM capability, and the three DIMMs per channel assist for the DDR4 expander maximizes the quantity of DRAM per expander (in comparison with competing options). The 5nm course of lowers the facility consumption, and the elements assist accesses from a number of hosts. The mixing of Arm Neoverse V2 cores seems to be a primary for a CXL accelerator, and permits delegation of compute duties to enhance general efficiency of the system.
Whereas Marvell introduced specs for the Structera elements, it does seem that sampling is at the very least a couple of quarters away. One of many fascinating points about Marvell’s roadmaps / bulletins in recent times has been their concentrate on creating merchandise tuned to the calls for of high-volume prospects. The Structera product line is not any totally different – hyperscalers are hungry to recycle their DDR4 reminiscence modules and apparently cannot wait to get their palms on the expander elements.
CXL is simply beginning its sluggish ramp-up, and the hockey stick section of the expansion curve is unquestionably positively not within the close to time period. Nonetheless, as extra host methods with CXL assist begin to get deployed, merchandise just like the Structera accelerator line begin to make sense from a server effectivity viewpoint.
