Intel broke its silence yesterday on technology aspects of Micron’s decision to pull the plug on the 3D XPoint memory that the two manufacturing partners had developed to fill the performance gap between NAND flash and DRAM.
“We were as surprised as anyone when Micron made their announcement,” said Jim Pappas, director of technology initiatives at Intel, during a live panel discussion yesterday at the SNIA Persistent Memory and Computational Storage Summit.
Panel moderator Dave Eggleston, principal of Intuitive Cognition Consulting, referred to Micron’s March news as the “elephant in the room” at the outset of the session entitled “What does the future hold for persistent memory?”
Micron announced on March 16 that it would immediately cease development of 3D XPoint and shift resources to new types of memory-centric products that use the emerging Compute Express Link (CXL) standard to connect compute, storage and any type of memory. Persistent memory technologies such as 3D XPoint target some of the most demanding workloads, including high-performance computing, databases, virtualized infrastructure, AI and analytics.
Intel’s Pappas, who chairs the CXL board, said Micron’s shift away from 3D XPoint in favor of CXL with other memory technologies was a “non sequitur.”
“In my mind, 3D XPoint and CXL are made for each other,” Pappas said.
3D XPoint on CXL
Pappas revealed Intel product plans “have 3D XPoint going on to CXL in the future.” He stressed that he is not part of Intel’s product teams but asked the general manager if he could mention the roadmap, which Intel has yet to publicly disclose. Intel sells its 3D XPoint-based SSDs and persistent memory modules under the Optane brand name.
Although Micron may have left Pappas with the impression that it saw 3D XPoint and CXL as “one versus the other,” Micron was well aware that was not the case. Late last year, Saeed Raja, senior director of product management and marketing in Micron’s emerging products group, sang the praises of new interfaces such as CXL in the context of 3D XPoint DIMMs that he said were on the company’s roadmap.
Raja said CXL could open the door to 3D XPoint memory modules that could work with AMD and Nvidia processors, not simply Intel’s CPUs through its proprietary DDR-T interface.
Micron chooses different CXL path
CXL is an open standard that allows multiple CPU vendors and multiple memory companies, like Micron, to innovate.
“It is absolutely true that you can attach 3D XPoint on CXL, and that hasn’t changed,” Rajeeb Hazra, senior vice president and general manager of Micron’s compute and networking business, said in a recent exclusive interview.
“The question really is how much value did we see in attaching and focusing on 3D XPoint versus other alternatives that CXL enabled,” Hazra said. “As we looked at our technology capabilities, talked to customers and looked at multiple options, we came to the conclusion that it is better for our customers and better for us to go down a path of different kinds of things attached to CXL in the memory space than 3D XPoint, and, therefore, we terminated our 3D XPoint program.”
Hazra said Micron found 3D XPoint to be an expensive investment without a “broad, big market” in the near-term foreseeable future. He said Micron is working on “better alternatives” to address customers’ problems but declined to disclose what they are.
Mark Webb, principal at MKW Ventures Consulting, speculated during yesterday’s panel discussion at the SNIA event that Micron would focus on putting NAND and DRAM on the CXL bus for hyperscale and data center customers. Meanwhile, Intel would likely need to subsidize Optane/3D XPoint technology — and “make it up on the system sales” — to keep prices at less than half the cost of DRAM, Webb said.
“One of the challenges with Micron not ramping the technology is that it probably won’t hit, anytime in the next three years, the kind of volume it would need to get totally competitive costs,” Webb said. “But the technology itself can easily be much more cost-effective than DRAM if you get the scale up probably to something around two to four times what it is right now.”
Persistent memory market to soar
Intel’s Optane/3D XPoint is a type of phase change memory (PCM). A joint report from Objective Analysis and Coughlin Associates projects the market for persistent memory will exceed $36 billion by 2030, with PCM and magnetoresistive RAM (MRAM) leading the way.
Micron and Intel began their joint development work on 3D XPoint in 2012 as a lower cost alternative to faster DRAM and a higher performance option than cheaper NAND flash. Micron later bought out Intel’s interest in the Utah-based fabrication plant that’s dedicated to 3D XPoint production, and the partners ended their joint development program last year. Intel continued to buy 3D XPoint wafers from Micron under a contractual agreement but faces decisions now that Micron is selling the Utah fab.
Intel’s brief public statement on March 16 was simply: “Micron’s announcement doesn’t change our strategy for Intel Optane or our ability to supply Intel Optane products to our customers.”
“We’ve known this for some time that we will be doing our own manufacturing. We’re already manufacturing parts in New Mexico, and that will continue,” Pappas said yesterday, adding that the company is confident it will be able to make the 3D XPoint it needs to supply the market.
Pappas mentioned the $20 billion investment in new factories that CEO Pat Gelsinger announced within weeks of joining Intel. He said manufacturing across different and new fabs “opens up new places where we could do persistent memory.”
Competitors could emerge to challenge Intel in the persistent memory space. Pappas noted that any type of persistent memory could, with a controller, attach to CXL. He said he is aware of at least a half dozen companies, including startups and established vendors, exploring the prospects.
“There will be controllers that come out, just like lots of controllers came out for SSDs,” Pappas said. “No company had an exclusive SSD controller. So, there’s nothing stopping it. Now, we have a standardized part that we could connect it to — probably any CPU architecture — via CXL. And we know that it’s got native support for persistent memory.” TechTarget