qupada

The estimated training time for GPT-4 is 90 days though.

Assuming you could scale that linearly with the amount of hardware, you’d get it down to about 3.5 days. From four times a year to twice a week.

If you’re scrambling to get ahead of the competition, being able to iterate that quickly could very much be worth the money.

Fortunately, there’s an extension that solves that: https://microsoftedge.microsoft.com/addons/detail/ajgodcbbfnpdbopgmfcgdbfhabbnilbp

I’d be curious to see how much cooling a SAS HBA would get in there. Looking at Broadcom’s 8 external port offerings, the 9300-8e reports 14.5W typical power consumption, 9400-8e 9.5W, and 9500-8e only 6.1W. If you were considering one of these, definitely seems it’d be worth dropping the money on the newest model of HBA.

I’m definitely curious, would only personally need it to be NAS + Plex server for which either of the CPUs they’re offering is a bit overkill, but it’s nice that it fits a decent amount of RAM, and you’re not forced to choose between adding storage or networking.

p is stored in the balls…

The phrase “cutting off your nose to spite your face” comes to mind.

Not in so much detail, but it’s also really hard to define unless you’ve one specific metric you’re trying to hit.

Aside from the included power/cooling costs, we’re not (overly) constrained by space in our own datacentre so there’s no strict requirement for minimising the physical space other than for our own gratification. With HDD capacities steadily rising, as older systems are retired the total possible storage space increases accordingly…

The performance of the disk system when adequately provisioned with RAM and SSD cache is honestly pretty good too, and assuming the cache tiers are adequate to hold the working set across the entire storage fleet (you could never have just one multi-petabyte system) the abysmal performance of HDDs really doesn’t come into it (filesystems like ZFS coalesce random writes into periodic sequential writes, and sequential performance is… adequate).

Not mentioned too is the support costs - which typically start in the range of 10-15% of the hardware price per year - do eventually have an upward curve. For one brand we use, the per-terabyte cost bottoms out at 7 years of ownership then starts to increase again as yearly support costs for older hardware also rise. But you always have the option to pay the inflated price and keep it, if you’re not ready to replace.

And again with the QLC, you’re paying for density more than you are for performance. On every fair metric you can imagine aside from the TB/RU density - latency, throughput/capacity, capacity/watt, capacity/dollar - there are a few tens of percent in it at most.

Being in an HPC-adjacent field, can confirm.

Looking forward to LTO10, which ought to be not far away.

The majority of what we’ve got our eye on for FY '24 are SSD systems, and I expect in '25 it’ll be everything.

There’s some space occupied by the servo tracks (which align the heads to the tap) in LTO, but if we ignore that…

Current-generation LTO9 has 1035m of 12.65mm wide tape, for 18TB of storage. That’s approximately 13.1m², or just under 1.4TB/m².

A 90 minute audio cassette has around 90m of 6.4mm wide tape, or 0.576m². At the same density it could potentially hold 825GB.

DDS (which was data tape in a similar form factor) achieved 160GB in 2009, although there’s a lot more tape in one of those cartridges (153m).

Honestly, you’d be better off using the LTO. Because they’re single-reel cartridges (the 2nd is inside the drive), they can pack a lot more tape into the same volume.

We’ve done this exercise recently for multi-petabyte enterprise storage systems.

Not going to name brands, but in both cases this is usable (after RAID and hot spares) capacity, in a high-availability (multi-controller / cluster) system, including vendor support and power/cooling costs, but (because we run our own datacenter) not counting a $/RU cost as a company in a colo would be paying:

• HDD: ~60TiB/RU, ~150W/RU, ~USD$ 30-35/TB/year
• Flash: ~250TiB/RU, ~500W/RU, ~USD$ 45-50/TB/year

Note that the total power consumption for ~3.5PB of HDD vs ~5PB of flash is within spitting distance, but the flash system occupies a third of the total rack space doing it.

As this is comparing to QLC flash, the overall system performance (measured in Gbps/TB) is also quite similar, although - despite the QLC - the flash does still have a latency advantage (moreso on reads than writes).

So yeah, no. At <1.5× the per-TB cost for a usable system - the cost of one HDD vs one SSD is quite immaterial here - and at >4× the TB-per-RU density, you’d have to have a really good reason to keep buying HDDs. If lowest-possible-price is that reason, then sure.

Reliability is probably higher too, with >300 HDDs to build that system you’re going to expect a few failures.

They’re not really particularly low power.

Quick search suggests around 8W power consumption with a 2 ohm heater, which at the approximately 4V of a charged Lithium-Ion battery (V=IR, P=VI) checks out to around a 2A draw.

Similar results suggest the batteries inside are in the neighbourhood of 0.75Ah (3.7V nominal) = 2.8Wh. I don’t know how much of that capacity actually gets used during the “lifespan” of the vape, but I’d guess half would be a good estimate. In any case, probably safe to assume you need to pack around 2Wh in at minimum.

