Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Kioxia and Dell cram 10 PB into slim 2RU server

blocksandfiles.com

Post reply on HN
Print view View original

61–70 of 116 posts

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#61

Post by bigyabai »

fgfarben wrote: post #38
evil-olive wrote: post #30

Earlier quoted context omitted.

no, absolutely not. orbital datacenters are never going to happen, it doesn't matter whether you try to frame them as compute or storage or whatever else. the extreme density of these SSDs is actually an anti-feature in the context of spacecraft hardware. the RAD750 CPU [0] for example uses a 150nm process node. its successor the RAD5500 [1] is down to 45nm. that's an order of magnitude larger than chips currently ma…

[flagged]

On Grahams hierarchy of disagreement, your comment is 3 tiers lower than the parent comment.

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#62

Post by crote »

trvz wrote: post #20
retired wrote: post #16

[flagged]

Not quite yet. The interesting thing here is ~256TB in a single drive, but it's in E3.L form factor. I have about 160TB on hard drives that I'm waiting to offload onto a single SSD. But that needs to come with a connector that has adapters to USB-C, so I can attach it to my Macbook Neo. Hopefully they get it a bit more dense soon and into the 2.5" NVMe form.

E3.L is just fancy-shaped PCIe, is it not? What's stopping the standard off-the-shelf NVMe-to-USB converter chips from being used?

Given this disk is going to cost something like $40k, what's another $500 for having a Chinese hw eng throw one of those chips together with an E3 connector on a PCB for you, and 3D printing a neat housing?

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#64

Post by buzer »

tempest_ wrote: post #60
buzer wrote: post #42

Earlier quoted context omitted.

The list price seems to be ~40M. https://www.dell.com/en-hk/shop/servers-storage-and-networki... Select the 40 slot chassis and put 40 of those 245TB disks in. Comes out at ~HK$317M. Of course HK prices might also be higher than what Dell USA offers. Now how heavy the discounts you can get I don't know.

My understanding is no one actually pays Dell sticker price. They list that price on the website but if you talk to whatever they are calling their sales reps you get the real price.

I know, I'm just not sure how high discount you are going to get in this kind of system. My understanding is that the discounts normally tend to be something like 30-40% so it would still be within same magnitude (10M+).

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#65

Post by wildzzz »

fancyfredbot wrote: post #10

The very first sentence of this article mistakes Terabytes and Petabytes. I used to dismiss the entire article as poor quality on seeing a mistake like this. But these days it also feels like an indicator the article was written by a human and might actually have something interesting to say. Sadly not in this case though - the Kioxia drives are interesting, but the fact that Dell has put some in a box is much less s…

All it really means is that big corps that have already existing sales relationships with Dell will be purchasing them in the next fiscal year. Anyone else that needed this level of storage density has already built their own boxes.

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#66

Post by wildzzz »

evil-olive wrote: post #30
NitpickLawyer wrote: post #12

There's been a lot of talk about orbital DCs lately, but with these levels of density, orbital CDNs might be a more obvious usecase. It would be interesting to see if something like Starlink can use something like this to cache media content and reduce their overall data moving through the constellation. It could even be worth it to have some satellites in higher orbits (even GEO if the ground hw can reach it) dedica…

no, absolutely not. orbital datacenters are never going to happen, it doesn't matter whether you try to frame them as compute or storage or whatever else. the extreme density of these SSDs is actually an anti-feature in the context of spacecraft hardware. the RAD750 CPU [0] for example uses a 150nm process node. its successor the RAD5500 [1] is down to 45nm. that's an order of magnitude larger than chips currently ma…

The RAD750 is like 20 years old and is the absolute king in high reliability in the most extreme radiation environments. LEO is much more forgiving and there's plenty of examples of commercial gear operating in it. You could definitely put this much storage into LEO along with some EDAC and be fine for a few years.

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#67

Post by DannyBee »

lostlogin wrote: post #15
gosub100 wrote: post #8

Earlier quoted context omitted.

