Meanwhile I'm still dreaming about any consumer and affordable 32TB or even 16TB portable SSD. Innovation and market for consumers are going backwards.
Funny thing is that one of the best you can get is the Crucial (Micron) 8TB one but even that one gets more expensive. I have the feeling it will be gone completely soon.
There was once a 2.5" SSD Mushkin Source 16TB SATA drive. At its cheapest it was ~1700 USD (or 1500 EUR). That was mid 2023 (like 3 years ago!).
Nowadays it feels like that this time and price region is like decades away in the future. I was hoping I can store more data in future on modern tech like SSDs and not less.
Note how that is still well in excess of what e.g. AWS EBS GP3 volumes offer (or at least used to, though even now their "80K IOPS" is measured with 64 KiB random transfers, whereas Micron measured that 42K IOPS with 4 KiB random transfers), which is what the person above is gesturing towards.
The same EBS GP3 used to be specified with 16K max IOPS at 16 KiB random transfers until pretty recently.
That is pretty awful write performance. Does anyone know more about this? I assume all of these hyperdense SSDs suffer from the same drawback. Also, I heard that the E3.L interface can support up to 16x lanes, but there are no practical commerical products at this point.
The u.2 form factor is slightly larger than a 2.5" drive. I can imagine the entire space in it taken by Flash chips. I can't imagine what cooling scheme do they employ for the chips in the middle.
Apparently TDP is 30 watts¹, according to the product brief. I would imagine it's a single PCB with flash chips on both sides then thermally bonded to the aluminum chassis. That should keep all chips at approximately the same temperature. On its own it could be easily air cooled, but with 24 in a 2U chassis you'll be having some decently hefty forced air over the drives.
1. For comparison, an HDD usually comes in around ~10 watts
The U.2 form factor is a 2.5" drive, not larger than it.
"U.2" does not change anything in the mechanical characteristics of a 2.5" drive, it just replaces the SATA or SAS electrical interface with a NVMe electrical interface.
You can mount a U.2 drive in any location intended for 2.5" drives, as long as its height can fit there.
However, 2.5" drives come in various heights. Many laptops and mini-PCs that accept 2.5" drives accept only some of the smaller heights and they do not accept the greater heights, like 15 mm, which are typical for enterprise SSDs and HDDs, regardless whether they have a NVMe, i.e. U.2, or a SAS interface or a SATA interface.
This new high-capacity U.2 SSD has the standard 15 mm height of the 2.5" form factor.
The transfer rates limit how much each chip can be active at any given time, so a heat-aware writing allocator can pick the least active blocks for the next writes and distribute the heat accordingly. Even if it’s not heat-aware, the tendency will be that the writes will be distributed over as many chips as there are, and so will be the heat generated.
Now, I would LOVE to see this much SLC flash on a direct to bus attachment setting.
Over the past few years the main improvement in SSD capacity has been due to them stacking an ever-increasing number of NAND layers in a single chip, with state-of-the-art SSDs already having over 300 layers.
No need to worry about cooling when each layer in the sandwich is only a fraction of a micrometer thick!
the u.2 form factor indeed evolved from chassis designs that were originally 2.5" drives. It's now kind of becoming obsolete with new designs using things like E1S, E1L (exactly the correct height to be slotted into a 1U server, it's like a slightly wider M.2, but meant to be insertable and removable), and E3S and E3L.
Note that the 245TB is an E3L, the half size version of it come in smaller size.
Getting rid of 30 watts of heat is trivial compared to say, 300 (I don't quite know how to read that ratio since a 2.5kW SSD seems a little high to me).
With a modern CPUs hitting 400W it's already a problem to fill a rack top down with servers like you could do before: too many heat to dissipate and transfer, too much power to provide in the first place.
Just imagine something like 2S 9565 in at least 2U machines: with 10 server x 2U x 2 CPU you would have 8kW in the processors alone and you didn't even fill half of the 42U rack.
God damn. I know somebody that became a multi-millionaire from web hosting in the 2000s and his entire data center back then could have been replaced with just one of these SSDs.
Want, but then need two for reduncancy... then a spare for recovery... why not 3 raid or zfs... imagine the resilver time on this. It's hit the limit of data surety surely.
> What accounts for the premium price/TB of these extremely high capacity enterprise-targeted drives?
Spare capacity, mostly. That’s why they have higher endurance. If you want to double the endurance of a given drive, tell the controller to allocate twice as many spare blocks and report less capacity than you would otherwise.
In this case, you are also paying a premium for the PCIe attachment instead of SAS, and a lot for price elasticity. You see, with drives like these you slash space and energy consumption in relation to HDDs by a large number, and that allows you to pay a premium for the device, because, at the end of its lifetime, it’ll have more than covered the cost difference in saved space and energy.
4-5x times what it would have been if not for the demand from AI. According to my rough calculation 4-8tb ssd drives were going to reach parity with hdd this year
The datasheet shows 3GB/s sequential write, which for 245.76TB means writing the whole drive takes around 22h45m. Odd that the endurance is specified as "1.0 SDWPD", which is almost meaningless since the drive takes roughly that long to write at full speed.
At scale, 1.9 times more energy is required for an HDD deployment
...but those HDDs are going to hold data for far more than twice as long. It's especially infuriating to see such secrecy and vagueness around the real endurance/retention characteristics for SSDs as expensive as these.
On the other hand, 60TB of SLC for the same price would probably be a great deal.
Perhaps their usual buyers just care less about retention?
Those drives aren't going to be used for cold storage, and it is basically a guarantee that there will be checksums and some form of redundancy. Who cares whether the data is retained for 10 or for 15 years after writing when you can do a low-priority background scrub of the entire drive once a month, and when there are already mechanisms in place to account for full-drive failure?
QLC retention reported to be around 1 year in unpowered state. I would assume, that drive does background refresh, though. No idea what effect it has on total drive lifetime. It is still mean that if you use it for cold storage it has to be powered.
Why is it mean? Why would you want to use a technology that is unsuitable for cold storage for cold storage? You won't even get the power / IOPS benefit if all it does is an infrequent replication of data and is then switched off.
I believe it has read speeds of 13GB/s, not 3 (unless you are referring to an equivalent array of 10 HDD). It will almost certainly be used to store training datasets and model weights. Which I assume are good use cases for fast sequential reads.
Can someone who knows explain what is the benefit of having all that data in one ssd instead of splitting it up into hundreds of individual drives? Does the single ssd benefit is more performance or does it really tuen out to be cheaper than hundreds of individual drives?
It’s about density in a datacenter. With this you have 1PB in 4 drives, fitting in a 1u rack, which is just incredible. Also these drives don’t use regular SATA or SAS, they use PCIe, so these drives are also quite fast in comparison. Density has a power efficiency aspect as well both in just having fewer drives and requiring fewer servers to put drives into.
A 42U rack filled with 1u servers with 8 drives each, will have 84PB of data. It feels like it was a few month ago where you could buy a rack with 1PB of storage, and that was awesome. Not anymore.
You’re actually right, it’s just that datacenters like density and will gladly split your data onto hundreds of these little amazing magical bits of technology rather than hundreds of less magical ones in the same physical volume.
DENSITY. Hyperscalers want to store as much data per rack and per data center as possible. They will eventually have hundreds of thousands of these drives.