shihab

I brought up Israeli-American donors because that’s what is relevant in the context of the story we’re discussing. We are talking about a war many right wing Israelis wanted for decades. If it were a general discussion about Citizens United and I focused on lobbying from only this group, perhaps your argument would have carried water.

Anyway, here’s Trump himself detailing the extraordinary access to White House this lobbying bought Adelsons:

https://www.reuters.com/world/us/trump-salutes-mega-donor-mi...

Exactly what part of my statement was dog whistling? Can you stop throwing around this serious accusation of antisemitism without any attempt to substantiate your claim?

Citizens United is an existential threat for USA. You cannot have Israeli-American dual citizens pouring $200 million dollars in elections. and that’s just her alone. This is simply not sustainable.

Another mid east war entirely on Israel’s behalf, another war Americans will pay tax for, die for- just so Israel can keep grabbing few parcels of lands from Palestine.

I think there are two steps here: converting video to sensor data input, and using that sensor data to drive. Only the second step will be handled by cars on road, first one is purely for training.

The article strictly talks about people who were pals with him _after_ his Pedophilia conviction. And please don't do this strawman "evil person eating babies", nobody sane is claiming that.

I work with GPUs and I'm also trying to understand the motivations here.

Side note & a hot take: that sort of abstraction never really existed for GPU and it's going to be even harder now as Nvidia et al races to put more & more specialized hardware bits inside GPUs

To the author (or anyone from vectorware team), can you please give me, admittedly a skeptic, a motivating example of a "GPU-native" application?

That is, where does it truly make a difference to dispatch non-parallel/syscalls etc from GPU to CPU instead of dispatching parallel part of a code from CPU to GPU?

From the "Announcing VectorWare" page:

Even after opting in, the CPU is in control and orchestrates work on the GPU.

Isn't it better to let CPUs be in control and orchestrate things as GPUs have much smaller, dumber cores?

Furthermore, if you look at the software kernels that run on the GPU they are simplistic with low cyclomatic complexity.

Again, there's a obvious reason why people don't put branch-y code on GPU.

Genuinely curious what I'm missing.

For example, NEON ... can hold up to 32 128-bit vectors to perform your operations without having to touch the "slow" memory.

Something I recently learnt: the actual number of physical registers in modern x86 CPUs are significantly larger, even for 512-bit SIMD. Zen 5 CPUs actually have 384 vectors registers, 384*512b = 24KB!

I'm not asking an academic program first published 8 year ago (e3nn) to beat actively developed CuEquivariance library. An academic proposing new algorithms doesn't need to worry too much about performance. But any new work which focuses on performance, that includes this blog and a huge number of academic papers published every year, should absolutely use latest vendor libraries as baseline.

I should note PETSc is a big piece of software that does a lot of things. It also wraps many libraries, and those might ultimately dictate actual performance depending on what you plan on doing.

To be practically useful, we don't need to beat vendors, just getting close would be enough, by the virtue of being open-source (and often portable). But I found, as an example, PETSc to be ~10x slower than MKL on CPU and CUDA on GPU; It still doesn't have native shared memory parallelism support on CPU etc.

Hi, I just wanted to note that e3nn is more of an academic software that's a bit high-level by design. A better baseline for comparison would be Nvidia's cuEquivariance, which does pretty much the same thing as you did- take e3nn and optimize it for GPU.

As a HPC developer, it breaks my heart how worse academic software performance is compared to vendor libraries (from Intel or Nvidia). We need to start aiming much higher.

Hi, I actually mentioned ISPC several times there. And although I strenuously avoided crowning one approach "better" over the other, it is worth pointing out that 1) Many of these benefits of ISPC can be had from explicit SIMD libraries like Google's Highway, and 2) ISPC (or any SIMT model) is a departure from how the underlying hardware works, and as the AI community is discovering with GPU, this abstraction can sometimes be lot more headache than its worth.

No. Assuming `k` is small enough, which in practice often is, the arithmetic intensity of this kernel is 25-90 Flops/Byte, way above the roofline knee of any modern CPU.

Hi, thanks for reading.

Re (b) I'm curious what that middle ground is. Is there any simple refactor to help GCC to get rid of this `if`? (Note, ISPC did fine here)

(c) Just to be clear, all the codes in benchmark figures (baseline and SIMD) were compiled with fast-math flags.

