Feeding the XPoint cuckoo and finding it a place in the storage nest

XPoint market

The two promotors say it is in-memory computing where XPoint will fit, and suggest applications such as gaming, Big Data analytics and high-performance computing. We think this means servers involved in analytics and HPC can, in effect, scale up memory with less need for multiple servers to do parallel IO-to-disk stores. If the data is in memory – or closer to compute, as they say – then access is not held up by network interface and disk latencies.

Up until now, in-memory computing has meant data being held in DRAM, whch is both expensive and volatile. This means its application use has been restricted to high-pay-off use cases like financial trading, where the potential profits can justify its eye-watering cost.

What will be the incentive for server manufacturers to develop and sell XPoint servers into the analytics and HPC markets? If they sell 1,000 servers to do an HPC or analytics job now, but will only need 100 XPoint servers to do the same job in future, then they have a revenue and pricing problem. That is unless they ae convinced that the overall XPoint server market will be substantially bigger than the existing HPC/analytics server market.

Micron and Intel have to convince server suppliers that they can sell more XPoint servers than traditional servers – or sell them more profitably – for this to work.

The server suppliers also have to work out how to use the new memory-storage hierarchy, what software changes are needed and what amounts of DRAM, XPoint, NAND and disk, if any, are needed. It's not simple.

Is there a storage supplier play here? Intel and Micron are talking about getting data closer to compute with XPoint. Its speed is such as to make having network interface latency between it and compute a ridiculous idea.

Unless the storage supplier is involved with NVMe fabric-accessed shared storage, then there is no play here. For EMC (DSSD) and Pure Storage – who both have NVME fabric initiatives – and any similar storage suppliers, then XPoint deserves immediate attention.

XPoint host software changes

If XPoint use as storage-class memory and as a tier alongside DRAM, NAND and disk requires software changes, then we are going to wait a decade or more for Windows to support it. Redmond's software development schedule has been slowed by the need for legacy system support.

Faster operating system support could come from Apple, open source software like Linux, and gaming software.

If Apple were convinced that XPoint Macs and iPads could take more market share from Windows PCs, notebooks and tablets, then it could rewrite Mac OS and IOS to support XPoint and rely on Apple customers' willingness to pay for an even-better Apple experience.

Such a calculation might even convince Apple to think about XPoint iPhones at some stage in the future, turning these devices into even hotter, more capable computing engines for such things as voice recognition.

The enterprise in-memory, HPC and analytics possibilities could be served with XPoint-supporting Linux and hypervisor software. Imagine what an XPoint-equipped HCIA (hyper-converged infrastructure appliance) from Nutanix or SimpliVity coud do compared to a standard X86/DRAM/NAND/Disk HCIA. If XPoint lives up to its billing, then an XPoint HCIA could offer a two to five times the application payload performance, or possibly even more.

Gaming software means Sony, Nintendo and Microsoft (Xbox) etc. If game developers could see that XPoint-equipped hardware would provide a significantly better gaming experience – meaning faster, more detailed, more immersive and broader environments – then they would be motivated to develop XPoint gaming hardware and software. They would also need to be convinced that gamers would pay the higher prices needed for such gaming.

XPoint now

XPoint changes nothing now, as at the time of writing it is largely vapourware. Some degree of sample chipmaking is going on, however, and Micron and Intel say they will independently develop products. They suggest we'll see chips in production in 2016, with product, including such things as NMVMe XPoint cards, we imagine, towards the end of that year.

Intel and Micron have to give out XPoint chip and device speeds and prices. Until they do that we are in an XPoint fantasy and the dream will fade. Come on guys! Give us die size, random read and write IOPS, read and write bandwidth, latency, capacity, workload, MTTF, etc. You know the score. ®