Sabrent Plot Ripper SSD

[wONKEyeYEs]

Sabrent has released a new SSD line called Plit Ripper.

It has extended TBW for Chia farming use.

 

Quotes

Quote

Plotripper Series SSD's Endurance Is Enhanced To Another Level With LifeXtension,

allowing you to upshift your plotting to next gear. You do not have to do anything,

you can simply fire up your laptop and get to work.

 

Next it will be mining with RAM, it's the never ending story.

 

https://www.sabrent.com/plotripper/

[Chris Pratt]

Der8aurer did plotting with a RAMdisk. The problem is you need at least 128GB of RAM, and that only lets you do one plot at a time.

[Moonzy]

At least Chia is pushing for higher endurance SSD in consumer space, that's something great

 

GPU mining pushed for GPU without display ports, that's backwards evolution

[RejZoR]

27 minutes ago, Moonzy said:

At least Chia is pushing for higher endurance SSD in consumer space, that's something great

 

GPU mining pushed for GPU without display ports, that's backwards evolution

Not really. Despite this, once this "storage" mining stops, you'll have bunch of badly worn drives on sale as second hand products. Also you won't be able to buy one new when this crap goes full swing. For now Chia doesn't seem to be same rage as Ethereum or Bitcoin. And luckily upgrading storage isn't as necessary as graphic cards that get old much faster. I mean Is till have Samsung 850 Pro 2TB that I bought maaaaaany years ago (7 years I believe?) and it's serving me just the same as day 1. It's still as fast it's still huge by all metrics even today.

[leadeater]

12 minutes ago, RejZoR said:

Despite this, once this "storage" mining stops, you'll have bunch of badly worn drives on sale as second hand products

Yep, unlike GPUs that other than fans don't actually wear like many think SSDs do so Chia means you will have no resell value as it's very highly it'll die anyway before you sell it and anyone that cares to ask for a SMART readout would run for the hills once they see it.

[WolframaticAlpha]

1 hour ago, leadeater said:

Yep, unlike GPUs that other than fans don't actually wear like many think SSDs do so Chia means you will have no resell value as it's very highly it'll die anyway before you sell it and anyone that cares to ask for a SMART readout would run for the hills once they see it.

I doubt it if the market for used SSD's is that large. I wouldn't buy a used SSD. Used HDDs are still fine

[leadeater]

2 minutes ago, WolframaticAlpha said:

I doubt it if the market for used SSD's is that large. I wouldn't buy a used SSD. Used HDDs are still fine

Used SSD market is fine, or was. Most SSDs only ever get 10% or less wear on them.

[WolframaticAlpha]

1 minute ago, leadeater said:

Used SSD market is fine, or was. Most SSDs only ever get 10% or less wear on them.

Did not know that. But I would rather cheap out on the GPU/CPU tha cheaping out on the SSD. 

 

 

 

Used SSD market is gonna get rekt.

[porina]

2 hours ago, Chris Pratt said:

Der8aurer did plotting with a RAMdisk. The problem is you need at least 128GB of RAM, and that only lets you do one plot at a time.

128GB doesn't sound like enough. According to Chia site it uses 256GB of temporary storage while plotting, which would fit in 256GiB of ram with spare for OS (note units). 

 

2 hours ago, James Evens said:

Did they finally disclosed prices? Announcement  was like 1-2 weeks ago.

I was wondering pricing too. It would be no surprised if it cost proportionately more to achieve this.

 

2 hours ago, Moonzy said:

At least Chia is pushing for higher endurance SSD in consumer space, that's something great

Not if it costs much more, in which case you can achieve the same effect by over-provisioning anyway.

[Kisai]

1 hour ago, porina said:

128GB doesn't sound like enough. According to Chia site it uses 256GB of temporary storage while plotting, which would fit in 256GiB of ram with spare for OS (note units). 

 

256GB is not something you can get on DDR4 systems that aren't Xeon/EPYC server-class machines (eg Mac Pro.) So I think we might unintentionally see this with DDR5 systems, where if Chia takes off, people max out the RAM on a system just to do this.

