Samsung PCIe 4 ‘Never Die’ SSDs: Machine Learning, Built-i
Samsung has announced new PCIe 4.0 SSDs, with a trio of new, baked-in technologies that could represent a significant improvement for drive features and reliability. The company is debuting 19 models based on the PM1733 and PM1735, with transfer speeds of up to 8,000MB/s.
According to Samsung, both of these drive controllers include new fail-in-place (FIP) technology. The company claims this is a true milestone in storage technology “by ensuring that SSDs maintain normal operation even when errors occur at the chip level, enabling a never-dying SSD for the first time in the industry.” SSDs outfitted with FIP software can detect faulty chips before they fail, scan for any data in the damaged sectors, and relocate that data into other chips. Samsung cites the example of a 30.72TB SSD with 512 NAND chips detecting an error and moving data out of these sectors before problems can occur.
Finally, Samsung claims to have implemented machine-learning tools to predict cell characteristics and detect variation in its circuit patterns through big data analysis. According to the company, this improved level of error control and analysis is required by the increasing layers of 3D NAND (we’re up to over 100) and the move from three-bit to four-bit cells. The implication of this phrase is that these new PCIe 4.0 products are TLC and QLC-derived, not MLC. Other manufacturers have been bringing up QLC as well or are already shipping it to customers.
The two drive families in question are based on the PM1733 and PM1735 controllers and will be available in a variety of form factors and mounts. The U.2 drives are capable of 6.4GB/s and 3.8GB/s, while the HHHL card-type drives are capable of 8.0GB/s and 3.8GB/s. Exact specs were not provided for various drives beyond the chart shown above. Drive endurance is defined as being either 1 drive write per day (DWPD) or 3, for a five year period in both cases.
For now, PCIe 4.0 is a feature that’s only available on AMD systems as far as the traditional x86 market is concerned. These drives will still work on Intel systems at reduced speeds, of course, and none of the additional features that Samsung has added appear to be tied to PCIe 4.0, which means they’ll all work equally well on Intel servers. As you may have guessed from the size capacities and references to virtualization, these drives are explicitly intended for the enterprise and server markets — though we’re curious about how long it’ll be before the AI and possibly the never-dying capabilities start to show up in consumer products as well. If AI can be used to reduce variability and improve yield we’ll likely see it across both enterprise and consumer products given enough time. Other capabilities, like the ability to divide the drive into 64 virtualized sub-drives, are likely to stay in the enterprise segment.
I should note that we also have questions about ‘Never dying’ SSDs and would really like to see additional materials on how this technology works, when it works, and how much additional protection it practically offers. Unfortunately, Samsung doesn’t seem to have written any publicly available materials on the topic.
Continue reading
Nvidia Unveils ‘Grace’ Deep-Learning CPU for Supercomputing Applications
Nvidia is already capitalizing on its ARM acquisition with a massively powerful new CPU-plus-GPU combination that it claims will speed up the training of large machine-learning models by a factor of 10.
New AI Writes Computer Code: Still Not Skynet, But It’s Learning
The Singularity is now in private beta. But you still have to care about syntax errors.
Google to Make Chrome ‘More Helpful’ With New Machine Learning Additions
Google is looking to make notifications in Chrome less annoying, and it wants to anticipate a user's behavior as well.
Google’s AutoML Creates Machine Learning Models Without Programming Experience
The gist of Cloud AutoML is that almost anyone can bring a catalog of images, import tags for the images, and create a functional machine learning model based on that.