An anonymous reader quotes a report from ExtremeTech, written by Joel Hruska: According to a report from Chinese tech site Expreview, the WD SN550 Blue — which is currently one of the best-reviewed budget SSDs on the market — has undergone a NAND lobotomy. While the new SSD variant performs on-par with the old drive that WD actually sampled for review, once you exhaust the SLC NAND cache, performance craters from 610MB/s to 390MB/s. The new drive offers just 64 percent of the performance of the old drive.
This is unacceptable. It is unethical for any company to sample and launch a product to strong reviews only to turn around and sell an inferior version of that hardware at a later date without changing the product SKU or telling customers that they’re buying garbage. I do not use the term “garbage” lightly, but let me be clear: If you silently change the hardware components you use in a way that makes your product lose performance, and you do not disclose that information prominently to the customer (ideally through a separate SKU), you are selling garbage. There’s nothing wrong with selling a slower SSD at a good price, and there’s nothing right about abusing the goodwill of reviewers and enthusiasts to kick bad hardware out the door.
As a reviewer of some twenty years, I do not care at all about the fact that SLC cache performance is identical. While I didn’t realize it at the time I wrote up the Crucial bait-and-switch on August 16, I’ve actually been affected by this problem personally. The 2TB Crucial SSD I purchased for my own video editing work is one of the bait-and-switched units, and it’s always had a massive performance problem — as soon as it empties the SLC cache, it falls to what I’d charitably call hard drive-level performance. Performance can drop as low as 60MB/s via USB3.2 (and ~150MB/s when directly connected via NVMe) and it stays there until the copy task is done. The video upscaling projects I work on regularly generate between 300-500GB of image data per episode, per encode. Achieving ideal results can require weaving the output of 3-5 models together. That means I generate up to 1.5TB of data to create a single episode. God help you if you need to copy that much information to or from one of these broken SSDs. It’s not literally as bad as a spinning disk from circa 2003, but it’s nowhere near acceptable performance.
Read more of this story at Slashdot.