[SOLVED] Samsung 980 vs 970 EVO Plus M.2 SSDs

[jimlau]

Any sense which 1 of these is better? The 980 seems to have better stats except for the sustain write speed.

If I do more reading than writing, should I give more importance to the reading stats? That would suggest the 970 EVO Plus.

Thanks.

 

[Darkbreeze]

I'd just get the one that was less expensive, assuming both are new. Either of them is likely to already be fast enough in practically every metric to saturate the bus in benchmarks and in real world use, 99% of the time you'll be very unlikely to ever see those kinds of speeds anyhow since the majority of what we do is random operations anyhow. The number of times we are actually reading or writing extremely large sequential or single files isn't very often and that's where you'd see those big numbers. The rest of the time, any of those drives is more than fast enough for regular operations and having a faster drive isn't particularly likely to greatly change that. Fast is fast, it doesn't need to be overcomplicated.

 

[jimlau]

Any sense which 1 of these is better? The 980 seems to have better stats except for the sustain write speed, which is better.

If I do more reading than writing, should I give more importance to the reading stats? That would suggest the 970 EVA.

Thanks.

 

[jimlau]

I'd just get the one that was less expensive, assuming both are new. Either of them is likely to already be fast enough in practically every metric to saturate the bus in benchmarks and in real world use, 99% of the time you'll be very unlikely to ever see those kinds of speeds anyhow since the majority of what we do is random operations anyhow. The number of times we are actually reading or writing extremely large sequential or single files isn't very often and that's where you'd see those big numbers. The rest of the time, any of those drives is more than fast enough for regular operations and having a faster drive isn't particularly likely to greatly change that. Fast is fast, it doesn't need to be overcomplicated.

I do read several thousand music sample files when I use a digital workspace for music composing. Would you say I still wouldn't likely notice a difference?
Thanks.

 

[Lutfij]

They are essentially the same, pick the one that's cheaper in your region/store. The 980 doesn't come in a 2TB flavor, while the 970 Evo Plus does.

 

[USAFRet]

Any sense which 1 of these is better? The 980 seems to have better stats except for the sustain write speed, which is better.

If I do more reading than writing, should I give more importance to the reading stats? That would suggest the 970 EVA.

Thanks.

The 980 is just a continuation of the 970 line.
You'll see zero user facing difference.

The 980 Pro is PCIe 4.0, and requires a relevant motherboard.
Benchmarks faster. But again, depending on what you're doing, little if any actual difference.

 

[sonofjesse]

Both are good drives you would be happy with either.

 

[Darkbreeze]

I do read several thousand music sample files when I use a digital workspace for music composing. Would you say I still wouldn't likely notice a difference?
Thanks.

Yes, I would still say that you wouldn't notice much if any difference, even you HAD a PCIe 4.0 capable motherboard, other than maybe on benchmarks. If the price difference is less than say ten or fifteen bucks, then go with the newer drive. Otherwise, I don't really see that you stand to gain a lot by going one way or the other.

 

[sonofjesse]

Both are great drives, can't go wrong with either.

 

[anticeon]

970 Evo is far better ...
the answer is DRAM, their speed might be similar but durability might be different

 

[Darkbreeze]

970 Evo is far better ...
the answer is DRAM, their speed might be similar but durability might be different

Do you have some evidence that supports this statement?

 

[USAFRet]

Do you have some evidence that supports this statement?

The 980 (not Pro) is DRAM-less.
https://www.tomshardware.com/reviews/samsung-980-m2-nvme-ssd-review

But for "durability", it is no 'worse' than most other similar drives.

Samsung 980 PCIe 3.0 SSD | Samsung Semiconductor Global

Samsung 980 is the smart choice for a gaming booster or seamless workflow. Experience all the speed advantages of NVMe M.2 SSD. Available from 250GB to 1TB.

www.samsung.com

Warranty numbers:
250GB - 150TBW
500GB - 300TBW
1TB - 600TBW

Any of which is totally irrelevant for normal consumer use.
150TBW is well beyond what happens in actual consumer use in a decade.

 

[USAFRet]

970 Evo is far better ...
the answer is DRAM, their speed might be similar but durability might be different

Define "durability".
Actual numbers, please.

