Hewlett-Packard Enterprise completed the original prototype of The Machine in November 2016, but it was canceled shortly after. Memristors are not ready yet to conquer the world, yet the idea of reducing our dependence on rigid processors and architectures in favor of giant memory banks and optical interfaces remains strong. That enabled the comeback of The Machine, now with 160 terabytes of memory, setting a new record.
The first computers and programs were designed from scarcity. Memory was expensive, as were the higher-density storage media. We were all forced to be efficient and economical, even if it meant sacrificing speed. Today the story is very different. A 64 GB memory card does not exceed 20 euros, new smartphones have a RAM capacity on par with desktop computers, and some companies already promise up to 100 TB of solid-state storage for data. Yet in the business world we can see much more advanced and extreme goals. The concept of unified memory presents itself as a first-class solution to the most critical problems of the Web, and behind that solution we find The Machine.
https://www.youtube.com/embed/AE5VitamCEQHewlett-Packard Enterprise began working on The Machine in 2014, but pulled the plug two years later after finishing its first prototype. At the time, a single server carried eight terabytes of memory. However, this reinterpretation of The Machine goes much further. It is a new architecture that seeks to eliminate delays associated with bottlenecks, latency, and internal transfers by placing everything at the software level within a huge memory bank available without restrictions. The prototype has a total of 160 terabytes of unified memory (read: “a single address space”), and 40 Cavium ThunderX2 (ARM) processors with 32 cores each, bringing the final number to 1,280. To this we must add support for photonic interfaces, a technology that will enable truly impressive scaling in the future.
Right now, The Machine has its 160 terabytes distributed across 40 nodes, but HPE’s official announcement throws out staggering numbers. They have suggested a maximum of 4,096 yottabytes, something like our entire current digital universe… multiplied 250,000 times. So the question: how long must we wait to see this technology in data centers? Between three and five years.
