Micron® 9300 NVMe™ SSD

Pavilion Data Systems has a mantra, a basic set of directives:

  1. Disrupt the data storage
  2. Enable true OpenChoice of NVMe SSDs
  3. Deliver the industry’s first Hyperparallel Flash Array.

With several members of our team having deep experience defining and launching performance storage devices (including the first PCIe SSDs, the first NVMe SSDs and multiple follow-on generations of NVMe) of we have strong opinions about the new drives that we qualify for OpenChoice. Typically, we keep these opinions to ourselves.

Occasionally, a drive comes along that creates more water-cooler buzz than others.  At the moment, that drive is the Micron 9300.  In particular, our team is excited about the 9300’s:

  • Capacity
  • Consistent read & write bandwidth
  • Latency measurements


Packing 15.36TB of TLC NAND, the 9300 is the biggest NVMe drive we’ve tested. 72 of these into a 4U chassis gives us more than 1PB of usable capacity providing a significant driver for NVMe-oF and disaggregation;  separating storage from CPU to fine-tune storage utilization and improve node recovery.

First, the utilization of NVMe drives in scale-out clusters using Direct-Attached Storage is inconsistent at best.  Cluster A has drives that are 80% utilized, while Cluster B may have only 30% drive utilization.  However, with the DAS approach, you still buy more servers and drives for Cluster A.  These islands of storage become even more problematic when more storage capacity is required even if no more CPU horsepower is needed.  With Pavilion Data’s disaggregated NVMe-oF, you can create RAID-6 volumes to the exact size for each server and offer thin provisioning to assure maximum device utilization.

As you deploy larger drives as DAS, suddenly node recovery in the event of a drive failure wreaks havoc with application performance.  In many large DAS deployments, NVMe drive size is now capped at 2TB since the recovery of 2TB can take as long as 50 minutes and rebuilding sharded data across the cluster creates heavy network traffic that impacts cluster performance.

By default, the failure of several SSDs in a Pavilion system has zero impact on cluster operations.  With our SWARM Recovery feature, the rebuild time for a single drive average 5 minutes per terabyte.

SWARM Recovery is unattainable with JBOFs, DAS, or Software-Defined Storage.  With up to 20 storage controllers, we harness as much processing power as you define to rebuild a specific drive within a RAID group.  The controllers work in parallel to reconstruct the drive as a background task, meaning no node is off-line during a rebuild, DAS recovery-related network traffic is eliminated, and cluster high-availability is never compromised.

Consistent Read & Write Bandwidth

There’s a reason most All-Flash Arrays do not show performance specs.  Legacy dual-controller architectures can’t deliver the bandwidth of a large number of NVMe SSDs.  When you push, they will quote 4K Read IOPS.  However, Write performance is rarely referenced.  It’s also true that most NVMe SSDs have nearly a 2x difference between Read throughput and Write at 128K sequential throughput.

The Micron 9300 has 3.5GB/sec bandwidth for both Read and Write.  Not only are these exceptional numbers, but the balanced approach is also precisely what the “real world” needs.  A great example of this is in AI/ML training.  Traditional methods use a sequential process to load data to an SSD, then separately push that data to a GPU, then back again.  With a balanced approach to Read/Write, it is possible to reduce AI training times significantly.

As Pavilion Data continues to push the envelope of Hyper-Parallel Flash Array performance (currently 120GB/sec Read and 90GB/sec Write), the Micron 9300 may help us further increase Write performance for hungry ingest workloads.

Lowest Latency SSD We’ve Measured

The most buzz around our shop came from latency testing.  We view latency as the most critical metric in NVMe-oF.  A fabric is, well, a fabric and altering the speed of light is still out of the question.  In our efforts to keep driving down NVMe-oF latency (our current RDMA-based spec is 40 microseconds) any latency decreases are welcome.  When we ran the 9300 through our testing, we found that it had lower latency and greater capacity than any other NVMe SSD to date.

Why We’re Excited About the Micron 9300

Big, consistent, and fast.  That’s the Micron 9300 NVMe SSD.  Its combination of massive capacity balanced read/write throughput and incredibly low latency are ideal for OpenChoice, and our Hyperparallel Flash Array. This combination can manage your high-demand, real-time applications across a variety of your critical workloads.