dram caches haven't ever appear to be size correlated. They tend to increment up on a generational basis; but have always been small enough I've assumed it came down to whatever the smallest chip size on the market was.
SSDs need a lot of dram for their capacity to hold mapping tables, HDDs just use it to buffer unwritten data to optimize write patterns. You don't want a really large one there because it increases the data loss risk from a power failure.
Micron lists the following sizes (converted from Mb to MB. Chips are sold by the bit): DDR2: 64MB-512MB (note that plenty of consumer devices still use 64MB. Also all chips are available in 4x, 8x, and 16x wide). DDR3: 128MB-2GB DDR4: 512MB-2GB
DDR4 has a significant price premium over DDR3 (for now). If it takes a new controller, they may have switched to DDR3 along the way, it is impossible to tell.
Remember, this is more a read-ahead buffer than anything else. Hard drives work better the longer the read length is, and the point of diminishing returns is somewhere around 64-256k. Playing with the numbers a bit, I'm guessing the thing reads whole tracks at a time (in the MB range) until given a different sector to read (in which it might only read ahead 64k).
Hybrid drives used to be a thing (I think I saw one on sale recently, so they aren't quite dead), and had 8GB of flash NAND storage. This is purely for the consumer market (and then only for the laptop market, mainly for those with only one drive bay and no m.2 slot), I'm sure that anyone in the enterprise market wanting a "real cache" for these will build a cache out of SSDs.
The cache is just buffer, so the amount needed has more to do with how active the user (or rather, workload) is rather than how large the drive is. I would hope that 256MB would be plenty to keep up with most workloads out there lol. Large drives like this will be mostly used for archive purposes rather than 'hot media', so this is probably an overkill amount.
Man... I have ~14TB in my home server with all 3TB HDDs in a RAID 6... to think I could replace it with 2 drives in a RAID 1 now hahaha
Cache should scale with platter size, not drive size IIUC. There are 3TB Seagates with 256MB cache, but they also use the same platter size (but fewer of them).
Saying a drive is PMR is extremely unnecessary and uninformative. Every HDD made in the last decade uses PMR. You may as well mention that the drive has a spindle and read/write heads.
The reason why it's important to note this is a PMR drive is because other manufacturers with this level of capacity typically employ SMR, which comes with a set of drawbacks that many consider unacceptable. This drive being PMR should lead to more consistent read/write speeds, possibly reliability improvements, etc.
Would you prefer if the article referenced the drives as being "non-SMR" instead? Stating that they're PMR is cleaner, and easily understood by the type of people interested in this drive.
I am happy to see a writer use "SMR" to distinguish them from other PMR-based drives because of the performance decrease of SMR drives. I personally would never buy a SMR drive, even for my cold/near-line storage.
Insisting they should all be written as "PMR" is just being pedantic.
I think I'm with you, Takeshi7. Like the others, I mostly want to know if a drive is shingled, so I can avoid it. If the correct term for "non-shingled" is CMR then that should be used. If PMR is ambiguous then it should be avoided. I certainly don't want to buy a drive described (correctly) as PMR only to find that it's shingled. The fact that we've muddled things up to now doesn't mean we should continue to do so. Especially on a tech website.
I hope SSHDs make a come back when QLC NAND comes out. The largest SSHD is just 2TB I believe. A 14TB SSHD with say 1GB of QLC NAND I think would work well as a combination. I'm not saying this because TLC AND QLC NAND SSDs wouldn't be able to meet someones need for bulk storage. But one common dismissal of SSHDs was always that the NAND part was very tiny. If we're going to make 3.5" HDDs, I can see QLC taking a more active role in them to really make them shine further.
There is simply no benefit. Cache is created to optimize data transfer from the system to media. The size increase of modern HDDs is reached by adding multiple platters. Still reading / writing operations are done to one platter / one side of it at a time. This is why sequential transfer rate is not changing with the size increases within one HDD generation. Caches are adjusted only when platter density allows faster speed. Otherwise you will just put lots of data into a cache that will be waiting for being written to media and take a risk that it disappears due to power loss.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
22 Comments
Back to Article
Marlin1975 - Friday, July 13, 2018 - link
Isn't the Cache on the small size for that large a drive? There are 3tb drives with 256mb of cache.DanNeely - Friday, July 13, 2018 - link
dram caches haven't ever appear to be size correlated. They tend to increment up on a generational basis; but have always been small enough I've assumed it came down to whatever the smallest chip size on the market was.SSDs need a lot of dram for their capacity to hold mapping tables, HDDs just use it to buffer unwritten data to optimize write patterns. You don't want a really large one there because it increases the data loss risk from a power failure.
wumpus - Thursday, July 19, 2018 - link
Micron lists the following sizes (converted from Mb to MB. Chips are sold by the bit):DDR2: 64MB-512MB (note that plenty of consumer devices still use 64MB. Also all chips are available in 4x, 8x, and 16x wide).
DDR3: 128MB-2GB
DDR4: 512MB-2GB
DDR4 has a significant price premium over DDR3 (for now). If it takes a new controller, they may have switched to DDR3 along the way, it is impossible to tell.
