Last year Marvell announced updated client NVMe controllers that we have not yet encountered in the retail SSD market, but now that the transition to PCIe gen4 is underway those controllers are already due for replacement. The new family of controllers reflect shifts in the market that Marvell is expecting, and are intended more for OEM SSDs than retail products. (Silicon Motion and Phison have almost completely displaced Marvell from the retail consumer SSD market.)

As NAND flash interface speeds and per-die capacities are increasing, Marvell is betting that mainstream client NVMe products can get away with just four NAND channels rather than eight. They're also making DRAMless SSDs (optionally with NVMe Host Memory Buffer support) a bigger part of their strategy. Those two changes combined means controllers can be physically much smaller, and Marvell expects shorter M.2 cards like the 22x30mm size to become much more popular now that they can offer higher performance and capacities up to 2TB (when using QLC NAND).

Marvell's new generation of client NVMe controllers consists of three products: DRAMless controllers with two or four lanes of PCIe gen4, and one controller with DRAM support and four lanes of PCIe gen4. All three controllers have four NAND channels, but the largest 88SS1321 that has the DRAM interface also has twice as many chip enables on the NAND channels and thus can support higher capacities than the DRAMless 88SS1322 and 88SS1323.

Marvell Client NVMe SSD Controller Comparsion
  88SS1321 88SS1322 88SS1323 88SS1084 88SS1100 88SS1093
 
Market Segment Consumer, Entry-level Datacenter Mainstream Consumer Mainstream Consumer High-end Consumer Consumer &
Enterprise
Manufacturing
Process
12nm FFC 28nm
CPU Cores 3x Cortex R5 4x Cortex R5 3x Cortex R5
DRAM DDR4, LPDDR4 No DDR4
LPDDR4
DDR4
LPDDR3
Host Interface PCIe 4.0 x4 PCIe 4.0 x2 PCIe 3.0 x4 PCIe 3.0 x4
NAND Interface 4 channels,
1200MT/s
4 channels,
800MT/s
8 channels, 800MT/s 8 channels, 533MT/s
Sequential Read 3.9 GB/s 3.9 GB/s 3.5 GB/s 3.0 GB/s 3.6 GB/s 3.2 GB/s
Sequential Write 3.3 GB/s 3.3 GB/s 3.0 GB/s 2.6 GB/s 3.0 GB/s 2.0 GB/s
4KB Random Read 690k IOPS 500k IOPS 450k IOPS 450k IOPS 780k IOPS 300k IOPS
4KB Random Write 500k IOPS 350k IOPS 300k IOPS 400k IOPS 650k IOPS 250k IOPS
Announced August 2019 June 2018 August 2014

The sequential IO performance of the new 4-channel controllers is only slightly better than Marvell's earlier 8-channel controller, and random IO has taken a step backward. Marvell isn't aiming to saturate a PCIe 4 x4 link, though the smallest 88SS1323 with only a PCIe 4 x2 link does hit the speeds we're used to seeing from PCIe 3 x4 SSDs.

Instead, Marvel is touting that they have the most power-efficient PCIe Gen4-capable SSD controllers, addressing concerns raised by AMD's latest chipsets and the Phison E16 SSD controller about PCIe 4 being a power hog. Marvell's new DRAMless controllers run at less than 2W with a PCIe 4 x4 link active, which isn't much more than the NAND flash itself requires. This is made possible by Marvell's jump to 12nm fabrication, compared to 28nm that has been the standard for most NVMe controllers. Even though these controllers are using a relatively advanced fab process, Marvell says they will allow for very cost-effective SSDs, especially when used in DRAMless configurations.

Aside from the faster PCIe and NAND interfaces, the new generation of controllers are architecturally similar to their predecessors, with a handful of Arm Cortex R5 CPU cores and the same fourth-generation LDPC engine used by last year's controllers from Marvell.

Marvell is currently sampling the new controllers, and will be showing them off next week at Flash Memory Summit.

 

 

Source: Marvell

Comments Locked

17 Comments

View All Comments

  • evilspoons - Thursday, August 1, 2019 - link

    Is that actually a 2230 M.2 card in the photo? I bought a Thinkpad as a work computer with three 2230 M.2 slots like five years ago expecting to be able to put SSDs in them but just ended up sticking with the 2.5" SATA drive the system came with until I left that job.
  • eye4bear - Friday, August 2, 2019 - link

    Having had almost 2.5 years of up and down problems with a Marvel chip in my Linksys WiFi router, I hate to see them anywhere near anything I might own in the future. I will not buy anything Linksys or with a Marvel chip in it ever again.
  • DyneCorp - Friday, August 2, 2019 - link

    Ah yes, the age old "my experience sucked, so wahh".

    People like this really are quite special.

    Could you please explain how you could better produce microprocessor chips?
  • Billy Tallis - Friday, August 2, 2019 - link

    Marvell is selling off that division.
  • peevee - Friday, August 2, 2019 - link

    It is so sad that x4 controllers are not simply twice as fast as x2. Product management fail. A couple more R5 cores would achieve that easily, and they are incredibly tiny on 12nm, package size would not increase.
  • DyneCorp - Friday, August 2, 2019 - link

    Why is it sad? What are you even on about?

    These are client drives. There's absolutely no reason to increase anything as no one who uses these SSDs will be utilizing the interface to its full potential.

    Do you guys just post this stuff to get attention or something? I don't understand. Did you even read the article or the title where it clearly states "client"?

    Do you even know what a client is?
  • Billy Tallis - Friday, August 2, 2019 - link

    The CPU cores usually aren't the bottleneck for SSD controllers. They're designed so that ordinary IO avoids the CPU cores as much as possible and flows through mostly fixed-function blocks.

Log in

Don't have an account? Sign up now