SK Hynix has announced it will acquire Intel’s NAND memory business for $9B, including Intel’s NAND SSD foundry, its component and wafer business, and the manufacturing facility Intel built in Dalian, China. The announcement specifically states that Intel will retain its Optane business.
Not the Fastest Timeline
The two companies envision a protracted sale process. SK Hynix and Intel hope to have regulator approval for the merger by late 2021. Once approval has been granted, SK Hynix will pay Intel $7B for its NAND business, associated IP, employees, and the Dalian foundry. Let’s say this happens by early 2022. The agreement then calls for an extended period in which SK Hynix owns most (but not all) of Intel’s NAND IP.
In March 2025, SK Hynix pays Intel the last $2B and receives “the remaining assets, including IP related to the manufacture and design of NAND flash wafers, RD employees, and the Dalian fab workforce.” SK Hynix is buying the building in 2022, but it won’t be responsible for the workers until March 2025. Under this arrangement, Intel will build NAND at Dalian until the end of Q1 2025 and will retain all of the IP related to the manufacture of NAND until the last part of the transaction is finalized.
The Intel Non-volatile Memory Solutions Group (NSG) had a rough 2019, which isn’t surprising, given that the entire NAND flash market took a beating that year. What’s surprising is that Intel would sell the business now, given the bump in revenue NSG has enjoyed during the pandemic. First, here are results for the last few years and the total revenue provided by the unit.
Here’s a more specific breakdown of last quarter, contrasting the massive growth in NSG with the much smaller growth in the Programmable Solutions Group (PSG).
PSG is responsible for Intel’s FPGA business, including the Agilex and Stratix brands. NSG is responsible for NAND flash and Optane (which Intel will retain). Intel is reportedly planning to get out of the NAND business due to the sagging price of flash, which raises lots of questions about… well, everything, really. The recent bust in NAND flash prices wasn’t all that surprising; both NAND and DRAM are subject to boom-and-bust cycles. Intel is well aware of this.
First, foundries build out more capacity to meet existing high demand. Then, as the market absorbs that capacity, prices fall. Eventually, supply becomes constrained, either by total worldwide production capacity or by manufacturing difficulties and low yields, all of which can raise prices. Manufacturers then find new approaches to the technical problems (boosting yield) or build new fabs (boosting absolute output), and the cycle repeats again.
Intel may be attempting to offload its memory business while the unit looks particularly strong. In its Q2 2020 conference call, the company said:
NSG’s record quarterly revenue of approximately $1.7 billion was up 76% year-on-year on strong NAND bit growth and improved pricing. Q2 was an all time record for quarterly revenue for our memory business. The business also returned to profitability this quarter, generating approximately $300 million in operating income.
Intel had previously said this year that the firm needed to generate more attractive returns from the NAND business and that it was considering a partnership of some kind, but that was before the business unit had one of its strongest quarters of all time. Intel already reduced its overall investment in memory technology when the company sold its share of the Optane / 3DXPoint-focused joint venture back to Micron for $1.5B. This $9B deal with SK Hynix dwarfs that sale.
It’s also possible (if unlikely) that Intel is reacting to ongoing pressure from the US government for US companies to draw down their investment and exposure to China, particularly in semiconductors.
What Will Happen to PLC NAND?
Intel had previously told ExtremeTech that it was continuing to work on penta-level cell (PLC) NAND, with a maximum capacity of five bits of data per NAND cell. This additional capacity increase would boost per-bit capacity by 1.25x, but at the cost of endurance. Erasing flash memory is a destructive process, and NAND cells can only be erased and rewritten so many times before they lose the ability to hold the necessary number of charge states. The more charge states a NAND cell is asked to hold, the fewer times it can be erased before becoming damaged.
Thus far, manufacturers have dealt with the decrease in longevity by over-provisioning drives and reducing write amplification with better algorithms. In order for this to be effective, the cost of overprovisioning the drive must be lower than the die cost penalty of shipping an equivalent amount of QLC or TLC storage. Yield plays a part in the overall cost analysis as well.
It’ll be interesting to see if SK Hynix picks up the proverbial torch on PLC development or if the company sticks to QLC. One interesting thing about the chart above is that the penalty for moving to the next “bit” increases the lower your initial process node for every transition except QLC – PLC. With 5Xnm technology, SLC is only 1.1x more reliable than MLC. At 1Xnm, it’s 3.5x more endurant. Similarly, the gap between TLC and QLC is much smaller at 5Xnm than it is at 1Xnm. It is not clear if PLC’s very different figures reflect the early stage of development or a shift in technology. SK Hynix will command a significantly larger share of the NAND market by the time Intel’s fabs are integrated into its own production.
- With QLC, SSD Manufacturers Target Hard Drives’ Final Strength: Raw Capacity
- Intel Working on 144-Layer QLC NAND, PCIe 4.0 for Optane: Rumor
- Future Versions of Windows 10 Will Warn You Before Your SSD Fails