 

However, keep in mind that Chia is not optimized for RAM, it's optimized for DISK, and likely there will be diminishing returns to trying to do plot farming, though at this point 256GB of RAM is cheaper than any GPU available. Just you can't put 256GB on an Intel non-Xeon system at present, and only AMD's ThreadRipper/EPYC can do it. So you're looking at investing between $6,000 an $20,000 to do this. Where as 256GB SSD's themselves are around $80. So you could probably just demolish 200 SSD's for the same cost as buying a system to do plotting in RAM to save demolishing any.

 

 

[RollinLower]

31 minutes ago, Kisai said:

256GB is not something you can get on DDR4 systems that aren't Xeon/EPYC server-class machines (eg Mac Pro.) So I think we might unintentionally see this with DDR5 systems, where if Chia takes off, people max out the RAM on a system just to do this.

 

However, keep in mind that Chia is not optimized for RAM, it's optimized for DISK, and likely there will be diminishing returns to trying to do plot farming, though at this point 256GB of RAM is cheaper than any GPU available. Just you can't put 256GB on an Intel non-Xeon system at present, and only AMD's ThreadRipper/EPYC can do it. So you're looking at investing between $6,000 an $20,000 to do this. Where as 256GB SSD's themselves are around $80. So you could probably just demolish 200 SSD's for the same cost as buying a system to do plotting in RAM to save demolishing any.

 

 

also, a big part of CHIA farming is plotting in parallel on the same system. there is no reason to just do 1 plot at a time. i do up to 40 plots per day on my 14 core Xeon system for example.

so eventough RAM would be mad fast for read/writes, the fact you can only fit 1 plot at a time on there means you would still be better off with a normal SSD setup.

[leadeater]

1 hour ago, Kisai said:

Just you can't put 256GB on an Intel non-Xeon system at present, and only AMD's ThreadRipper/EPYC can do it.

X299 can as well. X99 probably can too using DDR4 DIMM sizes larger than what existed at the time.

 

[WolframaticAlpha]

2 hours ago, Kisai said:

256GB is not something you can get on DDR4 systems that aren't Xeon/EPYC server-class machines (eg Mac Pro.) So I think we might unintentionally see this with DDR5 systems, where if Chia takes off, people max out the RAM on a system just to do this.

 

However, keep in mind that Chia is not optimized for RAM, it's optimized for DISK, and likely there will be diminishing returns to trying to do plot farming, though at this point 256GB of RAM is cheaper than any GPU available. Just you can't put 256GB on an Intel non-Xeon system at present, and only AMD's ThreadRipper/EPYC can do it. So you're looking at investing between $6,000 an $20,000 to do this. Where as 256GB SSD's themselves are around $80. So you could probably just demolish 200 SSD's for the same cost as buying a system to do plotting in RAM to save demolishing any.

 

 

putting in optane would be much more economical than using RAM

[RollinLower]

3 minutes ago, WolframaticAlpha said:

putting in optane would be much more economical than using RAM

for server systems optane is available in DIMM formfactor, you could drastically increase RAM size with that. tough very limited compatibility.

[porina]

2 hours ago, Kisai said:

256GB is not something you can get on DDR4 systems that aren't Xeon/EPYC server-class machines (eg Mac Pro.) So I think we might unintentionally see this with DDR5 systems, where if Chia takes off, people max out the RAM on a system just to do this.

Cascade Lake-X can, and it is pretty cheap on the used market since I assume previous owners of such have moved onto Threadripper.

 

2 hours ago, Kisai said:

However, keep in mind that Chia is not optimized for RAM, it's optimized for DISK, and likely there will be diminishing returns to trying to do plot farming

I agree, but this is a "you could" not a "you should" moment.

 

2 hours ago, Kisai said:

Where as 256GB SSD's themselves are around $80. So you could probably just demolish 200 SSD's for the same cost as buying a system to do plotting in RAM to save demolishing any.

Assuming USD, are you specifically looking at higher end ones? 500GB-class SSDs here start around that much, including tax.

 

I haven't done a detailed cost analysis, but I have to wonder if just using a LOT of HDs individually in parallel would be economic for plotting. 

 

1 hour ago, leadeater said:

X299 can as well. X99 probably can too using DDR4 DIMM sizes larger than what existed at the time.

Only Cascade Lake-X can, Skylake-X can't. At least, according to ARK. I didn't check back older.