You can't buy this stuff anymore. They are leased and rented through layers of middlemen.

> anymore Could you ever buy it?

The 64's and 128's could be bought, and were not even totally unreasonable, price wise, at intro (IE Concretely: a 64TB was like 6000 bucks last year. You could get them easily. It's now 25,000 for the same SSD.

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#68

Post by crote »

dijit wrote: post #32
trvz wrote: post #20

Earlier quoted context omitted.

Not quite yet. The interesting thing here is ~256TB in a single drive, but it's in E3.L form factor. I have about 160TB on hard drives that I'm waiting to offload onto a single SSD. But that needs to come with a connector that has adapters to USB-C, so I can attach it to my Macbook Neo. Hopefully they get it a bit more dense soon and into the 2.5" NVMe form.

I've been waiting with bated breath for a SATA 3.5" SSD with high capacity. I might be waiting forever, because clearly there's nothing coming. Though I'm not sure if it's because it's technically difficult (high power consumption to keep the flash lit?) or something else. I'm aware that it leaves performance on the table for the chips, and probably that means that unit economics means that for the yeild: OEMs would…

There's just no benefit to using SATA for it. Even a PCIe Gen3 x1 link provides a significantly higher performance - and those aren't exactly rare these days. Why invest a huge amount of time and money into building a controller chip which is significantly worse than its competitors in every way? Even if you're very interested in backwards compatibility it would make more sense to go for PCIe-based U.2 instead. And 3.5" is just a waste: look at 2.5" SATA SSD teardowns, they are mostly empty space.

If you really want a classic hotswap bay form factor, something like the Aoostar WRT Max allows installing a bunch of M.2 SSDs in a 3.5" bay. The QNAP TBS-h574 gives you five swap bays for M.2 SSDs - albeit in a cute custom form factor. Just want a whole bunch of storage? The Asustor Flashstor Gen 2 has up to 12(!) screw-down M.2 slots. At 8TB per slot that's 96TB of storage - significantly more than the 40TB 4-bay NAS you are proposing.

Or wait for someone to build a NAS which accepts EDSFF SSDs. But don't count on it, because there's no market for prosumer-level EDSFF drives (nobody has a bay for it, and M.2 is a far more attractive option for most people), so there's no market for EDSFF-compatible NASes either. Unless you plan on shipping them with M.2-to-EDSFF adapters - but at that point why not save a whole lot of space by going directly for M.2 instead?

It's a bit like asking for a Mini-ITX motherboard with an SP5 socket so you can plug in the latest Epyc CPU, and wanting it to have DDR3 memory slots: even if it's technically possible at all, what market is it supposed to serve?

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#69

Post by DannyBee »

bracketfocus wrote: post #13
louwrentius wrote: post #5

What would this cost?

They are likely 200USD+ per TB, so one 250TB drive would be ~50,000USD. There’s probably bulk pricing, but if you bought 40 drives separately thats 2,000,000USD in storage alone.

It's actually now at least 400 per tb.

The 64's are 25k, up from 6k a year ago. I have to imagine the 128's or 256's are at least 500/tb

Re: Kioxia and Dell cram 10 PB into slim 2RU server

#70

Post by nine_k »

killerstorm wrote: post #57
evil-olive wrote: post #30

Earlier quoted context omitted.

no, absolutely not. orbital datacenters are never going to happen, it doesn't matter whether you try to frame them as compute or storage or whatever else. the extreme density of these SSDs is actually an anti-feature in the context of spacecraft hardware. the RAD750 CPU [0] for example uses a 150nm process node. its successor the RAD5500 [1] is down to 45nm. that's an order of magnitude larger than chips currently ma…

It's possible to run modern GPU on a sattellite: https://www.starcloud.com/starcloud-1 Some error rate is acceptable for uses which aren't "mission-critical".

Redundancy and error correction should be built into such huge arrays anyway. Its amount just needs to be increased.
Post reply on HN