 

[anticeon]

Do you have some evidence that supports this statement?

its simply using Logic, HMB only using RAM as low as 128MB.
Meanwhile 1tb of NVME usually have 1gb of DRAM.

if we are in budget its better to buy NVME with LP/DDR3 DRAM which is have similar price as DRAMLESS NVME.

except samsung controller have some magic to dynamically allocated the RAM upto 1-2 GBs to replace the function of DRAM

 

[USAFRet]

if we are in budget its better to buy NVME with LP/DDR3 DRAM which is have similar price as DRAMLESS NVME.

In a blind test, I defy you to tell the difference between a Samsung 980 and a 970 EVO/EVO Plus.
Either in user facing performance (today), or the mythical "longevity" ( a decade from now).

 

[deesider]

its simply using Logic, HMB only using RAM as low as 128MB.
Meanwhile 1tb of NVME usually have 1gb of DRAM.

if we are in budget its better to buy NVME with LP/DDR3 DRAM which is have similar price as DRAMLESS NVME.

except samsung controller have some magic to dynamically allocated the RAM upto 1-2 GBs to replace the function of DRAM

What logic is this?
Using DRAM is generally faster because the drive can buffer data, but makes no difference to the number of writes to NAND. Thus there is no difference in durability.

 

[anticeon]

What logic is this?
Using DRAM is generally faster because the drive can buffer data, but makes no difference to the number of writes to NAND. Thus there is no difference in durability.

We can't calculate the real write to the NAND. there is an article that explain if we write a 10gb single file to SSD it will write more than a 10gb, but if we write a small 4kb files with total 10gb of size, the Write to the SSD might become twice or thrice of 10gb.
the caching is so complicated and different with each controller of SSD.
so more dram make the write less to the NAND cells.

 

[USAFRet]

We can't calculate the real write to the NAND. there is an article that explain if we write a 10gb single file to SSD it will write more than a 10gb, but if we write a small 4kb files with total 10gb of size, the Write to the SSD might become twice or thrice of 10gb.
the caching is so complicated and different with each controller of SSD.
so more dram make the write less to the NAND cells.

And in the context of normal consumer use, still irrelevant.

In all the SSD's you've owned and used, have you come anywhere close to the warranty TBW number?
If so, was that before or after the warranty expired due to age? 3-5 years?
Have you ever had an SSD die from too many write cycles?

 

[Darkbreeze]

And in the context of normal consumer use, still irrelevant.

And, besides which, the logic itself seems less than substantive.

 

[deesider]

We can't calculate the real write to the NAND. there is an article that explain if we write a 10gb single file to SSD it will write more than a 10gb, but if we write a small 4kb files with total 10gb of size, the Write to the SSD might become twice or thrice of 10gb.
the caching is so complicated and different with each controller of SSD.
so more dram make the write less to the NAND cells.

This table on Wikipedia lists the factors that affect write-amplification, and it does not include the use or quantity of DRAM as a factor: https://en.wikipedia.org/wiki/Write_amplification#Factors_affecting_the_value

Perhaps your source is outdated due to controller improvements.

Also, you can see that the endurance of the 980 (DRAMless) and the 980 Pro are the same. So would it not be logical to conclude that the presence of DRAM does not affect endurance?

 

[Maxxify]

DRAM absolutely impacts write amplification. Whenever you do writes on the drive you need to also update the mapping table, which temporarily is kept in DRAM, so you can both defer writes and combine writes to reduce WA, particularly with smaller I/O. The 980 can use host memory buffer (some system memory) for this, although it's very limited in capacity and has higher latency. DRAM is also particularly useful with maintenance and garbage collection as you track block-level metadata (e.g. wear-leveling) and especially if the drive is fuller where you have fewer prepared blocks; this also can reduce WA indirectly.

Also, TBW - "endurance" as you say for the 980 and 980 PRO - has nothing to do with real endurance. It's arbitrary on consumer drives and set only for warranty purposes, amount of writes or warranty period. DC/enterprise drives will factor TBW for the DWPD - drive writes per day - as they must target a specific amount of writes, although even this may not match real endurance but rather the guaranteed warrantied amount. Also, of course, if the drive is tracking by host writes it's even more meaningless.

In any case, a 980 PRO will be faster and more durable than a 980, at least thanks to the DRAM but also controller. However for consumer usage you probably wouldn't notice either.

