en4bz

OpenGL and Vulkan are provided by the GPU vendor you can't statically link them.

Whiteboard coding almost always devolves into leetcode which also requires at home study. You're going to be spending evenings and weekends coding something in either case.

Haven't most plans moved to PPO since Obamacare. They usually allow you to just skip primary care nowadays.

Flatbuffers, Cap'n Proto, or SBE over RDMA.

I'm assuming you're referring to the golang model of statically linking everything. That's not really doable when many popular libraries are (L)GPL'd like glibc and libstdc++. It also doesn't work if you want to provide a shared library and need to be compatible with every possible system. That's not my opinion it's just a deficiency of bazel.

bazel has provisions for patching stuff from http_archive if you supply a patch

That's a hack.

It's literally the default. How can the default be an anti-pattern. I doubt you're using C/C++ because you don't seem to understand the issue.

You're assuming you have control over all the deps and can set `linkstatic` on them.

qux is now forced to be static. Consumers should be able to choose whatever they want. I don't know what end users want to do.

The logic is backwards though. I may have multiple consumers of a library some of which may want static some of which may want dynamic. You need to create/import new targets to do this even though the original target creates both static and dynamic libs by default.

Non-Hermetic is the default for C/C++. And if you plan on using system provided libraries to support multi OSes then you can't use it.

None, bazel's caching implementation is broken because they don't even know or specify what constituents a build hash/key. See this issue from 2018 that's still open [1].

[1] https://github.com/bazelbuild/bazel/issues/4558

No, they are extremely broken. You can only choose:

- Link everything static.

- Link everything dynamic.

- Link user libs as static and system libs as dynamic.

There is no easy way to link a single user lib static/dynamic without resorting to hacks/workarounds like re-importing the shared library or defining weird intermediate targets. It's completely broken.

CUDA Driver API or Runtime API remoting?

That's the biggest problem with this model. With inference it's better to just use a dedicated model server. For training it's better to deploy on a massive dedicated machine. The only real use case left over is experimentation and debug for devs or students.

I doubt this does multi-server. All the GPUs probably have to be on the same machine.

Infiniband avoids the network stack. Has ~2us latency these days over LAN.

bazel is trash.

100% agree with this. There's nothing in the new standards past 17 that actually helps me in real life day to day work. Most of the new features are solely for code-golfers (see ranges). Mean while co-routines and other concurrency extensions have languished for years as TSs with only partial support. Here's the full list of TSs [1], half of which only have partial or no support despite being 5+ years old.

[1] https://en.cppreference.com/w/cpp/experimental

It's mostly run by code-golfers which is much worse.

All the jobs you listed have professional associations that require passing exams. If you required the same level of testing and exams for devs I doubt half the software industry would pass.

Kids theses days hate using technology that's older than themselves? Kerberos is amazing and it make me sad to see all these people reinventing stuff that was solved 30 years ago. The number of companies selling SSO and doing it poorly (see this weeks okta hack) is unfortunate.

You can use rseq on linux to avoid this.

If you've ever tried reading the source for glibc or libstdc++ you know this is terrible advice.

So `dlmopen`?

I wonder if Chronic Wasting Disease has finally jumped to humans given that NB has a large number of Deer/Moose hunters.

Tying to the previous point, currently ZLUDA does not support asynchronous execution. This gives us an unfair advantage in a benchmark like GeekBench. GeekBench exclusively uses CUDA synchronous APIs

Any "professional" application solely use async APIs so while these numbers may look impressive something like tensorflow or pytorch would either not run or be incredibly slow.

If they go through arbitration they'll end up like DoorDash [1].

[1] https://www.latimes.com/business/story/2020-02-11/doordash-a...

The A100 is on TSCM 7nm.

All this on Samsung 8nm (~61 MT/mm2). They didn't even feel the need to use TSMC 7nm (~100 MT/mm2). Probably keep the price down and to reserve capacity at TSMC for the A100.

This is like the anime hero/villain (depending on your perspective) equivalent of fighting at half power.