 

44 minutes ago, WolframaticAlpha said:

putting in optane would be much more economical than using RAM

The Optane SSDs are probably more accessible than the Optane as ram systems. I'm still using a 900p 280GB as the boot SSD in my desktop gaming system. It's rated for 5100 TB endurance. For context, a 2 TB 980 Pro is only rated for 1200 TB endurance. So that Optane drive has about 30x higher rated cycle life than a current high end consumer SSD.

 

I do wonder if Chia plotting leans more towards sequential or random accesses. That might indicate how much plotting you can get done at once on any particular storage device. 

[lostcattears]

Ahhh think of all the lost CHia coins lost in cyberspace due to astronomical amounts of failures of storage devices

 

eh they might be regulated to being one of the thousands of useless coins... There is no hype around it. so no growth.

[Kisai]

4 hours ago, porina said:

Cascade Lake-X can, and it is pretty cheap on the used market since I assume previous owners of such have moved onto Threadripper.

But you're throwing away the "energy efficient" argument if you start using 65w TDP+ CPU's, Sure yes you can use the 99 chipsets, they're just the xeon socket 2011 chipset except "consumer model"

 

4 hours ago, porina said:

I agree, but this is a "you could" not a "you should" moment.

 

Assuming USD, are you specifically looking at higher end ones? 500GB-class SSDs here start around that much, including tax.

 

I didn't look at anything larger, I just know that a cheap source of 256GB NVMe's are off-lease laptops, so I looked up the current retail price for the same/similar part. I don't know why anyone would want to use these specifically, but that set a baseline to do the math.

 

As far as cost-effectiveness goes, what I think Sabrent did with their drives is they just over-provisioned it in reverse. Where as a typical SSD might be like 10-20% over provisioned, the sabrent plot-ripper is probably 80% overprovisioned, and they just used bigger chips, or they tuned the caching so that the drive itself has as large write cache as possible on the profile used. Too early to really tell, but, but from experience, SSD's in data centers don't last very long in web servers because so many processes are log-heavy. TLC drives aren't going to last 3 years. Keep in mind that all SSD's scale in performance linearly with size. So if you make the drive big enough, it's a lot harder to destroy, since less of the drive is constantly "turn over" in writes.

 

Another thought that did come to mind later after I posted this, is perhaps a specially designed board of PCIe 4.0 x 4 lane (so something with 28 lanes would have 7 of them) might actually use a software-raid with hot-spares setup where the plots are kept on redundant drives so they don't go poof if the drive wears out before the plotting is finished. I can think of a few scenarios, but it still seems like the only way the "Chia is green" argument works is if there are no mechanical drives in the system that have to stay on. Like what if you had 3 SSD's, one being a hotspare, two on the active raid and just keep going until you destroy one of the SSD's, let the hotspare take over, and just keep cycling drives out until you fill up all the plots you want, e-waste recycle the plot SSD's.

 

Now, again, I do not advocate for destroying SSD's purposely for this. Ideally what will come along is a type of SSD that has RAM-like durability and performance with some level of power-loss prevention, which may very well just be RAM in an enclosure (eg similar to Hyperdrive5) with a LiPO battery. Which means 4 x 64GB ($500-600/ea, or about $2000-$2400 for the RAM alone) with present DDR4 modules. 

 

Hmm interesting. Apparently NVDIMM-P is a thing. I wonder who is developing those.

 

[porina]

4 hours ago, lostcattears said:

Ahhh think of all the lost CHia coins lost in cyberspace due to astronomical amounts of failures of storage devices

The plot files are kinda disposable. You're just after total capacity. If some die, no big deal, you can replace it. Backups would be inefficient unless you use something even cheaper than HD. Tape? If it takes longer to backup/restore than to create new plot files, it isn't economic.

 

Coins wont get lost unless you lose your private key, equivalently the random words use to generate it. Anyone serious about crypto will take multiple backups. Being a small amount of text, it is easily backed up.

