SSD capacity

[_rubik]

even smaller industries such as bicycle shops have to have their Christmas orders placed by June if they expect to have it in time

Don't get me started on the cycling industry manufacturing shortages. The local bike bubble here nearly rivaled the Pokemon craze over covid. If only I could sell my many-year-old heavily used HDD for 2x retail like I could a dirty road bike.

I haven't looked, but the industry was in freefall just before Christmas (which is really rare), sales tanked and a bunch of companies were on the edge of going under.
Not sure why but I guess they expected the party to keep going forever. I was extremely temped to upgrade but I just didn't want to saddle myself with payments for a year and work has me working long hours anyhow so..

Check Jensenusa, they always have killer deals.

That's the story of SO many quarantine friendly activities. "Hey, the line is going up, guess we better revise our 10 year outlook to include this pace". One of the many places cycling, tech, and cycling techies overlap

[Axiom_]

I'm praying that this AI craze will also drive storage tech since those training datasets are many many PB

Just speculating, the AI boom is more likely to drive up prices in the short term. And on the tech side of things, it seems like the real winner is Nvidia while storage tech is seemingly moving at its own pace, determined largely by cloud service demand.

[tp4tissue]

Just installed the 4tb sn850x, so roomy.

[tp4tissue]

Some testing.

No Heatsink 53C idle.. Heatsink 35C idle.

[phinix]

Some testing.

No Heatsink 53C idle.. Heatsink 35C idle.

which heatsink did you use?

[phinix]

Is this the price rise everyone has been shouting lately?
21st March 4TB SSD P3 Crucial £133, today - £248...

[tp4tissue]

which heatsink did you use?

Just the normal aluminum block.

Tp4 has considered the heatpipe / fin-types, but couldn't justify it, because there's no sustained load for at home use. Bringing down the idle temp is enough.

Don't think we can fully saturate even a small heatsink because the average load is far too short. Tried a 200GB transfer, heat wouldn't peak, only went up to 39 C. How often do we even have 200GB to load.

[_rubik]

I'm praying that this AI craze will also drive storage tech since those training datasets are many many PB

Just speculating, the AI boom is more likely to drive up prices in the short term. And on the tech side of things, it seems like the real winner is Nvidia while storage tech is seemingly moving at its own pace, determined largely by cloud service demand.

I'm going to go out on a limb and say that people needed petabyte scale storage clusters well before this craze and will continue to need it for all-time. Plenty of industries rely on similar scales of data, I don't think we'll see a shortage and price spike the same as we say for crypto and AI on GPUs

Also speculating here, but I don't know of many folks training their own models on PB scale datasets either. That sort of training is ludicrously expensive and, outside of the startups getting fun money thrown in their direction, most folks are using off-the-shelf models for most of their heavy lifting.

[Axiom_]

I'm going to go out on a limb and say that people needed petabyte scale storage clusters well before this craze and will continue to need it for all-time.

Indeed. I'd attribute the demand to cloud storage/ cloud computing. It was hyped a while ago and continues to grow rapidly.

Plenty of industries rely on similar scales of data, I don't think we'll see a shortage and price spike the same as we say for crypto and AI on GPUs

That's true. The other influencing factor is that there are more manufacturers in the storage space in contrast to only a handful in the GPU space.

Also speculating here, but I don't know of many folks training their own models on PB scale datasets either. That sort of training is ludicrously expensive and, outside of the startups getting fun money thrown in their direction, most folks are using off-the-shelf models for most of their heavy lifting.

The trend seems to be towards the deployment of cloud-based, pre-trained foundation models which are then fine-tuned by businesses using local datasets. Due to the latter, SMB/enterprise demand for storage is likely to grow more than usual. My personal take is that individual PC users only ever account for a small share of hardware demand.

[tp4tissue]

Well, the only reason the average person would need so much storage is arrrghhh matey movies.  That doesn't run into petabyte, maybe around 500TB.

