exged

Once you take SIMD into account, lookup tables frequently lose. Vector gathers are not cheap at all.

What you want for asymmetric cryptography is "always hard to solve, easy to verify". Only a very small number of problems are believed to fall in this category, including factoring and discrete logarithm.

It's far more common to have problems that are "sometimes easy to solve, sometimes hard to solve, easy to verify". That describes this problem as well as most NP-hard problems that with good heuristic solutions. The problem is those "sometimes easy to solve" cases break your cryptography!

That's for model inference. For training OpenAI said they used around 3000 PetaFLOPS / days on the largest GPT-3 model. That translates to about 300 Nvidia A100 GPUs if you want training to finish in a month (any slower and your researchers are not going to be able to make much progress). A system like that would cost at least $5M, probably more like $10M.

SVE doesn't really solve #2 and #3 by itself. You can still design a core with huge SIMD units that require a ton of power and area, resulting in downclocking and relatively poor performance on non-SIMD workloads.

#1 is a big deal though. It's a huge burden to rewrite SIMD code over and over for every new instruction set, so people (and compiler implementers) just don't use it too often. Then you're paying power and die area for completely useless SIMD units.

It's pretty simple - for regular users Chrome is faster, has more features, uses less power, and has better website support. It is also aggressively marketed by the largest advertising company on the planet.

It wasn't always this way. For example, FF beat Chrome in battery life until a few years ago, and FF extensions used to be superior in functionality Chrome extensions. Chrome eventually caught up in the few areas they were seriously lagging in, or Firefox shot itself in the foot trying to morph itself into a copy of Chrome (eg. multiprocess architecture, WebExtensions).

Today there is little reason to use any browser other than Chrome / Chromium or Safari (which is still superior on Mac / iOS in a few aspects).

ML contributes to a significant fraction of revenue at three of the world's largest companies (Amazon, Google, Facebook - largely through recommendations and ad ranking). It also drives numerous features that other tech companies build into their products to stay competitive (think FaceID on iPhone). Hard to argue that it doesn't move the needle...

Liquid hydrogen tanks are definitely a different problem, but I don't know if "orders of magnitude safer" is necessarily true. Many spacecraft have run into problems involving liquid hydrogen storage for a variety of reasons:

- Liquid hydrogen is so cold that many metals weaken

- Exposure to hydrogen itself weakens many metals (hydrogen embrittlement)

- Liquid hydrogen expands rapidly if exposed to heat, so venting is needed

- Hydrogen leaks through the smallest of seams

Now imagine trying to build a tank that prevents all that from happening, and it has to be lightweight for usage on spacecraft / aircraft.

High-level IC roles (in the companies that have them) are still leadership roles. They are just technical leadership roles instead of people management roles. They may not have direct reports, but they are still responsible for getting big projects done, usually by organizing a large number of engineers.

It's desirable, but typically hardware devices have DMA-capable buses that can read / write to arbitrary physical memory. So a buggy or malicious WiFi driver may be able to use the WiFi DMA hardware to write into the GPU driver's memory, and no pure software mechanism can stop it from doing so.

IOMMUs solve this problem by giving hardware devices their own virtual address spaces.