Intel previews Optane persistent memory, QLC SSD and even PLC

Intel has previewed a range of faster, higher-capacity, server persistent memory and flash storage solid-state storage (SSD) products to be released in the next year. They include Optane DC PMEM (datacentre persistent memory), Optane SSD, as well as “classic SSD” with quad-level cell (QLC) flash.

The Optane DC PMEM product line – codenamed Barlow Pass and announced at its Seoul storage event last week – will comprise cards that can use DIMM DDR4 slots with price and performance somewhere between DRAM and SSD.

The new generation of Optane SSDs, codenamed Alder Stream, will offer performance 50% better than existing P4800X units.

Intel will also ship SSD next year in the form of QLC 144-layer 3D Nand that will have capacities in excess of 128-layer products from competitors, according to Intel.

As with the preceding Intel persistent memory generation, which was built for Cascade Lake Xeon-equipped servers, Optane DC PMEM will be available in capacities of 128GB, 256GB and 512GB.

Paired with DRAM, that pushes memory up to a maximum of 3TB (four times DC PMEM of 512GB plus four times DDR4 of 256GB) in a mono-socket workstation or double in a bi-socket with Cooper Lake (14nm generation) and Ice Lake (new generation 10nm) CPUs (also due in 2020).

Optane DC PMEM has two main benefits.

First, it offers twice the capacity of conventional DDR4 at the same price, although not as speedily. Latency ranges from 180ns (nanoseconds) to 1,000ns, while that of DDR4 is around 70ns. Even so, DC PMEM latency is 1,000 times better than that of conventional SSD and can bring huge acceleration to machines where hard drives are used as virtual memory.

Second, using Intel’s Application Direct Mode, data isn’t lost during a reboot, such as an operating system update, and in any situation the server can restart much more quickly because its data is already in memory.

The two benefits are important for servers that run large databases, and complement the launch in June of Intel’s Distributed Asynchronous Object Storage (DAOS), which allows storage directly in memory of data usually accessed in object storage on a storage array. That’s also useful for servers running virtual machines (VMs).

Meanwhile, however, use cases for DC PMEM on workstations are yet to be determined. Beyond 2020, Intel is talking about the adoption of Optane DC PMEM in a composable memory format for desktops and laptops.

Key to the performance of NVMe Optane DC SSD is a controller that stops latency from increasing as the volume of input/output (I/O) builds. For example, if latency on a P4800X SSD is currently around 10ms at 100,000 IOPS, it is likely to climb to 20-60ms with 500,000 IOPS (its performance ceiling).

Latency of the new Optane DC SSD will remain between 8ms and 12ms for a range of IOPS between 100,000 and 800,000 IOPS.

Also, the new DC SSD will benefit from 3D Xpoint components in four layers instead of two as in the current generation, which will bring a doubling of capacity. P4800X modules currently attain 1.5TB in capacity so 3TB is likely from the next generation.

While waiting for the arrival of its 144-layer SSD, Intel will fill the gap, at the end of this year, with 3D Nand QLC but with 96 layers. That succeeds the 660P SSD in M.2 with NVMe connectivity which has 64 layers.

According to Intel information, this new product will at first be sold with similar capacities to existing versions (1TB) but with better performance. Read throughput will go from 1,200Mbps to 1,800Mbps, writes from 1,300Mbps to 1,890Mbps and write IOPS from 37,000 to 47,600.

These figures are two times lower than triple-level cell (TLC) SSD and three times lower than multi-level cell (MLC) SSD. But QLC is aimed mainly at cost-conscious use cases. These Intel SSDs come in at around $100 per terabyte, while a Samsung Pro 970 of 1TB in 3D Nand MLC pushes 3Gbps but costs around $300 per terabyte.

Intel also took advantage of the Seoul event to talk about penta-level cell (PLC) SSD, in which each cell can switch five bits. These flash chips will be lower performance and give a lower lifespan, but will cost less. They are destined for internet of things (IoT)-type use cases on devices deployed on production sites where hard drives are less frequently written to than datacentre hardware.