News Samsung seemingly caught swapping components in its 970 Evo Plus SSDs - Ars Technica

[Steltek]

The number of companies willing to screw over their customers for a few extra bucks just seemingly never ends...

Apparently, the controller substitution they slipped in without notice results in write speeds that drop as much as half over the original drives and is especially noticeable when writing large files.

Samsung seemingly caught swapping components in its 970 Evo Plus SSDs

The Phoenix controller in Samsung's own product images isn't present on newer drives.

arstechnica.com

 

[coercitiv]

Apparently, the controller substitution they slipped in without notice results in write speeds that drop as much as half over the original drives and is especially noticeable when writing large files.

I doubt it's the controller, since AFAIK it's the same controller used in 980 Pro.

So far the only problem I have with Samsung is they didn't add a "v2" on the packaging. Other than that it seems they made an effort to actually improve the drive by increasing the SLC cache.

 

[KentState]

I doubt it's the controller, since AFAIK it's the same controller used in 980 Pro.

So far the only problem I have with Samsung is they didn't add a "v2" on the packaging. Other than that it seems they made an effort to actually improve the drive by increasing the SLC cache.

To that point, it could possibly a firmware issue that can be resolved by an update. The Elpis is used on the 980 Pro so it doesn't look like a step back to a worse component. If anything, the controller is being more conservative with how much of the capacity it treats as SLC. Feels like Ars was trying to do a "gotcha" article without doing a thoroughly analysis. I may have missed where they reached out to Samsung for clarification.

 

[Steltek]

To that point, it could possibly a firmware issue that can be resolved by an update. The Elpis is used on the 980 Pro so it doesn't look like a step back to a worse component. If anything, the controller is being more conservative with how much of the capacity it treats as SLC. Feels like Ars was trying to do a "gotcha" article without doing a thoroughly analysis. I may have missed where they reached out to Samsung for clarification.

Samsung simply couldn't source the controller used in the 970 Evo Plus any more due to the semiconductor shortages, so they switched to one they could source. That is fine, can't fault them for that, as you can't sell what you can't produce.

However, in the end, Samsung knew that the controller wasn't a good match for the design of the 970 Evo Plus. There is no way they couldn't have not known, given that they re-designed it to implement the new controller, tested it, and manufactured it. Then, they turned around and intentionally sold and marketed it as the 970 Evo Plus without a version change or notice to the public. The only noticeable difference is a change to the manufacturer model number in tiny print on the drive itself which wasn't printed on the box.

I'm sorry, but what they did was certainly no better than what WD or even ADATA did. Maybe a firmware update will fix it, but maybe it won't. And, who knows how long it will take Samsung to get around to release this currently theoretical firmware fix. And, even if Samsung eventually fixes it, it doen't negate the fact that they purposely baited and switched customers into buying a product that wasn't what it represented itself to be solely to protect their sales revenue and market share.

 

[DeathReborn]

In my view Samsung, WD etc should make use of LE, SE or V2 or similar. Yes they'd have to either make new boxes etc or stickers & update some forms etc but at least that would show it's different to the reviewed model.

 

[Justinus]

Samsung simply couldn't source the controller used in the 970 Evo Plus any more due to the semiconductor shortages, so they switched to one they could source. That is fine, can't fault them for that, as you can't sell what you can't produce.

You do understand that Samsung designs and fabs all of these controllers in house, right?

They didn't suddenly run out of supply of the old 970 Evo Plus controller, they purposefully discontinued it and started using the newer, probably better in every measurable metric Elpis controller they debuted on the 980 "Pro".

Complain all you want, but the drive performs better in every conceivable metric except direct to TLC writes. They aren't using fewer NAND dies on the 970 Evo Plus vs the 980 Pro, so it's not a channel saturation issue. Likely they have a new firmware they put on the controller for this use case and it has some kinks that need to be worked out.

See the 980 "Pro" and it's write cache flushing bug that was only fixed about 6 months after launch.

I'd really like to see @Billy Tallis chime in after contacting Samsung for some actual journalism, unlike whatever the hell that Ars article is trying to be.

