SK Hynix, Solidigm, Fadu And Microchip Announcements At The 2023 FMS
This is the second of a few articles covering announcements and keynote talks at the event. In addition to the keynote talks there was a full day of tutorial talks on Monday and many parallel sessions and a large exhibit hall from Tuesday through Thursday. In this piece we will look at keynote talks by SK hynix and its Solidigm subsidiary as well as controller companies Fadu and Microchip (with a cameo appearance by Micron).
Jungdal Choi, Executive VP of NAND Development at SK hynix spoke about the company’s “4D” NAND technology. SK hynix currently offers the highest NAND flash stack count at 238 layers as well as a 30% reduction in the size of logic under the NAND cells from the company’s previous generation product. The company started mass production of 238-layer NAND products in March 2023, first in mobile devices as shown below.
SK hynix 512Gb TLC Launch (238-layer NAND)
The company announced that it had made the first 321-layer NAND flash, shown in the image below. This involved correctly overlaying 3 plugs of cell holes and developing a stress free material to enable higher layer stacks.
SK hynix 321-layer 4D NAND
SK hynix plans to start transferring its 321-layer NAND into mass production starting in 1H 2025. The product provides 41% bit growth, 13% lower read latency, 12% faster program performance and 10% better read power efficiency.
In order to provide higher capacities and performance required for AI workflows the company is developing multi-site cells and shared bit lines to 5 bit per cell (PLC) write times that equal three bit per cell (TLC) NAND (effectively treating the single PLC cell as two 2.5 bit cells. SK hynix sees this as a path enabling higher bit per cell (e.g. PLC) implementation in future NAND generations.
The company said that their PCIe Gen 5 PS1010 and PS1030 enterprise SSDs using their proprietary controller architecture began shipping in June 2023 with significant performance improvements. The company is also working on HW/SW co-designs and what they call an h-TPU to introduce even faster PCIe Gen6 products from SK hynix and Solidigm by 2026. They are making PCIe Gen4 client SSDs today with PCIe Gen5 products planned for 2024. The Solidigm Synergy software can increase performance and value even more.
SK hynix is introducing UFS 4.0 products and is preparing UFS 5.0 products that include the company’s H-TPU technology as well as E2E HW automation to provide a 2X improvement in sequential performance and a 5X improvement in random performance. It is using these features along with its V7 4D NAND to provide future UFS products to the automotive market.
SK hynix is looking into enhanced key-value computational storage devices for AI data needs. It also has CXL products (shown below) for memory expansion modules, including bandwidth memory expansion (BME) and capacity memory expansion (CME) as well as computational memory solutions (CMS). SK hynix also has its AiM processing in memory (PIM) products for AI acceleration, which are sampling at customers.
SK hynix CXL Products
The SK hynix talk was followed by a talk by Robbie Frickey, a Solidigm Fellow on storage from the core to the edge. He indicated that the amount of data stored at the edge is growing faster than that in traditional data centers. He thinks that SSD shipped capacity will grow driven by increasingly read and data-intensive workloads that require high performance. As an example of the growth of data at the edge he said that advanced production automotives will generate 2-4TB that need to be processed. The image below shows projections for the growth of data centers and edge servers to support the growth of vehicle generated data.
Solidigm Estimates for Transportation Storage
He made the argument that higher density storage is needed at the edge and near the edge that favor SSDs and that SSDs may make further inroads on HDDs if data center users start to favor higher density SSDs (particularly using QLC flash). Below are Solidigm’s QLC SSD products for enterprise and data center applications.
Solidigm QLC SSDs
Fadu presented the following keynote. The make SSD controllers. The image below shows developments in Open Compute Project (OCP) data center NVMe SSDs, which is including more players as time goes by. Telemetry and debug are increasingly important features for data center applications and OCP has developed improved methods for performing these functions.
Fadu OCP Data Center NVMe SSD Spec History
Fadu also spoke about the importance of flexible data placement (FDP) to reduce write amplification (which wears out the flash memory cells), which has been a big issue in bringing QLC SSDs into the data center. OCP FDP methods are the result of Meta and Google cooperation on finding to reduce write amplification. Standard SSD wear can be reduced by 2-3X using FDP, allowing longer SSD life and the write performance also improved by about the same amount. The OCP is now working on open-source SSD qualification testing to standardize this approach between SSD customers. Fadu talked about their FC5161 PCIE Gen5 controller. Details on this controller are shown below.
Fadu FC5161 PCIe Gen 5 Controller
Fadu said that using their controller reduced the operating temperature of the SSD controller by 20 degrees C and reduced controller power from 9-10W down to 7W. Temperature is important because thermal throttling reduces SSD performance if the temperature gets too high. They also said that this controller provides lower write amplification and higher write throughput, presumably using the OCP FDP methods. Fadu said that customers can develop the implement their own firmware for the company’s controllers starting in 2024.
The company has additional plans for creating PCIe Gen6 and CXL memory controllers and CXL switches in the next few years as shown below.
Fadu Future Controller and Switch Plans
Samer Haija from Microchip spoke about CXL during their keynote. CXL is a standardized communications channel running over PCIe for system memory and co-processor expansion using shared memory. It allows sharing of memory between processors and will help make sure that expensive memory is more fully utilized. Samer then introduced Steve Pawlowski from Micron. Steve spoke further about CXL and in particular CXL 3.0 that enables memory shared on a switched network. As shown below, this enables scale up and scale out shared memory systems.
Micron and Microchip Plans for Memory Centric Computing
SK hynix announced 321-level NAND by 2025 and CXL products as well as UFS QLC Solidigm products. Fadu and Microchip spoke about their controllers including CXL and PCIe 5.0 products.