For the `sum` example, Bend has a huge disadvantage, because it is allocating 2 IC nodes for each numeric operation, while Python is not. This is obviously terribly inefficient. We'll avoid that soon (just like HVM1 did it). It just wasn't implemented in HVM2 yet.
Note most of the work behind Bend went into making the parallel evaluator correct. Running closures and unrestricted recursion on GPUs is extremely hard. We've just finished that part, so, there was basically 0 effort into micro-optimizations. HVM2's codegen is still abysmal. (And I was very clear about it on the docs!)
That said, please try comparing the Bitonic Sort example, where both are doing the same amount of allocations. I think it will give a much fairer idea of how Bend will perform in practice. HVM1 used to be 3x slower than GHC in a single core, which isn't bad. HVM2 should get to that point not far in the future.
Now, I totally acknowledge these "this is still bad but we promise it will get better!!" can be underwhelming, and I understand if you don't believe on my words. But I actually believe that, with the foundation set, these micro optimizations will be the easiest part, and performance will skyrocket from here. In any case, we'll keep working on making it better, and reporting the progress as milestones are reached.
While that's true, Python would be using big integers (PyLongObject) for most of the computations, meaning every number gets allocated on the heap.
If we use a Python implementation that would avoid this, like PyPy or Cython, the results change significantly:
% cat sum.py
def sum(depth, x):
if depth == 0:
return x
else:
fst = sum(depth-1, x*2+0) # adds the fst half
snd = sum(depth-1, x*2+1) # adds the snd half
return fst + snd
if __name__ == '__main__':
print(sum(30, 0))
% time pypy sum.py
576460751766552576
pypy sum.py 4.26s user 0.06s system 96% cpu 4.464 total
[Edit: just noticed the previous comment had already mentioned pypy]
> I'm aware it is 2x slower on non-Apple CPUs.
Do you know why? As far as I can tell, HVM has no aarch64/Apple-specific code. Could it be because Apple Silicon has wider decode blocks?
> can be underwhelming, and I understand if you don't believe on my words
I don't think anyone wants to rain on your parade, but extraordinary claims require extraordinary evidence.
The work you've done in Bend and HVM sounds impressive, but I feel the benchmarks need more evaluation/scrutiny. Since your main competitor would be Mojo and not Python, comparisons to Mojo would be nice as well.
I'm personally putting a LOT of effort to make our claims as accurate and truthful as possible, in every single place. Documentation, website, demos. I spent hours in meetings to make sure everything is correct. Yet, sometimes it feels that no matter how much effort I put, people will just find ways to misinterpret it.
We published the real benchmarks, checked and double checked. And then you complained some benchmarks are not so good. Which we acknowledged, and provided causes, and how we plan to address them. And then you said the benchmarks need more evaluation? How does that make sense in the context of them being underwhelming?
We're not going to compare to Mojo or other languages, specifically because it generates hate.
Our only claim is:
HVM2 is the first version of our Interaction Combinator evaluator that runs with linear speedup on GPUs. Running closures on GPUs required colossal amount of correctness work, and we're reporting this milestone. Moreover, we finally managed to compile a Python-like language to it. That is all that is being claimed, and nothing else. The codegen is still abysmal and single-core performance is bad - that's our next focus. If anything else was claimed, it wasn't us!
People are naturally going to compare the timings and function you cite to what's available to the community right now, though, that's the only way we can picture its performance in real-life tasks.
> Mojo or other languages, specifically because it generates hate
Mojo launched comparing itself to Python and didn't generate much hate, it seems, but I digress
In any case, I hope Bend and HVM can continue to improve even further, it's always nice to see projects like those, specially from another Brazilian
> our code gen is still on its infancy, and is nowhere as mature as SOTA compilers like GCC and GHC
Yet people still misinterpret. It is frustrating because I don't know what I could've done better
FWIW on HN people are inherently going to try to actually use your project and so if it's meant to be (long term) a faster way to run X people evaluate it against that implicit benchmark.