 

[Billy Tallis]

They aren't using fewer NAND dies on the 970 Evo Plus vs the 980 Pro, so it's not a channel saturation issue.

The updated 970 EVO Plus is using 512Gbit 128L dies, while the 980 PRO uses 256Gbit dies except on the 2TB model. So the 128L version of the 1TB 970 EVO Plus has half as many dies as the 92L version, or the 1TB 980 PRO. The 980 non-PRO also uses the 512Gbit 128L dies on at least the 500GB and 1TB models.

(I haven't actually been in touch with Samsung about this. I just read the part numbers off the NAND packages and looked them up in my spreadsheet. It's easy to miss the difference between K9DUGY8J5B and K90UGY8J5B if you don't know what to look for.)

 

[Insert_Nickname]

The updated 970 EVO Plus is using 512Gbit 128L dies, while the 980 PRO uses 256Gbit dies except on the 2TB model. So the 128L version of the 1TB 970 EVO Plus has half as many dies as the 92L version, or the 1TB 980 PRO. The 980 non-PRO also uses the 512Gbit 128L dies on at least the 500GB and 1TB models.

That makes sense then. Half the dies, half the direct-to-TLC performance. Better SLC cache performance due to the newer controller.

Just wish they'd added a v2 or similar.

Thanks for the info.

 

[Justinus]

The updated 970 EVO Plus is using 512Gbit 128L dies, while the 980 PRO uses 256Gbit dies except on the 2TB model. So the 128L version of the 1TB 970 EVO Plus has half as many dies as the 92L version, or the 1TB 980 PRO. The 980 non-PRO also uses the 512Gbit 128L dies on at least the 500GB and 1TB models.

(I haven't actually been in touch with Samsung about this. I just read the part numbers off the NAND packages and looked them up in my spreadsheet. It's easy to miss the difference between K9DUGY8J5B and K90UGY8J5B if you don't know what to look for.)

Do you have any speculation as to why the swap to Elpis nerfed the direct to TLC write speed?

 

[coercitiv]

Do you have any speculation as to why the swap to Elpis nerfed the direct to TLC write speed?

He just gave you the reason, the new drives are using half the number of dies.

 

[Justinus]

He just gave you the reason, the new drives are using half the number of dies.

With 512Gbit dies it would still need 16 dies for a 1TB drive, right? Isn't that plenty to saturate an 8 channel controller?

 

[Billy Tallis]

With 512Gbit dies it would still need 16 dies for a 1TB drive, right? Isn't that plenty to saturate an 8 channel controller?

For reads, yes. But on a single-die basis, writing to TLC NAND is typically 10-20x slower than reading, which is why SLC caching exists in the first place.

The industry is moving toward dividing 512Gbit dies into four planes that can operate more or less in parallel, rather than just two planes. This will allow a single 512Gbit die to sustain similar write speeds to two 256Gbit dies. But Samsung's 128L 512Gbit TLC is still just two planes per die, with a theoretical maximum write speed of 82MB/s (ignoring SLC caching): https://www.anandtech.com/show/16491/flash-memory-at-isscc-2021

So that gives a theoretical upper bound of 1312MB/s sustained write speed for the updated 1TB 970 EVO Plus, minus any write amplification due to garbage collection and SLC cache flushing.

 

[Justinus]

For reads, yes. But on a single-die basis, writing to TLC NAND is typically 10-20x slower than reading, which is why SLC caching exists in the first place.

The industry is moving toward dividing 512Gbit dies into four planes that can operate more or less in parallel, rather than just two planes. This will allow a single 512Gbit die to sustain similar write speeds to two 256Gbit dies. But Samsung's 128L 512Gbit TLC is still just two planes per die, with a theoretical maximum write speed of 82MB/s (ignoring SLC caching): https://www.anandtech.com/show/16491/flash-memory-at-isscc-2021

So that gives a theoretical upper bound of 1312MB/s sustained write speed for the updated 1TB 970 EVO Plus, minus any write amplification due to garbage collection and SLC cache flushing.

It seems like everyone's focusing on the controller swap and blaming it while simultaneously not realizing it's actually a NAND swap that accompanied it causing the direct to TLC write speed hit