 

2 hours ago, Kisai said:

But you're throwing away the "energy efficient" argument if you start using 65w TDP+ CPU's

We're in the scenario of large scale Chia farmers. The power of plotting machines is probably insignificant relative to the farming machines. What's the spinning power of a 3.5" drive these days? From my limited observation of running Chia the farming disks don't get a chance to spin down. A few W each adds up once you scale. If we assume the farmer is continuously adding capacity, you could work out a power to create per plot I suppose, and compare it to the power needed to farm the plots. That is beyond my capability. Bottom line is this is still far below GPUs.

 

2 hours ago, Kisai said:

As far as cost-effectiveness goes, what I think Sabrent did with their drives is they just over-provisioned it in reverse. Where as a typical SSD might be like 10-20% over provisioned, the sabrent plot-ripper is probably 80% overprovisioned, and they just used bigger chips, or they tuned the caching so that the drive itself has as large write cache as possible on the profile used.

I have a gap in my plotting experience: just how many simultaneous plots can SSD like bandwidth support? You can then factor that in to the usable space to determine how many simultaneous plots can be done. I too also thought the logical ways they can try to improve endurance is to either run in SLC mode or overprovision. Maybe a bit of both. SLC mode does take up a ton of space though. For example, say you have a 1TB TLC drive, you go down to 1/3 capacity working in SLC mode. That's then only enough for one plot.

 

If it is just overprovisioning, you can pretty much do the same by buying a bigger normal SSD. It'll come down to the price.

 

2 hours ago, Kisai said:

Another thought that did come to mind later after I posted this, is perhaps a specially designed board of PCIe 4.0 x 4 lane (so something with 28 lanes would have 7 of them) might actually use a software-raid with hot-spares setup where the plots are kept on redundant drives so they don't go poof if the drive wears out before the plotting is finished.

Over-complicated for the benefit. If a SSD dies during plotting, sure you lose whatever is in progress. Adding any form of redundancy is just going to add cost and/or slow it down, and reduce value. Using a SSD as backup is wasteful. Using a HD as backup will impact performance as it probably wont keep up. Besides, you probably want to use the HDs for farming anyway.

 

2 hours ago, Kisai said:

I can think of a few scenarios, but it still seems like the only way the "Chia is green" argument works is if there are no mechanical drives in the system that have to stay on.

I think Chia is being promoted as greener than GPU mining, but I don't think we'll escape using power as long as we have to do computation and network activity.

 

Since it is capacity based, HDs remain the obvious choice to prove capacity balanced with performance. Can't just turn them off. I assume tape will be too slow. Any other bulk storage methods I might have overlooked? Don't think optical has the capacity.  

 

2 hours ago, Kisai said:

Ideally what will come along is a type of SSD that has RAM-like durability and performance with some level of power-loss prevention, which may very well just be RAM in an enclosure (eg similar to Hyperdrive5) with a LiPO battery. Which means 4 x 64GB ($500-600/ea, or about $2000-$2400 for the RAM alone) with present DDR4 modules.

Optane gets closer to ram, at least it is far beyond what flash can do. As illustrated earlier my 900p has about 30x rated cycle life over a 980 Pro. Still not unlimited, but a significant improvement. As expensive Optane is, it is still far cheaper than ram per capacity. At the time I think I paid around £400 for the 900p 280 GB. Today I'd be looking at around £1200 for 256GB of ram (8x32GB).

[Fasterthannothing]

14 hours ago, leadeater said:

Used SSD market is fine, or was. Most SSDs only ever get 10% or less wear on them.

10% or less is bull crap I've killed tons of SSDs including Samsung drives over the years. Heck the Samsung drive was used for nothing but booting Linux. You use a drive long enough you will kill it. Second hand SSDs are stupid.

[Kisai]

3 hours ago, SlidewaysZ said:

10% or less is bull crap I've killed tons of SSDs including Samsung drives over the years. Heck the Samsung drive was used for nothing but booting Linux. You use a drive long enough you will kill it. Second hand SSDs are stupid.

The only second-hand SSD's you should ever consider are NVMe drives. Never SATA ones. SATA ones are only good for disposable purposes since they cap performance at the SATA interface. Consider SATA SSD's the equivalent of buying 4200RPM drives while NVMe drives are 10,000RPM.

