By: Michael Feldman

Changes are in store for “The Machine,” the R&D effort at Hewlett Package Enterprise that recasts the computer as a memory-centric system. According to a blog post by HPE CEO Meg Whitman, the project, along with HP Labs, the unit that manages the The Machine's development, will be placed under the Enterprise Group. The decision is related to the exit of HPE CTO Martin Fink, who will be retiring from the company at the end of the year.  

Fink, who heads HP Labs, and who spearheaded the research and development of The Machine, has been with Hewlett Packard (and now HPE) for more than 30 years. He announced the original vision for the project back in 2014. Now that HP Labs is being folded into the Enterprise Group, Antonio Neri, who heads that division, will be in charge of The Machine, along with the other R&D efforts the lab is involved with. Just for a little perspective, the Enterprise Group brought in about $27.9 billion of revenue in FY2105, while R&D spending at the company chewed through about $2.3 billion over the same 12 months

Putting R&D under the control of a profit center might seem counter-intuitive, but in the particular case of The Machine, it makes a certain amount of sense. Although the project isn’t expected to yield a full-blown commercial offering before the end of the decade, some of the component technologies will be incorporated into HPE’s server portfolio well before that. In theory, those transitions are likely to happen more smoothly if the business and R&D organizations are more closely aligned.

At the top of that list for commercialization is the silicon photonics technology that will debut in The Machine prototype, which is scheduled to be unveiled later this year. The silicon photonics uses optical interconnects to link together nodes in such a way as to enable large amounts of distributed memory to be shared as a global pool. The technology, which is driven by HPE’s T1 optical module, is capable of transferring data at the rate of 1.2 terabits/second. That's than 10 times the speed of an 100G InfiniBand or Ethernet connection.

 

 

In the prototype, 240 GB of non-volatile memory, what HPE is calling “fabric memory,” will be globally accessible across the 2,500 cores that will power the rack-sized system. The fabric aspect comes into play because the super-speedy X1 will be used to knit together the shared memory into a globally available resource. An additional 80 GB of DRAM will be spread across the rack to be used as local memory in each node.

When The Machine was first envisioned by Fink two years ago, all memory in the system was going to be provided by memristors, which was being lauded as a “universal memory.” The idea was that the non-volatile memristors were so fast and scalable, they would replace both NAND and DRAM, offering a vast amount of ultra-fast general-purpose storage. Manufacturing the devices proved problematic, however, so the company changed the architecture of The Machine, at least for the time being, to employ a mix of conventional DRAM, and current non-volatile memory components.

The photonics technology has fared much better. T1 silicon has been taped out, and its appearance in the prototype is a precursor to its commercial launch, which HPE expects to happen in 2017. Other componentry from the project, including software, will find their way into commercial offerings as well. And if HPE ever figures out how to perfect memristors, it’s a good bet they’ll be in every machine the company sells.