[_rubik]

Also speculating here, but I don't know of many folks training their own models on PB scale datasets either. That sort of training is ludicrously expensive and, outside of the startups getting fun money thrown in their direction, most folks are using off-the-shelf models for most of their heavy lifting.

The trend seems to be towards the deployment of cloud-based, pre-trained foundation models which are then fine-tuned by businesses using local datasets. Due to the latter, SMB/enterprise demand for storage is likely to grow more than usual. My personal take is that individual PC users only ever account for a small share of hardware demand.

Oh I was only talking about enterprise demand. I agree, consumers have and will always get the dregs of enterprise. Training an AI at home (at this point) is like running a Kubernetes cluster at home. Sure, it mimics the patterns of a production cluster but the scale is effectively a rounding error.

My point is the, even in the enterprise context, few companies are training their own models. And even if they are, they're rarely training from the ground up. More often than not, they're extending a base model. Storage development and therefore prices will continue to increase because we're spinning off more data as a society in a day than we can ever hope to capture, not because we're all training openai-scale models.

Maybe you can argue that we have an increased interest in capturing the more "mundane" data for training purposes, but so much of that data is noise that filtering the raw bits down into a useful signal will always bring us back to the real bottleneck: compute.

The only evidence I can really point to is that researchers were working on transformers for _years_ before this wave. We've only just hit a point where it's computationally feasible to scale the architecture and train the model.

I also acknowledge that, judging by your post history, we're likely on opposite sides of the bullish-bearish spectrum.

[tp4tissue]

Ok, so 1 scenario where Tp4 thinks a bigger SSD heatsink is warranted. If someone is using the SSD slot that gets hot air from the GPU, usually slots toward the bottom, those slots get crazy hot when the GPU is going full out.  These should be either water cooled or at least actively cooled. A makeshift airguide/ heat shield to direct the hot gpu air away from the drive may be easiest.

[phinix]

Looks like we will get bigger SSDs soon, phinix is now a happy hippo

Bring on 8TB SSD for $200!

https://www.techpowerup.com/321557/samsung-readies-290-layer-3d-nand-for-may-2024-debut-planning-430-layer-for-2025

[tp4tissue]

Production cost to ssd is already lower than HDD, right now they're just fleecing the customers.

[TomahawkLabs]

Looks like we will get bigger SSDs soon, phinix is now a happy hippo

Bring on 8TB SSD for $200!

https://www.techpowerup.com/321557/samsung-readies-290-layer-3d-nand-for-may-2024-debut-planning-430-layer-for-2025

The biggest winner here IMO is the users. I've been into PCs since the late 00s. Hard drive wiring was always a pain, now every Mobo has 2 NVME slots and there are now power or data wire to run. Much roomier and easier to manager. I can also see a future in which case manufactures opt out of having 2.5" drives even. Just use NVME.

[Leslieann]

I don't expect case companies to do away with 2.5 drives any time soon, while NVME is great, expandability is low.
Besides, it's pretty easy to hide a mount for a pair of them in a case, you can even just put a set of mounting holes in the case bottom/front rear and call it a day.

But also, case companies are slooooooow to move (due to tooling costs), they're still selling cases designed 20 years ago with new face plates, and they HAAAATE removing options.  Drive bays and such are feature or selling point. It's really only been since the Lian Li 011 that has really changed how cases look and work but even that has drive bays behind the mobo. Expect mobo makers to kill sata ports (and adding ore nvme) before cases really start dropping drive bays.

[tp4tissue]

It's actually expensive now to get cases with good 3.5 drive support.

Wish they'd make a 22bay one.

[Coreda]

Enterprise drives are where the highest capacities are at.

- 15TB used NVMe SSDs from data centers could be bought for $700 last year before prices rose (again).
- The Soldigm 61TB (NVMe) model was selling new for $3700 late last year before the price rises pushed it to $6400.
- Someone on Reddit was selling a 100TB Nimbus SSD (SATA only) with 70% wear for $3k.