 

[deesider]

DRAM absolutely impacts write amplification. Whenever you do writes on the drive you need to also update the mapping table, which temporarily is kept in DRAM, so you can both defer writes and combine writes to reduce WA, particularly with smaller I/O. The 980 can use host memory buffer (some system memory) for this, although it's very limited in capacity and has higher latency. DRAM is also particularly useful with maintenance and garbage collection as you track block-level metadata (e.g. wear-leveling) and especially if the drive is fuller where you have fewer prepared blocks; this also can reduce WA indirectly.

I don't disagree with the principles you have mentioned, but only the magnitude of their effect. Surely the mapping table itself can only have a miniscule effect on wear levels, and it is not as if a modern controller naively writes 4k blocks as they arrive. That sort of issue was resolved many years ago - including through the use of host memory.

Also, TBW - "endurance" as you say for the 980 and 980 PRO - has nothing to do with real endurance. It's arbitrary on consumer drives and set only for warranty purposes, amount of writes or warranty period. DC/enterprise drives will factor TBW for the DWPD - drive writes per day - as they must target a specific amount of writes, although even this may not match real endurance but rather the guaranteed warrantied amount. Also, of course, if the drive is tracking by host writes it's even more meaningless.

Maybe it would interesting to see some actual endurance testing on current drives - I have seen any comparisons since that TechReport 'experiment' 6 years ago. I assume all drives easily meet their official TBW limits, so perhaps testing beyond that is somewhat misguided, since as you point out the price paid is for the warranty period and no more.

 

[Maxxify]

If you're asking about the difference in write amplification, that is write amplification factor as the ratio between NAND and host writes, for SATA drives it would not be uncommon to see DRAM-less drives having twice the wear rate of DRAM drives, e.g. 3.0 vs. 1.5 WAF. It depends on the workload and drive characteristics, although these numbers are very low with the understanding that consumer workloads are read-heavy with lots of idle time, although random writes are not untypical for OS usage. I say drive characteristics because even something as "simple" as erase block size (e.g. larger with QLC) can impact WAF greatly, for example quadrupling pages per block could double or triple WAF with concern to garbage collection. I have documentation to back this up as I covered it on my discord server recently. As for NVMe drives, yes HMB is a good option, but it has far higher access latency than on-PCB DRAM (I also have a patent with SPECIFIC timings on this that I've shared on my discord) which can reduce efficiency particularly with a fuller drive. In the grand scheme of things I don't think halving endurance is even serious within the typical 3-year DRAM-less warranty period, though.

Actual flash endurance is based on many factors but if we're looking at the raw flash it largely comes down to architecture, assuming similar yield goals. As such we have a reasonable idea on guaranteed endurance as it's on the datasheet, e.g. Micron's 96L B27B is 3000 PEC, including in SLC mode, e.g. 30K/40K PEC depending. You'll note that Micron's 96L B27A is rated for a lower PEC than B27B. This is due to geometric differences, for example the block size is much smaller on B27B; reference my point above about this, keeping in mind the influence of DRAM on GC (since you are migrating lots of pages from multiple blocks). In any case, you would see 1500/3000 for BiCS for example, and Samsung's older 3D TLC is rated for 10K+ PEC (but it's difficult to find these datasheets), although actual endurance will be higher still. Generally I would expect 3K PEC from 3D TLC as a guarantee with 5-10K PEC being a realistic result and up to 20K PEC for the best replacement gate TLC, keeping in mind Micron's 128L TLC was rated for 5000 PEC.

That being said, again, PEC here would refer to NAND writes or indirectly the average block erase count (which factors into WA/WAF, but we're assuming ideal wear-leveling). However even this simple example is incomplete as you have good blocks and bad blocks due to the uneven etching aspect ratio (pillars are conical not cylindrical) plus other things like HCI (hence dummy WLs). That is to say we can dismiss host writes but cannot treat all word lines as equal, either. But further, consumer drives rely on SLC caching which impacts endurance as static SLC uses a separate wear zone (but the best blocks of each die) while dynamic uses the same as native flash and thus has additive wear potential (and lower relative endurance as it rotates through all native flash based on wear data). Thus testing endurance properly is complicated. But as you can see with the one Chia drive, 10K PEC with TLC and dynamic SLC caching is achievable (using FortisMax vs FortisFlash).