 

Even then, it doesn't make much sense to buy previously used SSD's since there's no way to check wear before buying them. Like my z87 Motherboard supports NVMe, but the board itself doesn't have M2, so I had to buy a bracket for it and stick it in the x4 slot. Still, it gets 3GB/sec where the previously existing SATA SSD gets 530MB/sec. At any rate anything smaller than 256GB will have rather poor performance regardless of it's age or wear or interface. 

[leadeater]

11 hours ago, SlidewaysZ said:

10% or less is bull crap I've killed tons of SSDs including Samsung drives over the years. Heck the Samsung drive was used for nothing but booting Linux. You use a drive long enough you will kill it. Second hand SSDs are stupid.

I'm sorry but I simply do not believe you. First why did the SSD fail, highly unlikely NAND wear that is for sure. Second which models were they. I only buy Samsung SSDs, I have 10 and zero have failed. These are mostly 840 & 850 Pros with two 840 EVOs and the EVOs which are OS boot devices have next to zero wear on them.

 

Here are two Read Intensive 5 years old SSD used in a server for our backup application Index Caching

 

These SSDs have worse DWPD than either the Samsung EVOs or Pros.

 

7 hours ago, Kisai said:

The only second-hand SSD's you should ever consider are NVMe drives. Never SATA ones

Please no, NVMe SSDs for a really long while had worse DWPD rating than the actually decent SATA ones that were around for ages. SATA SSD performance is excellent, you'll almost never see a real world difference between NVMe and SATA on a desktop system for every day usage.

 

Spec sheet and benchmark throughput figures are literally garbage for any real comparisons of storage devices.

 

7 hours ago, Kisai said:

Even then, it doesn't make much sense to buy previously used SSD's since there's no way to check wear before buying them

Yes there is, every vendor has an app that can read the wear usage of their SSDs or a standard SMART reader then just math out the bytes written, job done. If they won't give it to you then don't buy it, if it's significantly different return for refund.

 

20 hours ago, porina said:

Only Cascade Lake-X can, Skylake-X can't. At least, according to ARK. I didn't check back older.

ARK only considers DIMMs at the time, if large ones come out which almost always work then ARK doesn't get updates much of the time to reflect that. Talking UDIMMs of course.

[porina]

3 hours ago, leadeater said:

ARK only considers DIMMs at the time, if large ones come out which almost always work then ARK doesn't get updates much of the time to reflect that. Talking UDIMMs of course.

I stand corrected. As an alternate data point I looked at the web page for my X299 TUF Mark 2, and yes it does claim 256 GB max mem support for Skylake-X and newer CPUs.

 

I'd still be cautious on a case by case basis, as I recall there were compatibility problems when 32 GB modules were first introduced. 

[Stahlmann]

On 6/6/2021 at 9:42 AM, WolframaticAlpha said:

Used SSD market is gonna get rekt.

Thing about used SSDs is that you have the SMART system that allows users to see the rest of the expected life span.

This allows potential buyers to check the condition of the drive before making the buying decision when testing it in person.

Afaik there is no easy way to change the SMART values.

 

I'd actually love it if this would be a feature on GPUs. GPUs don't have a "lifespan" afaik, but just seeing the manufacturing date and the total operating time would be at least an indicator of a GPU that has been used for mining. 

[StDragon]

On 6/5/2021 at 11:46 PM, wONKEyeYEs said:

 

 

Next it will be mining with RAM, it's the never ending story.

 

https://www.sabrent.com/plotripper/

That's the whole point with proof-of-work crypto currency. It's all about valuation predicated on scarcity.

 

In order to boost valuation, there must be a PC hardware arms-race, be with with compute cycles (GPU or CPU), storage, or RAM capacity.

[StDragon]

On 6/6/2021 at 2:42 AM, WolframaticAlpha said:

Used SSD market is gonna get rekt.

I don't want to derail the conversation, but this does bring up an interesting point. Crypto mining is brutal on hardware. I would think there's an opportunity to provide some HW validation testing certification, bound to the device's serial number, with a program to publish records publicly as to the health status. This of course would only be useful for the 2nd hand market.

 

So far all we've got is SMART status from SSD vendor diags via screenshots; and that's assuming the seller would go through such trouble (most wouldn't). And I'm not sure how feasible this would be for GPUs unless there's a way to test for VRAM and GPU calculation errors.