Remember, this is more a read-ahead buffer than anything else. Hard drives work better the longer the read length is, and the point of diminishing returns is somewhere around 64-256k. Playing with the numbers a bit, I'm guessing the thing reads whole tracks at a time (in the MB range) until given a different sector to read (in which it might only read ahead 64k).
Hybrid drives used to be a thing (I think I saw one on sale recently, so they aren't quite dead), and had 8GB of flash NAND storage. This is purely for the consumer market (and then only for the laptop market, mainly for those with only one drive bay and no m.2 slot), I'm sure that anyone in the enterprise market wanting a "real cache" for these will build a cache out of SSDs.
CaedenV - Friday, July 13, 2018 - link
The cache is just buffer, so the amount needed has more to do with how active the user (or rather, workload) is rather than how large the drive is. I would hope that 256MB would be plenty to keep up with most workloads out there lol.Large drives like this will be mostly used for archive purposes rather than 'hot media', so this is probably an overkill amount.
Man... I have ~14TB in my home server with all 3TB HDDs in a RAID 6... to think I could replace it with 2 drives in a RAID 1 now hahaha
saratoga4 - Friday, July 13, 2018 - link
Cache should scale with platter size, not drive size IIUC. There are 3TB Seagates with 256MB cache, but they also use the same platter size (but fewer of them).takeshi7 - Friday, July 13, 2018 - link
Saying a drive is PMR is extremely unnecessary and uninformative. Every HDD made in the last decade uses PMR. You may as well mention that the drive has a spindle and read/write heads.c4v3man - Friday, July 13, 2018 - link
The reason why it's important to note this is a PMR drive is because other manufacturers with this level of capacity typically employ SMR, which comes with a set of drawbacks that many consider unacceptable. This drive being PMR should lead to more consistent read/write speeds, possibly reliability improvements, etc.Would you prefer if the article referenced the drives as being "non-SMR" instead? Stating that they're PMR is cleaner, and easily understood by the type of people interested in this drive.
takeshi7 - Friday, July 13, 2018 - link
The term for a non-SMR drive is CMR, or conventional magnetic recording.SMR drives are still PMR, that's why PMR is an irrelevant term. The tech press really needs to learn that.
romrunning - Friday, July 13, 2018 - link
I am happy to see a writer use "SMR" to distinguish them from other PMR-based drives because of the performance decrease of SMR drives. I personally would never buy a SMR drive, even for my cold/near-line storage.Insisting they should all be written as "PMR" is just being pedantic.
Lolimaster - Friday, July 13, 2018 - link
Maybe you need to learn than being simple sometimes is best than trying to be overly formal.PMR simply stick after the change to PMR recording then with new tile based arranged SMR for the sake being easy to get it.
Arbie - Sunday, July 15, 2018 - link
I think I'm with you, Takeshi7. Like the others, I mostly want to know if a drive is shingled, so I can avoid it. If the correct term for "non-shingled" is CMR then that should be used. If PMR is ambiguous then it should be avoided. I certainly don't want to buy a drive described (correctly) as PMR only to find that it's shingled. The fact that we've muddled things up to now doesn't mean we should continue to do so. Especially on a tech website.nismotigerwvu - Friday, July 13, 2018 - link
I think it's mentioned to emphasize that this is a conventional drive, albeit He filled, rather than some exotic technology (HAMR, SMR, MAMR...ect).nandnandnand - Friday, July 13, 2018 - link
Do you make this complaint on every HDD article AnandTech runs? Because I'm getting deja vu. BTW, STFU.takeshi7 - Friday, July 13, 2018 - link
Yes. Yes I do. And I'll continue making the complaint until AnandTech stops being wrong.returnzer0 - Friday, July 13, 2018 - link
They're not wrong. By your own statements, you just think it's unnecessary.LordSojar - Friday, July 13, 2018 - link
This drive doesn't have a spindle. It's new Spindleless Technology, brought to you by IMPOSSIBRU Tech. I hear it works well, just like warp drive.CheapSushi - Friday, July 13, 2018 - link
I hope SSHDs make a come back when QLC NAND comes out. The largest SSHD is just 2TB I believe. A 14TB SSHD with say 1GB of QLC NAND I think would work well as a combination. I'm not saying this because TLC AND QLC NAND SSDs wouldn't be able to meet someones need for bulk storage. But one common dismissal of SSHDs was always that the NAND part was very tiny. If we're going to make 3.5" HDDs, I can see QLC taking a more active role in them to really make them shine further.PeachNCream - Friday, July 13, 2018 - link
Probably a bad idea given the low number of P/E cycles of current QLC and the high number of write operations a disk cache is likely to absorb.Lolimaster - Friday, July 13, 2018 - link
Dunno why it's so difficult to include more dram cache, at least 512MB/1GB for drives 4TB and up.adr0106 - Sunday, July 15, 2018 - link
There is simply no benefit. Cache is created to optimize data transfer from the system to media. The size increase of modern HDDs is reached by adding multiple platters. Still reading / writing operations are done to one platter / one side of it at a time. This is why sequential transfer rate is not changing with the size increases within one HDD generation.Caches are adjusted only when platter density allows faster speed. Otherwise you will just put lots of data into a cache that will be waiting for being written to media and take a risk that it disappears due to power loss.