Why do SandForce SSD controllers have a bad reputation?

[patrickjp93]

I was never off, as my statements were correct the whole time. Its you, that spread bullshit about pretty much everything disscussed in this thread

 

PS:

Still waiting on those quotes for 730

No, just no, and I provided the quotes in the original and following posts. You've been wrong about the 730, the endurance of TLC, and techreport's less-than-superb reputation when it comes to testing products.

[hojnikb]

No, just no, and I provided the quotes in the original and following posts. You've been wrong about the 730, the endurance of TLC, and techreport's less-than-superb reputation when it comes to testing products.

No you didn't provice any quotes, that specifically call the 730 a sandforce drive. Endurace wise i was correct the whole time, as everyone with common knowlege can confirm. And numerous tests of TLC.

Techreports test is just fine.

[hojnikb]

V-Nand isnn't accessed the same way traditional 2-D NAND is. I suspect it's actually SLC-like. I'm betting Samsung is aiming towards a 4TB SSD, which would finally make HDD makers move their asses.

 

Its not. At the very basic level it still works as a regular flash (still needs to have different states to store data) the only mayor difference is, that v-nand can have multiple layers (thats why its called 3d), while with regular nand, you need to stack dies on top of each other to get the "same" effect. And since v-nand can have upwards of 32 layers, it's more dense than 2d nand.

But as i've said, at a very basic level its still floating gate transistors, electrons trapped inside and differentiating differente voltage states for data. And with 850PRO, they are using 4 different states, so effectivly its MLC.

[hojnikb]

PS:

4tb SSD would really require updated controllers, 3D nand really isn't necessery for that.

2D nand at 1X nm is already dense enough for 4TB drives (you can stack 16 128Gbit dies and you need 16 of them -- easily doable in a 2.5" formfactor). Its the controllers, that keep us from having such large drives. Its pretty hard to keep track of so many blocks, thats why most solutions cap out at 1TB.

[patrickjp93]

No you didn't provice any quotes, that specifically call the 730 a sandforce drive. Endurace wise i was correct the whole time, as everyone with common knowlege can confirm. And numerous tests of TLC.

Techreports test is just fine.

No, no, and no! You lying piece of crap! TLC, at least made by Samsung, is only about 12-18% worse, not 70%, on endurance.

 

And Techreports is not only inaccurate, it's also outdated. software testing can't give you the truth. You need to work on the hardware. Reporting bad sectors is not indicative of anything. It reports a bad block once it's reached a pre-determined number of writes. It will continue writing to that block until the leakage current passes a certain amount and then it will finally be a "dead" block.

[patrickjp93]

Its not. At the very basic level it still works as a regular flash (still needs to have different states to store data) the only mayor difference is, that v-nand can have multiple layers (thats why its called 3d), while with regular nand, you need to stack dies on top of each other to get the "same" effect. And since v-nand can have upwards of 32 layers, it's more dense than 2d nand.

But as i've said, at a very basic level its still floating gate transistors, electrons trapped inside and differentiating differente voltage states for data. And with 850PRO, they are using 4 different states, so effectivly its MLC.

but that way in which data is accessed in VNAND is very different. It's much like capacitors in series which have less impedance as you increase the number of capacitors. You get a chain of data from one string of cells instead of being able to get 1 data item from 1 cell at a time.

[hojnikb]

No, no, and no! You lying piece of crap! TLC, at least made by Samsung, is only about 12-18% worse, not 70%, on endurance.

 

And Techreports is not only inaccurate, it's also outdated. software testing can't give you the truth. You need to work on the hardware. Reporting bad sectors is not indicative of anything. It reports a bad block once it's reached a pre-determined number of writes. It will continue writing to that block until the leakage current passes a certain amount and then it will finally be a "dead" block.

 

Thats not true at all. It will report a bad block, once it cant reliably store electrons inside or they become trapped due to wear. Its the pre determined p/e that caused lockup on some drives though in their testing. But thats a different thing entirely.

[patrickjp93]

PS:

4tb SSD would really require updated controllers, 3D nand really isn't necessery for that.

2D nand at 1X nm is already dense enough for 4TB drives (you can stack 16 128Gbit dies and you need 16 of them -- easily doable in a 2.5" formfactor). Its the controllers, that keep us from having such large drives. Its pretty hard to keep track of so many blocks, thats why most solutions cap out at 1TB.

Actually it's thermal performance that's been the problem with stacking 2D nand, and not the controllers. Any algorithm built for one size will scale if it was designed with a lick of sense. It's no different than stable matching or sorting. Both are trivial problems.

[hojnikb]

but that way in which data is accessed in VNAND is very different. It's much like capacitors in series which have less impedance as you increase the number of capacitors. You get a chain of data from one string of cells instead of being able to get 1 data item from 1 cell at a time.

Thats not how V-NAND works at all. Actually, thats not how NAND works at all.

[patrickjp93]

Thats not true at all. It will report a bad block, once it cant reliably store electrons inside or they become trapped due to wear. Its the pre determined p/e that caused lockup on some drives though in their testing. But thats a different thing entirely.

that is across all chips, not one block. Bad blocks are only blocks which have met pre-determined estimates. They're not dead/read-only until the leakage current reaches a certain point.

