7 ms·
Performance figures given in the article are about same as those for Samsung 960 Pro NVMe [1]. Would have expected this to have an opportunity for higher speeds
by jpalomaki 9y ago
Performance figures given in the article are about same as those for Samsung 960 Pro NVMe [1]. Would have expected this to have an opportunity for higher speeds by splitting the write/read to multiple chips.
40GB of DDR is pretty interesting. Add there a battery or suitable capacitor and you could use that as a reliable write cache. With SSDs I guess you could design it so that the capacitor would have enough power for flushing the data from the cache to the flash chips even if external power is lost.
[1] http://www.samsung.com/semiconductor/minisite/ssd/product/consumer/ssd960/ http://www.samsung.com/semiconductor/minisite/ssd/product/co...
- hateful 9y agoI would guess that maybe there's an overhead by adding additional NAND Channels? I wonder at what point the performance suffers because you have to split everything into too many channels?
- wtallis 9y agoThe NAND channel count within SSDs is usually limited by a combination of the host interface bottleneck (no point in using 64 channels if 12 is sufficient to saturate the uplink) and by the cost of increasing pin count on the controller (a current 32-channel controller is a 1517 ball BGA). Increasing the channel count also adds to the power draw, and most products try to stay within 25W maximum.
- wtallis 9y ago> Would have expected this to have an opportunity for higher speeds by splitting the write/read to multiple chips. This drive uses a 12Gb/s SAS interface rather than a ~32Gb/s PCIe x4 interface. SAS drives are still preferred in many segments of the enterprise storage market, especially where capacity is a higher priority than performance. > 40GB of DDR is pretty interesting. Add there a battery or suitable capacitor and you could use that as a reliable write cache. Supercapacitors are standard issue in enterprise SSDs like this. Most of the DRAM is used to cache the mappings between logical block addresses and physical NAND flash pages and blocks. With 40GB of DRAM but only 32TB nominal raw capacity of NAND flash on the drive, there's quite a bit of leftover DRAM for write caching. (The normal ratio is 1GB of DRAM per 1TB of NAND.)
- IntronExon 9y agoJesus, I remember being blown away by 16MB of HDD storage years ago. I am older than dirt! Now it’s 30TB, and solid state... no flying cars, but the future has definitely come through in some regards. Edit: These responses have filled me with a really pleasant nostalgia. It’s a pleasure to have such a simple, yet profound shared experience; the “We’ll never fill that up!...” club. Drinks are on me.
- op00to 9y agoI remember my father installing a 250 MB "hardcard" into our computer at home, and marveling at how we'd NEVER fill it up.
- jsgo 9y agosimilar when I was younger, but it was one of the smaller gig hard drives where I was like "wow, they've hit the point I'll never be able to fill this thing." Those were innocent times.
- berbec 9y agoI've said "Wow, I'll never fill THAT drive" many times in my life, and wouldn't you know it, I'm always wrong. I stopped saying that years ago; guess I can learn.
- noja 9y agoIs the power draw higher when splitting to more chips?
- jagger27 9y ago40GB seems like a lot of cached writes to lose in a power loss. Assuming best-case write speeds of ~1,700MB/s, you'd need almost 30 seconds of power. Could an on-board capacitor satisfy that or is a RAID-style battery necessary? Though I suppose Samsung expects these drives to be put behind a RAID controller (SCSI 12-Gbit, not PCIe), likely with battery backup.
- wtallis 9y agoIntel's 8TB drive takes about 5 seconds to shutdown and flush its caches, so your estimates are definitely in the right ballpark. I'm not sure how many more capacitors can be squeezed into a 2.5" drive. I'll be cracking open that 8TB drive once I'm done testing it—but carefully, and after making sure it has had plenty of time to discharge.
- kevin_thibedeau 9y agoA small LiPO cell would suffice.