A Lithium AA battery (Li-FeS2 chemistry) gives you 3.4Ah @ 1.5V = 5.1Wh, but has a maximum discharge current of 2.5A (only 3.8W). The AAA is only 1.2Ah with 1.5A discharge, but two of them would give you 3.6Wh and 4.5W, closer to the target but still under.

You could probably arrange this in some sort of configuration whereby the batteries charge a capacitor and that runs the heater, at those kind of numbers it’d need to be at most a 2 seconds off for 1 second on deal, but that honestly seems like it should be fine for, y’know, vaping. Might just need to have an on/off switch to avoid draining the batteries when you’re not using it.

But I guess we’re at the point where manufacturing Li-Po cells happens in such vast quantities that the extra electronics to charge a capacitor from a 1.5V battery probably cost more.

What sets some of Boox’s models apart from the other e-readers is they’re full Android devices; you can install most apps from the Play Store. Perhaps not as great for battery life, but a world apart so far as functionality goes (and you can even install the other e-book vendors’ apps if you have existing purchased content).

In the “pocketable” size category, Palma which is a phone form-factor device (I have one of these, has been great), the Page looks very much inspired by the design of the Kindle Oasis, or the Tab Mini C has a colour e-ink display.

To expand on @doeknius_gloek’s comment, those categories usually directly correlate to a range of DWPD (endurance) figures. I’m most familiar with buying servers from Dell, but other brands are pretty similar.

Usually, the split is something like this:

• Read-intensive (RI): 0.8 - 1.2 DWPD (commonly used for file servers and the likes, where data is relatively static)
• Mixed-use (MU): 3 - 5 DWPD (normal for databases or cache servers, where data is changing relatively frequently)
• Write-intensive (WI): ≥10 DPWD (for massive databases, heavily-used write cache devices like ZFS ZIL/SLOG devices, that sort of thing)

(Consumer SSDs frequently have endurances only in the 0.1 - 0.3 DWPD range for comparison, and I’ve seen as low as 0.05)

You’ll also find these tiers roughly line up with the SSDs that expose different capacities while having the same amount of flash inside; where a consumer drive would be 512GB, an enterprise RI would be 480GB, and a MU/WI only 400GB. Similarly 1TB/960GB/800GB, 2TB/1.92TB/1.6TB, etc.

If you only get a TBW figure, just divide by the capacity and the length of the warranty. For instance a 1.92TB 1DWPD with 5y warranty might list 3.5PBW.

To make it worse, we have our own in New Zealand, which is the (worldwide) original of that format. The Aussie series is a spin-off.

https://en.wikipedia.org/wiki/Border_Patrol_(New_Zealand_TV_series)

You just know that if they did have a support email address, it’d just reply with “💩”.

This video about ex-Soviet RTGs of questionable radioactive source choice is quite a good watch

https://www.youtube.com/watch?v=NT8-b5YEyjo

NASA apparently used RTGs for deep space missions only, while in the same timeframe the Soviets scattered them all across the countryside, then promptly forgot about them.

T’ain’t enough. Gotta block everything they do, everywhere on the internet.

As someone so eloquently put it: you might not have a facebook profile, but facebook has a you profile.

If you’ve ever seen a “share on facebook” button on another website, they’ve been watching you.

Heh, came here to post that and you beat me to it by 40 minutes.

Other user’s comment up thread about “churning vomit water”… accurate.

Something not so far mentioned is Tree Style Tab.

If you habitually have a lot of tabs open, you’ll probably know how annoying it is finding things when each page title has been condensed down to 4-5 characters. On widescreen displays (especially 16:9), vertical pixels are also a lot more precious, while horizontal ones are plentiful.

For me (3840×2160 display, 200% scale), its vertical tab sidebar fits about 30 tabs before needing a scrollbar, and you get a full width title for each and every one.

It can be a bit of an adjustment at first, but I’ve been using this since the pre-WebExtensions days (since around Firefox 4.0), it’s definitely one of my must-haves.

Induction elements “cycle” on and off – hundreds or thousands of times per second […] There is no human perceptible duty cycle

See unfortunately what you’re describing here are good induction stoves, which is not the majority of what is on the market.

I’ve seen far too many of the bad kind, with duty cycles measured in the tens of seconds. Your 7/10 on the dial could be - like a non-inverter microwave - something in the neighbourhood of 7 seconds on / 3 seconds off. At that point they can actually be worse to use than old halogen glass cooktops, which at least remain hot during the off part of their thermostat’s cycle.

This is not even just cheap no-name crap either, have witnessed it with big-name-brand in-bench stovetops with four-figure pricetags.

If you’re doing something like poaching eggs (which typically calls for a wide, flat pan), you’ll actually see the water starting and stopping boiling in a cycle as it switches. Absolutely terrible.

Yeah, it’s joined Facebook and Twitter on that “do not click” list for me.

You’d think that quitting cold turkey would have been hard, but it somehow just hasn’t been.