They also have better endurance than consumer drives. Eg: the current Solidgm ~8TB (QLC, NVMe) model is twice the endurance of Samsung's QVO (QLC, SATA) equivalent. Their 61TB model endurance is 65PBW (petabytes written not terabytes).

[_rubik]

Enterprise drives are where the highest capacities are at.

- 15TB used NVMe SSDs from data centers could be bought for $700 last year before prices rose (again).
- The Soldigm 61TB (NVMe) model was selling new for $3700 late last year before the price rises pushed it to $6400.
- Someone on Reddit was selling a 100TB Nimbus SSD (SATA only) with 70% wear for $3k.

They also have better endurance than consumer drives. Eg: the current Solidgm ~8TB (QLC, NVMe) model is twice the endurance of Samsung's QVO (QLC, SATA) equivalent. Their 61TB model endurance is 65PBW (petabytes written not terabytes).

I'm all for buying used enterprise hardware, but not drives. The wear stats on those drive are insane, spinning rust or solid state. It's like buying a bike off a pro-tour cyclist in training

[Coreda]

I'm all for buying used enterprise hardware, but not drives. The wear stats on those drive are insane, spinning rust or solid state. It's like buying a bike off a pro-tour cyclist in training

They do have greater endurance though and if the use case is mostly write once read many then even a well-used drive that still has enough endurance left would suit. One can't even buy HDDs over 30-30TB while SSDs have already far surpassed that capacity.

Drives can die at any point, whether new or used. The only difference is the warranty. Various sellers of such used drives offer their own warranties, so up to the user if they're comfortable with that route. Eg: one can buy used 16TB helium HDDs for like a third the price, so you could buy three for a RAID and just replace if one happens to fail (or get a replacement if within the used seller warranty) vs a single one with a manufacturer warranty.

[tp4tissue]

Those wear levels can be overwritten and changed. Same with HDD hours. That's the risk you get with enterprise used.

[Coreda]

Those wear levels can be overwritten and changed. Same with HDD hours. That's the risk you get with enterprise used.

I know some sources (including 'factory recertified' from Seagate themselves) reset SMART data entirely. However judging by power on hours/etc in SMART data (ideally pre-provided by the seller) you can see if it makes sense for the number of years it's likely been in operation, which is how most do this (and also obviously running a full sector scan after purchase for bad sectors).

Everything I've read shows that only SMART resets using vendor specific low level tools are possible, rather than any specific values, hence why users can still gauge this.

[tp4tissue]

Everything I've read shows that only SMART resets using vendor specific low level tools are possible, rather than any specific values, hence why users can still gauge this.

But how do you then gauge whether they're setting up a plausible ruse using such tools? To them it's money, and they know the vast majority of buyers have such low loads that they wouldn't notice the cheat.

[Coreda]

But how do you then gauge whether they're setting up a plausible ruse using such tools? To them it's money, and they know the vast majority of buyers have such low loads that they wouldn't notice the cheat.

I mean we could consider the possibility if you want. If a seller of used server hardware has a model from a particular year (visible from the listing) and resets all SMART data (which is the only option from what I've seen of the tools) then to ramp up the power on hours artificially and be plausible they'd have to have it powered on for a similar amount of time as a realistic used lifespan (let's assume based on endurance spec or warranty period given sellers sell in bulk batches, suggesting scheduled replacements). So from a purely time consideration seems unrealistic, assuming such a lifecycle (ie: not drives from many years earlier).

However what if the drive hadn't run for an expected duration but for the sake of argument instead perhaps only a shorter duration and the seller reset the data (due to some concerning stat level) then left it powered on to clock up hours. To be plausible, since all the other stats were reset, including wear level, they'd have to also artificially induce those, too. Could it be done? Yes. Does inducing realistic-seeming stat levels artificially like wear on an already well-used drive while also offering seller warranties sound plausible? I may have to don a tp hat :p

For the users who aren't browsing sellers who post SMART stats, or who don't care, someone wouldn't need to go through all that rigmarole, especially when manufacturer refurb'd models are already known to reset all data. One can see just from existing sales that most people will buy based on return window or some kind of warranty period.