Remember you don't need comment trolls on your team, and you'll go insane taking them seriously. Focus on piquing the interest of motivated language nerds. I personally would have really appreciated a "look, were still 10x (or whatever) slower than Python, so now I need all the help I can get working on the codegen, etc." This would have given me quick perspective on why this milestone is meaningful.
Just a note: we are NOT 10x slower than Python. I think a lot of people got the wrong message from this thread. HVM is actually quite fast already. It is just that, on this specific program, Python was doing no allocations, while HVM was allocating a lot.
If you compare programs that do the same allocation, HVM already outperforms not just Python but even compiled languages like Haskell/GHC, due to using all cores. See the Bitonic Sort example. And that's extremely relevant, because real world programs in high-level languages allocate a lot!
I think I made a huge mistake of using a "sum" example on the GUIDE, since it is actually one of the few specific programs where HVM will do poorly today.
That's where the comparison to Python comes from: getting speedup on slow interpreters is not very _relevant_. Now if your interpreter has the same optimisations as Python (or v8 or JVM), even a small fraction of what you show would be impressive.
Having said this, the work your team did is a really challenging engineering feat (and with lot more potential). But I do not believe the current speedups will hold if the interpreter/compilers have the level of optimisation that exist in other languages. And while you do not claim it, people expect that.
I read the whole thing first, then commented, but people often read half of such a document, assume they've got all the important bits, and dive straight in.
(we used to have that problem at $work with new team members and our onboarding doc; I added a section at the bottom that was pure silliness, and then started asking people who claimed to have read it a question that would only make sense if they'd seen the joke ... generally followed by telling them to go back and finish reading and not to try that with me again ;)
If you add some disclaimer like "Note: Bend is currently focused on correctness and scaling. On an absolute scale it may still be slower than single threaded Python. We plan to improve the absolute performance soon." then you won't see these comments.
Also this defensive tone does not come off well:
> We published the real benchmarks, checked and double checked. And then you complained some benchmarks are not so good. Which we acknowledged, and provided causes, and how we plan to address them. And then you said the benchmarks need more evaluation? How does that make sense in the context of them being underwhelming?
> But keep in mind our code gen is still on its infancy, and is nowhere as mature as SOTA compilers like GCC and GHC.
Second paragraph of Bend's GUIDE.md:
> While cool, Bend is far from perfect. In absolute terms it is still not so fast. Compared to SOTA compilers like GCC or GHC, our code gen is still embarrassingly bad, and there is a lot to improve. That said, it does what it promises: scaling horizontally with cores.
Limitations session on HVM2's paper:
> While HVM2 achieves near-linear speedup, its compiler is still extremely immature, and not nearly as fast as state-of-art alternatives like GCC of GHC. In single-thread CPU evaluation, HVM2, is still about 5x slower than GHC, and this number can grow to 100x on programs that involve loops and mutable arrays, since HVM2 doesn’t feature these yet.
Yeah exactly. I read most of the readme and watched the demo, but I'm not interested in installing it so I missed this. I would recommend moving this to the first section in its own paragraph.
I understand you might not want to focus on this but it's important information and not a bad thing at all.
We'll add the disclaimer before the install instructions instead!
I think you can avoid this kind of criticism by setting expectations better - just plastering a banner at the top saying that it’s in early stage development and not optimized, but that the future is bright, for example. The current headline saying it’s the “parallel future of computation” isn’t really enough to make people understand that the future isn’t here yet.
Same goes for the README, the fact that it’s not production ready per-se really ought to be at the top to set people’s expectations properly IMO, since a lot of people will not read the whole wall of text and just jump straight into trying it out once they’re on your GitHub page.
They’re critical since they are led to have much higher expectations than what actually exists today.
That said, this is a cool project and wish you the best in making it really good!
Based on how you've made a nice marketing page and README that sounds like you want people to actually use this tool in practice, within that context correctness is a minimum requirement/table stakes for a language to be usable at all, but that alone doesn't make it "production ready" if it fails to practically achieve anything you'd realistically want to do with it better than old-school languages that people already know how to use.
