HPC

Micron fires off some in-memory flash data rockets

New unit will go beyond selling chips and standard components


Possible pathway

We futurologists at The Register Think Tank believe Micron will do this not by building and branding its own systems, taking on mainstream server and archival array vendors directly.

Instead it will seek partnerships with, for example, white box server manufacturers who sell through tier 2 and 3 server OEMs. Also tier 1 OEMS with Chinese connections would be attractive; Huawei would be a great partner for entry to the Chinese market, also Lenovo.

We noted that IBN and the Micron SBU have a very good relationship, and think IBM could use 3D TLC and storage memory in future FlashSystem products.

We were shown a demo server, a 1U rackmount box, in a so-called “petting zoo". It featured an X86 CPU with 64GB of RAM, running Linux, and 12 x 8TB SSDs. This 96TB server delivers about two million IOPS and has more than 15GB/sec of bandwidth. It is more than one year old so newer SSDs should boost performance further.

Our feeling was that Micron wants the server OEMs to move faster in its direction, and the petting zoo server box serves to encourage them to get their act together as well as evangelising Micron’s message to visitors to the building.

Supermicro is already offering storage-heavy servers so loading them up with flash is a minor extension. Storage is becoming server-ised with virtual SANs. We can see 3D TLC SSDs being used for higher capacity direct-attached storage (DAS) that is a good fit to virtual SAN-type systems.

If that flash can be used as storage memory, logically resident in the same address space as the server’s DRAM, then flash/DRAM-based in-memory computing becomes more affordable and so usable in a large range of use cases beyond the SAP HANA DRAM instance.

Thomas said storage memory “is the future [and] works perfectly in a smartphone".

Software need

However, system software is needed to accomplish this. First, mask the different DRAM and NAND access latencies, and two, if possible, obviate the need for upper-level software changes in client and server systems that let users run apps of their choice in these platforms.

Thomas said his SBU has a centre of software expertise in Austin, led by Steve Moyer, coming from Dell and Panasas before that. He is a firmware/filesystem designer by trade. The centre is filled with SW engineers, software rather than firmware engineers, and with expertise at the host-based software operating system level. The filesystem expertise was emphasised.

This group could also help with adding optimisations to upper-level software so it works better with the flash.

What Thomas and the SBU want to accomplish is nothing less than a transformation of servers towards being flash-intensive in-memory data processing rockets on the one hand, and, later we think, low-latency active archive data tubs on the other.

Fast and capacious 3D TLC NAND is our pick for the data rockets, and lower-cost 3D QLC is our choice for the fast-access archive flash vaults. ®

Similar topics

Narrower topics


Other stories you might like

Biting the hand that feeds IT © 1998–2022