IBM Takes On The Memory Crunch With New FlashSystem Lineup

February 16, 2026 Timothy Prickett Morgan

There is a shortage of main memory and flash in the world right now. This is due, in large part, to the explosion in investments in yottascale AI systems that burn tens of gigawatts of juice. Such systems require a huge amount of HBM stacked memory on their GPU and XPU accelerators as well as high performance DRAM and flash in their hosts and truly enormous shared flash storage to serve up datasets.

The biggest companies buy up most of the DRAM and flash chips a year or more before they are even made, but with demand skyrocketing, the memory makers are learning to wait and see who will make the highest bids for scarce capacity and therefore drive their revenue and profits even higher.

Shops that use IBM systems along with its FlashSystem all-flash arrays are fortunate in one respect: Big Blue also buys its own flash from SK Hynix and Samsung (and possibly Micron Technology, I am not sure) and has its own flash modules created for the FlashSystems. So in a way, IBM controls its own storage fate. Hopefully, IBM acquired its flash and DRAM allocations a long time ago and planned for growth, because with the new FlashSystem arrays that were announced last week, many of the features that boost the performance of the flash as well as make it last longer are going to be in high demand in a flash-strapped world. IBM could be facing its own shortage because demand far outstrips supply.

Big Blue was a little late to the flash party in the datacenter, but it got a leg up when it acquired Texas Memory Systems way back in August 2012. The company’s RamSan appliances used a mix of IBM PowerPC processors and Xilinx (part of AMD for many years now) FPGAs to handle the processing, encryption, compression, and de-duplication in those arrays. RamSan was rebranded FlashSystem and last week the fifth generation of these flash arrays were updated with a new generation of FlashCore Modules based on QLC NVM-Express flash drives. Which look like this:

There are three new FlashSystem arrays, and they complement the FlashSystem C200 array that was announced a year ago and that I discussed in detail here. They are:

The FlashSystem 5600, in announcement letter AD26-0002
The FlashSystem 7600, in announcement letter AD26-0004
The FlashSystem 9600, in announcement letter AD26-0003

All three will be available on March 6 in most countries. The exceptions are Japan, New Zealand, and Australia, which get them on March 20; Korea, which gets them on May 5; Taiwan, which gets them on June 12; and South Africa, which gets them on September 11.

The three different FlashSystem arrays announced last week differ from each other in terms of compute capacity on the host processors (which now use X86 rather than Power chips and which are called node canisters by IBM), cache memory on those controllers (which is DRAM on the host), how many I/O slots are available on the controllers, the size of the FlashCore Modules available, and the size of the enclosure wrapping around it all.

The entry-level FlashSystem 5600 has two controllers, each with 12 cores and either 128 GB or 256 GB of DRAM uses as storage cache. The node canisters in the FlashSystem 5600 have a 10 Gb/sec/25 Gb/sec port for system management and have two adapter slots to plug in 32 Gb/sec or 64 Gb/sec Fibre Channel or FC-NVM-Express ports, 40 Gb/sec/100 Gb/sec or 10 Gb/sec/25 Gb/sec iSCSI or TCP over NVM-Express ports, or 12 Gb/sec SAS ports.

The 1U form factor for the FlashSystem 5600 has room to house a dozen of IBM’s custom E3 EDSFF NVM-Express flash drives, which come in 6.6 TB, 13.2 TB, 26.4 TB, and 52.8 TB capacities. In addition to the flash drives, the FlashSystem 5600 can have up to two expansion modules loaded up with disk drives – yes, disk drives – that spin at 7,200 RPM and that come in 12 TB, 20 TB, or 24 TB capacities. The feature 4662-12G expansion enclosure has room for 12 drives while the feature 4662-92G expansion enclosure has room for 92 drives. You can mix the types. It is not clear how the disk storage is presented to systems, but presumably it is all seamless and integrated. IBM uses its Storage Virtualize (formerly known as the Storage Virtualization Controller) to provide thin provisioning, snapshotting, and replication for data across the FlashSystems. The Storage Insights dashboard employs AI to proactively monitor the performance and capacity on the arrays, and also interfaces to the anomaly and ransomware detection hardware that is in the FlashCore Modules.

The FlashSystem 7600 comes in a 2U form factor and can support up to 1.68 PB of raw flash capacity in that relatively small space with up to 32 of the custom E3 EDSFF drives that comprise the FlashCore Modules. The same 6.6 TB, 13.2 TB, 26.4 TB, and 52.8 TB capacities are supported as with the FlashSystem 5600. IBM has a technique called Variable Stripe RAID to do data protection across the flash in each module, and overprovisions the drives more than other vendors do to be able to provide higher drive writes per day. With the compression, de-duplication, and thin provisioning all added up, IBM can make that 1.68 PB of raw flash capacity in the FlashSystem 7600 look and feel like 7.2 PB.

Each node canister in the FlashSystem 7600 has a 16-core X86 processor and 384 GB or 768 GB DRAM cache options. Each canister has four I/O slots instead of the two used in the FlashSystem 5600 and offers the same I/O choices: 32 Gb/sec or 64 Gb/sec Fibre Channel or FC-NVM-Express ports, 40 Gb/sec/100 Gb/sec or 10 Gb/sec/25 Gb/sec iSCSI or TCP over NVM-Express ports. The FlashSystem 7600 does not support extensions to SAS disk drive enclosures.

With the FlashSystem 9600, the node canisters have 48-core processors and 768 GB or 1.5 TB DRAM cache options, plus the same I/O options and 32 FlashCore Modules like the FlashSystem 7600. The one big difference, aside from the array controller compute boost is that there is a 105.6 TB FlashCore Module that allows the effective maximum capacity in the 2U enclosure to be 11.8 PB.

Up to 32 of these three FlashSystem devices can be gridded up to act like one big virtual array, which is neat, with maximums of 77 PB for the FlashSystem 5600, 230 PB for the FlashSystem 7600, and 377 PB for the FlashSystem 9600.

These capacities all sound like sci-fi to us. We live in the future.

Here is a table that shows the basic feeds and speeds of the three arrays and how they compare to prior midrange and entry FlashSystem devices:

That aggregate maximum bandwidth for reads out of the new midrange and high-end FlashSystems – 1,760 GB/sec for the FlashSystem 7600 and 2,752 GB/sec for the FlashSystem 9600 – couple to a response time on I/O operations of under 50 microseconds – a 28.6 reduction compared to the fourth generation FlashSystems – are what caught my eye.

IBM has not released pricing on these three new flash arrays, and we do not expect the kind of aggressive prices we saw with the C200 last year, which was definitely aimed at making flash as cheap as nearline disk storage. In the current market, IBM seems to be positioning the bigger FlashSystems for capacity and performance in a market where both are going to be in short supply. And that is not a recipe for inexpensive flash – not one tiny little bit.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search