thulle

Any perceived difference isn't due to any technical superiority of vinyl though, more often it's the other way around: you can't dynamically compress audio as much and still have it playable from vinyl, so there's often a different, more dynamic, master/mix going onto the vinyl.

Sighted testing and different audio levels are other common sources of perceived vinyl superiority.

Many people enjoy the ritual of grabbing a vinyl and setting it up for playback, having physical access to coverart and maybe lyrics, which are entirely valid reasons to prefer vinyl, but that isn't audio quality.

I see hacspec using secret-integers: "wrapper around integer types for constant-timedness". Does Hacspec/Bertie have assurances about constant-timedness too?

Link seems broken.

> It is perfectly good enough for the error code enumeration to be statically randomized into hard coded constants.

A comment points out that they aren't randomized:

> The values used were chosen such that it takes a large number of bit flips to change from allowed to denied. Using random values doesn't really protect against this attack.

Some data for me:

Nitter: I get all 8 posts in the thread in 18 requests, 207 kB, 169 kB transferred.

X: I only get the first post, 128 requests, 11.45 MB, 2.11 MB transferred.

The 608.5 mm² 4090 with 76.3 B transistors isn't too far off from the 814 mm² H100 w 80 B transistors. Seems like mostly more area for memory interfacing?

Assuming they're not selling the 4090 at a loss and being generous that the whole cost is the chip cost, that makes for a $1600 chip. Scaled to the H100 we get $2140. That's a pretty minuscule fraction of a $35k card.

That makes sense though, are the chips themselves the bottleneck in production?

> 5nm dies are the most expensive part of the whole solution, meaning there is also a 65% pricing advantage (though some of this advantage is offset by more complex packaging and other cheaper dies that go into mi300x as well as more HBM chips).

That doesn't say anything about the fraction of the cost on the finished product though? Those numbers just say it could be something like 2% vs 6% of the cost of the card.

Does it matter? Is wafer cost that large of a fraction of the price of these insanely expensive cards?

Better output than the smaller llamas in my limited testing, but it's surprisingly slow:

Output generated in 101.74 seconds (0.98 tokens/s, 100 tokens, context 82, seed 532878022)

Output generated in 515.46 seconds (0.99 tokens/s, 511 tokens, context 27, seed 660997525)

Checking nvidia-smi it stalls at ~130W (out of ~470 W max) power usage, ~25% GPU usage and ~10% memory bandwidth usage. There's fairly much traffic on the pci-bus though, and the python process is stable at 100% usage of one core. GPU possibly limited by some thing handled in python? Pausing the GPU-accelerated video-decoding of a twitch stream it get a surprisingly large boost:

Output generated in 380.42 seconds (1.34 tokens/s, 511 tokens, context 26, seed 648992918)

> February 08, 2022

They'd need 20 months to manage to notify users?

Coming back home and checking 2.2.1 out zfs instantly started spewing write & checksum errors due to #15533. Both this and #15526 seem to have underlying issues from <2.2 that are just more easily triggered now. First one also confirmed on FreeBSD now.

Holding off on 2.2 seems recommended, and if you're keeping critical data on OpenZFS it might be a good idea to give the issues a glance. The 2nd one might have the same underlying solution as an issue that has given me system freezes when closing in on ~90% pool usage using 2.1 on top of LUKS.