zzzoom

It's amusing because i915-derived integrated graphics had the largest GPU share for a long time, but they still tried to make an x86-based GPU instead.

If they lose 6B but manage to keep their customers whole, I don't think it's that bad of a message. Too big to fail territory.

You have an interesting definition of "efficient and performant".

Similar technique around the same time presented at SIGGRAPH:

https://www.youtube.com/watch?v=FKXOucXB4a8

https://people.csail.mit.edu/mrub/VisualMic/

Is anthropic trying to overthrow any governments this month like Meta is?

Even 3.12 will be EOL half a year before RHEL 8's.

And "support" doesn't mean you get the same package selection as the original:

  $ yum search 'python3' 2>&1 | grep '^python3-' | wc -l
  1551
  $ yum search 'python3' 2>&1 | egrep '^python3(8|9|\.)' | wc -l
  279

Enterprise distros support their python for 10 years anyway. The lack of Python LTS releases results in most versions having longer lifetimes in practice than intended:

  - 3.14: Ubuntu 26.04
  - 3.12: RHEL 10, Ubuntu 24.04
  - 3.10: Ubuntu 22.04
  - 3.9: RHEL 9
  - 3.8: Ubuntu 20.04
  - 3.6: RHEL 8 (EOL in 2029)

I had to do this once, after a blackout during a BIOS flash bricked my motherboard. Fun stuff. I don't think the working motherboard donor knew what he was signing up for.

Meta will always do whatever it takes to get more users.

No way to avoid the memory cartel, even if CXMT catches up.

Each minute that a GPU isn't running is money evaporating

Wow, if only the Chinese knew that! So easy!

Matrix multiplication is one of the few operations that isn't regularly limited by memory bandwidth. BLAS implementations come with several heavily optimized, architecture-specific versions of sgemm.

Isn't the delta fed to an arithmetic coder?

Not as many as the ones that the US snatched in operation paperclip

I'd bet on the inverse: China scaling DRAM production until the price crashes, and the whole US stock market that is propped on top of that scarcity going down with it.

Prefill (GEMM) is compute bound, decode (GEMV) is memory bound.

EPYC 9965: 614GBps of 12-channel DDR5-6400

A100: 1935GBps of HBM2e

Most of those FLOPS are constrained by memory bandwidth.

A social safety net

And definitely not more evil than the workers at current Meta.

Is it? Spack has only one package that depends on chapel.

And state actor LLMs probably don't need convincing

Copy Fail was found using AI and made our last week miserable (thanks RedHat). Anyone with access to a SOTA model can point it at any package in your stack and go fish. You're underreacting.

> GabeN also had the wrong take in that it's a "supply problem" or whatever nonsense he said.

It is a supply problem. Steam regional pricing and game passes have demolished piracy in countries where people wouldn't have dreamed of paying for a game 15 years ago. And so did Netflix for a while with video, but then everyone had to jump on the bandwagon and now piracy is flourishing again.

If the image contains information required to bring up the device, why isn't that data shipped in firmware?

At some point SBCs that require a custom linux image will become unacceptable, right?

Right?

It's predatory pricing.

If any of those addresses sold a single sat the price would crash hard.

That's a genius idea, keep adding broken stuff into the standard until there's no choice but to break compatibility to fix it.

Imagine considering some bird poop staining the paint dangerous instead of the air pollution that's slowly killing you.