Intel overhyping flash-killer XPoint? Shocked, we're totally shocked

Storage tech's real performance figures fail to match marketing claims

Swedish eggs and meat dish - pyttipanna
Is the XPoint dish over-egged?

+Comment XPoint will substantially undershoot the 1,000-times-faster and 1,000-times-longer-lived-than-flash claims made by Intel when it was first announced – with just a 10-times speed boost and 2.5-times longer endurance in reality.

This became clear when Micron presented details of its QuantX-branded XPoint memory technology at the Santa Clara Flash Memory Summit this week. Summit attendees could see a 128GB prototype QuantX system with four XPoint, seemingly, dual-chip packages on it; meaning two 16GB die per package.

In contrast to the wildly optimistic Intel "1,000 times faster than NAND" claims when XPoint was launched, the Micron pitch presented a 10-times improvement over NAND in terms of IOPS and latency, and four times more memory footprint than DRAM per CPU.

Now that does seem more realistic. A Micron blog states:

It delivers mesmerizing PCIe SSD performance – measuring typical read latencies at less than 10µs and a five 9 QoS at less than 25µs. That's over 10X better than the best NAND flash-based SSDs!

El Reg doesn't think a 10-times improvement over NAND is worth an exclamation point, or classing it as "mesmerizing" – not when a 1,000-times improvement has been alleged before.

We understand Micron is arguing that raw XPoint cell performance is 1,000 times faster than NAND raw cell performance, but, when you add in everything else in an SSD, such as the controller, its firmware and the interface, the speed drops.

What did Micron tell the audience about QuantX latency?

  • Write latency is <20μs (An NVMe SSD has <200μs).
  • Read latency is <10μs (An NVMe SSD has <100μs).
  • A U.2 200GB to 1.6TB XPoint SSD achieves 900,000 IOPS vs a 1.6TB NVMe SSD's 200,000 IOPS.
  • A 900GB to 1.6TB XPoint SSD in a half height, half length PCIe 3.0 x8 card achieves 1.8 million IOPS vs a 1.6TB NVMe SSD's 200,000 IOPS.

The IOPS testing used 4KB blocks with random 70 per cent read and 30 per cent write operations. With the U.2 QuantX NVME SSDs, queue depths of four to six saturated a PCIe x4 link, which the NAND NVMe SSD couldn't do. On a PCIe card, the queue depth reached 8 to 12, varying with QuantX SSD capacity, before the link was saturated. A 200GB QuantX SSD saturated the link at a queue depth of 16.

Micron also provided an endurance number: QuantX SSDs will support 25 drive writes per day (DWPD) over five years. What do NAND SSDs provide?

  • Samsung SM1715: 10 DWPD for five years
  • Toshiba PX04P: 10 DWPD for five years
  • Intel DC P3700: 17 DWPD for five years

It looks as if QuantX XPoint has 2.5 times more endurance than ordinary NAND SSDs, and not the 1,000 times (which would mean 10,000 DWPD for five years) that the Intel marketing department keeps banging on about.

We understand data stored in QuantX DIMMs will be encrypted. Operating system and application code changes will be needed to get the full benefit of XPoint as system memory, as opposed to storage. The NVDIMM-P products from Netlist will, presumably, help here, alerting potential XPoint memory ecosystem partners to the need for change.

Price

QuantX products will cost four to five times more than equivalent flash capacity and about half the cost of DRAM. We could see XPoint being used to bulk out DRAM as storage-class memory, and so enable larger working sets to be held in memory, leading to faster application run times.

Products with capacities from 200GB to 1.6TB are expected and these will be for data centers and enterprises. We can expect Micron to be trying to sign up server and storage system vendors as OEMs, with array vendors needing an NVMe-over-fabrics link to server hosts to make XPoint SSD use worthwhile.

Expect the first QuantX products in the second 2017 quarter. We would expect all the server vendors: Cisco, Dell, Fujitsu, HDS, HPE, Huawei, Inspur, Supermicro, etc to look at qualifying them. We would also expect all the NVMe over Fabrics-class array vendors: Apeiron, Dell/EMC DSSD, E8, Excelero, Mangstor and Pavilion Data Systems, to check them out as well.

Reg comment

The big question is: does XPoint deliver enough of a performance boost over NAND to justify a four to five-times price premium?

There are other technologies available to fill the price-performance gap between DRAM and NAND. Flash DIMMs in NVDIMM-P guise for one. Everspin MRAM DIMMs for another. And then there's Samsung's mystery Z-SSD.

XPoint competition is mounting up and it has a relatively trivial endurance advantage over flash, and a less-than-expected performance advantage. It is remotely possible that Intel and Micron have over-egged their XPoint pudding and it could fall flat. ®

PS: QuantX is supposed to signify quantum leap, by the way.


Biting the hand that feeds IT © 1998–2017