Regarding (a), one of the points I wanted to get across was that it didn't feel that complicated to program in the end as I had thought. Porting to AVX-512 felt mechanical (hence the success of LLMs in one-shotting the whole thing).

This is a subjective opinion, depends on programmer's experience etc- so I won't dwell on it. I just wish more CPU programmers gave it a try.

Yeah N is big enough that entire data isn't in the cache, but the memory access pattern here is the next best thing: totally linear, predictable access. I remember seeing around 94%+ L1d cache hit rate.

For SIMD at least, the {mins[3], maxs[3]} representation aligns more naturally with actual instructions on x86. To compute a new bounding box:

new_box.mins = _mm_min_ps(a.mins[3], b.mins[3]);

If you are from ML/Data science world, the analogy that finally unlocked FFT for me is feature size reduction using Principal Component Analysis. In both cases, you project data to a new "better" co-ordinate system ("time to frequency domain"), filter out the basis vectors that have low variance ("ignore high-frequency waves"), and project data back to real space from those truncated dimension ("Ifft: inverse transform to time domain").

Of course some differences exist (e.g. basis vectors are fixed in FFT, unlike PCA).

I'd love to see a breakdown of what exactly worked here, or better yet, PR to upstream Abseil that implements those ideas.

AI is always good at going from 0 to 80%, it's the last 20% it struggles with. It'd be interesting to see a claude-written code making its way to a well-established library.

I was surprised by that defiant tone there in an official page. But it's missing actual numbers, which makes it all pretty strange.

I'm curious what you did with the "active sorting range" after a push/pop event. Since it's a vector underneath, I don't see any option other than to sort the entire range after each event, O(N). This would surely destroy performance, right?

To Koreans, they looked more like sauce bowls, leading them to conclude that the Japanese had starved themselves to stretch out the siege.

As a Bengali man, that's exactly how I felt when I came to USA and first visited japanese restaurants. Part of the reason we consume so much rice is that rice is kind of the main dish (not a side)- it literally takes up central and most of the space in your food plate.

https://commons.wikimedia.org/wiki/File:%E0%A6%87%E0%A6%B2%E...

Could you please elaborate on your example? Thanks.

One feedback from someone interested in using this about the examples: I have looked at several and they seem too high level to get a sense of the actual API (i.e. the expected benefit of using this library vs the development complexity of using it).

For example, the cloth bending simulation is almost entirely: at __init__, call a function to add a cloth mesh to model builder obj, pass built model to initializer of a solver class; and at each timestep: call a collide model function, then call another function called solver.step. That's really it.

That was government of qatar, this is Abu Dhabi (UAE). They had a diplomatic crisis, with full-scale blockade not so long ago.

I’m a big, big fan of acquired podcast, listened to every single of their episodes.

But I remember their entire Microsoft episodes felt like a lengthy defense of Steve Ballmer. There were too many instances of “here’s why this bad decision of Steve made sense given the circumstances” or “ here is how people underestimate the contribution of Steve on this good decision.” They were all well argued points, of course, but so numerous that I found myself wondering if the hosts does not have a relationship with Steve.

The existence of this interview does not help with that suspicion.

People don’t understand that other countries (primary suppliers of stem graduate students) do have lots of research positions, it’s just they don’t usually get first rate talent because USA is far more attractive for those people. Now they will

I don't believe you really can in GPU world. With CPU, if you ignore something important like cache hierarchy, the performance penalty is likely to be in double digits percentage. Something people can and do often ignore. With GPU, there are many many things (memory coalescing, warp, SRAM) that can have triple digits % of impact, hell maybe even more than that.

In one view, the fact that it's a software problem is actually a weakness of (GPU) hardware design.

In the olden, serial computing days, our algorithms were standard, and CPU designers did all sorts of behind-the-scene tricks to improve performance without burdening software developers. It wasn't perfect abstraction, but they tried. Algorithm led the way; hardware had to follow.

CUDA threw that all away, exposed lots of ugly details of GPU hardware design that developers _had to_ take into account. This is why, for a long time, CUDA's primary customers (HPC community & Natl labs) refused to adopt CUDA.

It's interesting that now that CUDA has become a legitimate, widely adopted computing paradigm, how much our view on this has shifted.