Cutting Hewlett-Packard Labs down to size
Addressing the 3D XPoint elephant in the storage-class memory room
Comment When Martin Fink resigned from his positions at HPE in August 2016, the announcement said: "Martin Fink, our chief technology officer and head of Hewlett-Packard Labs, will be retiring from HPE at the end of the year, after more than 30 years with the company." He was retiring, we were told, but he was just 51.
In Fink's time running HP Labs, and then Hewlett-Packard Labs after the HP-HPE split, he helped HP embrace Linux, launched its Helion public cloud service and spearheaded The Machine development and associated Memristor memory technology. The Helion cloud has been discontinued. The Memristor development is much-delayed and The Machine prototype supports Optane (3D XPoint) DIMMs while HPE has a ReRAM development effort with Western Digital, which actually inherited it when it bought SanDisk, with whom HPE set up the deal, in 2016.
The Machine project was reviewed publicly at the end of last year and its memory-driven computing ideas will find a home in the roadmaps for HPE's servers but likely will not result in an actual Machine product.
Fink's retirement announcement also said: "We will move Hewlett-Packard Labs into the Enterprise Group under Antonio Neri. This move will also help align our R&D work on The Machine with the business – particularly how we integrate key components like photonics and Memristor into existing product lines – by bringing together our innovation roadmap with our business roadmap."
Translation: the innovation roadmap was not well co-ordinated with the business roadmap before.
Pre-split HP Labs was run by Fink and covered all of the HP product areas, from printers to PCs to networking to servers and so forth. Post-split the printers-to-PCs HP had its HP Labs run by Shane Wall while HPE had Hewlett-Packard Labs run by Martin Fink, and that no longer had printer and PC/laptop/tablet-related activities. Fink, along with Enterprise Group GM Antonio Neri and other execs, reported directly to CEO Meg Whitman.
The Helion cloud was canned in October 2015. The Memristor development has consumed years of time and money and not resulted in licensable IP or revenue-earning product. By cutting a deal with SanDisk over ReRAM it seemed to many that HPE was signalling that Memristor tech development was being pushed on to the back-burner. The Machine was a huge research project that, so far, has not resulted in usable technology and the fruits of which will now be spread across HPE's server product set as it progresses towards memory-driven computing.
We feel Fink, at the age of 51, was persuaded to leave or pushed out so that Hewlett-Packard Labs costs could be better controlled, and Neri could be given overall control of it inside his Enterprise Group.
So Fink, facing a shrunken Labs structure and, we think, a demoting transfer into Neri's Enterprise Group, "retired". Just over four months later, he has un-retired, and taken up the CTO role at Western Digital.
We think Hewlett-Packard Labs will slowly shrink in size and influence and focus less on blue-sky developments and more on things that have a foreseeable ROI. Neri's Enterprise Group will pay for Hewlett-Packard Labs and will surely want a return on this money, although calculating its ROI must be an exercise in creativity.
This could result in an exodus of talent from Hewlett-Packard Labs as researchers and scientists there see career options becoming limited.
Vendor labs and memory tech
El Reg thinks the days are long gone when a server manufacturer can invent and productise a memory technology, if they ever existed in the first place after IBM and magnetic core memory. There are three main reasons we see for this.
One is that the market for such technology – chips of some sort – is made up of server manufacturers, and server maker A isn't likely to license memory technology from competing server maker B, not when alternatives from independent fabs are available. Secondly, the in-house demand for such chips is not likely to be high enough to create enough volume to make the fab cost affordable.
The technology and process development difficulties are so complex that specialised and experienced chip foundry partners are needed yet the main ones are focussed on cross-server OEM markets and not single-vendor products. Even though HPE partnered with SK Hynix to develop Memristor chips, the likelihood of other server manufacturers buying Memristor chips from SK Hynix was vanishingly low, particularly once Intel and Micron launched 3D Xpoint memory.
So the memory technology development focus has shifted to the DRAM and flash fab operators. By selling their product to many server vendors they get the volume they need to drive down per-chip costs.
Their job is to do a better job of productising their next-gen memory technology better, meaning faster and with a road to affordability, than their competitors.
There are too many candidate technologies offering their own claimed combination of DRAM-class speed, non-volatility, density, density roadmap, and affordability, and there will have to be a cull in the ranks of:
- 3D XPoint
- Carbon nanotube memory
Eight candidates to fill the DRAM-NAND gap is far too many. Fink has a fight on his hands to convince server vendors and others that Western Digital's storage memory candidate tech, ReRAM, is the one to choose. We think we can say HPE is on board the ReRAM train. If WD can sign up another major OEM then that would be a great honking signal that ReRAM's day has arrived.
The elephant in this storage-class memory room is Intel and Micron's 3D XPoint. We look forward very, very much to Western Digital's positioning and comparison data for its ReRAM versus 3D XPoint. XPoint has to be demolished by ReRAM or sidestepped. It is going to be fascinating to see what happens and how Western Digital takes on the challenge. ®