2 ms·

The end of this article they claim their tech can do 4 IOPS/watt. Thats the theoretical power consumption, not including any computation to decode the data et

The end of this article they claim their tech can do 4 IOPS/watt. Thats the theoretical power consumption, not including any computation to decode the data etc. I assume that one 'IO' means reading a 47 kilobyte page. That means reading at say 200 Megabytes per second (like a hard drive) would take 1 Kilowatt! Am I missing something?

magicalhippo 2y ago
They also say a regular HDD has 20 IOPS/Watt. Consider the WD HC570[1] which can do around 220 IOPS in mixed 4KB workloads, and consumes around 9W doing so. So this seems what they mean, and thus they're "only" a factor of five or six away from the efficiency of a modern HDD. [1]: https://documents.westerndigital.com/content/dam/doc-library/en_us/assets/public/western-digital/product/data-center-drives/ultrastar-dc-hc500-series/data-sheet-ultrastar-dc-hc570.pdf https://documents.westerndigital.com/content/dam/doc-library...
- londons_explore 2y ago
  I think the difference is that a hard drive can read sequential pages far quicker, and at far less energy per byte, whereas as far as I can see this tech cannot. Random and sequential reads would work out the same energy per byte.
  - magicalhippo 2y ago
    Right, to put some numbers to that the same HC570 can do peak 291MB/s which divided by 4KB pages means 74496 sequential IOPS. Hence why we love SSDs. However to their defense, they seem to be quite focused on this being for cloud vendors. And, as has been demonstrated[1], a sufficiently large number of sequential IOs effectively looks like random IOs to the the server. That being said, surely this is more for cold data, so I'd imagine the cloud vendors can optimize access to benefit from the high sequential IO bandwidth of HDDs. [1]: https://youtu.be/yHgSU6iqrlE?t=2578 https://youtu.be/yHgSU6iqrlE?t=2578

2 ms·

The end of this article they claim their tech can do 4 IOPS/watt. Thats the theoretical power consumption, not including any computation to decode the data et

by londons_explore 2y ago

magicalhippo 2y ago
They also say a regular HDD has 20 IOPS/Watt. Consider the WD HC570[1] which can do around 220 IOPS in mixed 4KB workloads, and consumes around 9W doing so. So this seems what they mean, and thus they're "only" a factor of five or six away from the efficiency of a modern HDD. [1]: https://documents.westerndigital.com/content/dam/doc-library/en_us/assets/public/western-digital/product/data-center-drives/ultrastar-dc-hc500-series/data-sheet-ultrastar-dc-hc570.pdf https://documents.westerndigital.com/content/dam/doc-library...
- londons_explore 2y ago
  I think the difference is that a hard drive can read sequential pages far quicker, and at far less energy per byte, whereas as far as I can see this tech cannot. Random and sequential reads would work out the same energy per byte.
  - magicalhippo 2y ago
    Right, to put some numbers to that the same HC570 can do peak 291MB/s which divided by 4KB pages means 74496 sequential IOPS. Hence why we love SSDs. However to their defense, they seem to be quite focused on this being for cloud vendors. And, as has been demonstrated[1], a sufficiently large number of sequential IOs effectively looks like random IOs to the the server. That being said, surely this is more for cold data, so I'd imagine the cloud vendors can optimize access to benefit from the high sequential IO bandwidth of HDDs. [1]: https://youtu.be/yHgSU6iqrlE?t=2578 https://youtu.be/yHgSU6iqrlE?t=2578