[hojnikb]

Actually it's thermal performance that's been the problem with stacking 2D nand, and not the controllers. Any algorithm built for one size will scale if it was designed with a lick of sense. It's no different than stable matching or sorting. Both are trivial problems.

 

Not true at all. And 1TB 840evo msata proves, that you can reliably stack 16 dies in one package. So considering one can put 1TB into 4 packages, you might aswell have 16 of them (easily fitterd inside 2.5" case). So really it comes down to controller, as i've said.

[patrickjp93]

Thats not how V-NAND works at all. Actually, thats not how NAND works at all.

That's not how 2D nand, but that is exactly how VNAND works. we met a few samsung engineers at Miami right before I departed for home. My explanation is theirs.

[hojnikb]

That's not how 2D nand, but that is exactly how VNAND works. we met a few samsung engineers at Miami right before I departed for home. My explanation is theirs.

Maybe u got the explaination mixed up or they dumbed down it so severly for you, that it makes no sense at all

[hojnikb]

that is across all chips, not one block. Bad blocks are only blocks which have met pre-determined estimates. They're not dead/read-only until the leakage current reaches a certain point.

Nope. If that was true, all of the drives in techreports article would die long ago.

[patrickjp93]

Not true at all. And 1TB 840evo proves, that you can reliably stack 16 dies in one package. So considering one can put 1TB into 4 packages, you might aswell have 16 of them (easily fitterd inside 2.5" case). So really it comes down to controller, as i've said.

No, and no. You're not asynchronously writing to all 16 layers at once, ever. If you did the latent heat would eventually compromise the cells at the halfway mark and above due to the large heat output combined with electrical flow. It's one thing to use a heat-based endurance correction system when cells are idle, but having that much electrical flow would inevitably melt some of the floating-gates without aggressive cooling.

[patrickjp93]

Nope. If that was true, all of the drives in techreports article would die long ago.

the drives don't die based on bad blocks. They die when finally all blocks cannot be reliably written to (measured by leakage current).

[patrickjp93]

Maybe u got the explaination mixed up or they dumbed down it so severly for you, that it makes no sense at all

No, that is their explanation verbatim, to a graduate-level class of computer engineering students with far more hands-on and theoretical experience with transistor tech than either of us. 

[hojnikb]

the drives don't die based on bad blocks. They die when finally all blocks cannot be reliably written to (measured by leakage current).

Yes they do. Once they burn thru all of the spare blocks (when you get a bad block, its replaced with a new block from spare area) drive will simply die.

[hojnikb]

No, and no. You're not asynchronously writing to all 16 layers at once, ever. If you did the latent heat would eventually compromise the cells at the halfway mark and above due to the large heat output combined with electrical flow. It's one thing to use a heat-based endurance correction system when cells are idle, but having that much electrical flow would inevitably melt some of the floating-gates without aggressive cooling.

I never said you could. But you can have 16 of them (as its evident from 84evo 1tb msata).

But since controllers can't keep up with so many blocks, they can't really make a consumer drive with more than 1TB of storage. You could raid more controllers on pcb (as it was done in the past) but a single consumer grade controller can't adress more than 1tb. Too many blocks to keep track, lots of dram requred for redirection tables etc, etc... page and block size need to go up, than we'll see bigger single controller solutions.

[patrickjp93]

Yes they do. Once they burn thru all of the spare blocks (when you get a bad block, its replaced with a new block from spare area) drive will simply die.

Nope, because a bad block is only estimated to be bad until considered dead. Bad is estimation (useful for wear-leveling) and dead is unwritable.

[patrickjp93]

I never said you could. But you can have 16 of them (as its evedent from 84evo 1tb msata).

But since controllers can't keep up with so many blocks, they can't really make a consumer drive with more than 1TB of storage. You could raid more controllers on pcb (as it was done in the past) but a single consumer grade controller can't adress more than 1tb

bullshit. It's a 64-bit arm chip slimmed down. It can address petabytes the same way any 64-bit architecture can.

[hojnikb]

bullshit. It's a 64-bit arm chip slimmed down. It can address petabytes.

It can't. Its something to adress a lot of ram (that is true and keep adress petabytes -- well way more by design) but entirely different to adress flash blocks.

 

And controller isn't 64bit, its really 32bit cortex r series

[hojnikb]

Nope, because a bad block is only estimated to be bad until considered dead. Bad is estimation (useful for wear-leveling) and dead is unwritable.

Thats not how bad blocks work.

[patrickjp93]

It can't. Its something to adress a lot of ram (that is true and keep adress petabytes -- well way more by design) but entirely different to adress flash blocks.

 

And controller isn't 64bit, its really 32bit cortex r series

You have a source for that? And if you do, maybe you could answer the seemingly obvious question: why the F@CK would anyone limit themselves to 32-bit architecture now by choice? If Samsung could build 4TB drives they would and they should. It would force every other SSD maker into deeper competitive waters and just plain make them look good.

[patrickjp93]

Thats not how bad blocks work.

that's, and I'm sorry, but you're wrong.