I am not a Python dev, but it seems that CPython's goal is not to be as fast as C, but just that it is a default runtime for Python [1] and the fact that C is in its name is just an implementation detail. Very different.
So the criticism leveled at the project appears to be valid.
And that's comparing a first-version interpreter against a SOTA runtime deployed in all browsers around the world and optimized by all major companies over 20+ years. If that's not useful to you, that's useful to me, which is why I wanted to share so it can be useful to more people.
The other link on the front page says:
"Welcome to the Parallel Future of Computation"
from reply below:
> I apologize if I got defensive, it is just that I put so much effort on being truthful, double-checking, putting disclaimers everywhere about every possible misinterpretation.
I just want to say: don't stop. There will always be some people who don't notice or acknowledge the effort to be precise and truthful. But others will. For me, this attitude elevates the project to something I will be watching.
Thank you. I understand in such an early irritation of a language there are going to be lots of bugs.
This seems like a very, very cool project and I really hope it or something like it is successful at making utilizing the GPU less cumbersome.
And you're just not fast enough for anyone doing HPC, where the problem is not identifying what can be parallelized, but figuring out to make the most of the hardware, i.e. the codegen.
> It is very important to reinforce that, while Bend does what it was built to (i.e., scale in performance with cores, up to 10000+ concurrent threads), its single-core performance is still extremely sub-par. This is the first version of the system, and we haven't put much effort into a proper compiler yet. You can expect the raw performance to substantially improve on every release, as we work towards a proper codegen (including a constellation of missing optimizations).
which seems to be pretty much exactly that?
It's at the bottom, though, so I can imagine people just skimming for "how do I get started" missing it, and making it more obvious would almost certainly be a Good Thing.
I still feel like reading the whole (not particularly long) README before commenting being angry about it (not you) is something one could reasonably think the HN commentariat would be capable of (if you want to comment -without- reading the fine article, there's slashdot for that ;), but I'm also the sort of person who reads a whole man page when encountering a new command so perhaps I'm typical minding there.
In my experience, to not be misunderstood it is more important to understand the state of mind/frame of reference of your audience, than to be utterly precise.
The problem is, if you have been working on something for a while, it is extremely hard to understand how the world looks to someone who has never seen it (1).
The second problem is that when you hit a site like hacker News your audience is impossibly diverse, and there isn't any one state of mind.
When I present research, it always takes many iterations of reflecting on both points to get to a good place.
I'm not informed enough to comment on the performance but I really like this attitude of not overselling your product but still claiming that you reached a milestone. That's a fine balance to strike and some people will misunderstand because we just do not assume that much nuance – and especially not truth – from marketing statements.
that sum example, transpiled to C and compiled takes 1m12.704s, so it looks like it's just the VM case that is having serious issues of some description!
Clearly this language is positioned at using the gpu for compute-heavy applications and it's still in its early stages. Recursion is not the target application and should not be a relevant benchmark.
The authors made obviously false claims. That this is an extremely fast implementation. When it's not. That's really poor form.
The authors never made any kind of false claims at all. You're reading a lot in to both their README and my post.
They've updated the README for a bit of clarity, but even re-reading the README as it was when I looked this morning (and even a few from before) it hasn't claimed to be fast. The claims are all related to the features that it does have, around parallelisation.
That's not how complexity theory works at all. If your constants are massive you can make anything look like it scales well.
Imagine a compiler that adds in a very large sleep that's inversely proportional to the number of cores. There you go. My compiler now emits code that scales linearly. Of course it's very slow. Where do I pick up my Turing award?
I never thought that CS education was this bad.
The readme is also blatantly lies: "It is not the kind of algorithm you'd expect to run fast on GPUs." Bitonic sort on GPUs is a thing and it performs well.
The gaslighting is just amazing here.
CPU, Apple M3 Max, 1 thread: 12.15 seconds
CPU, Apple M3 Max, 16 threads: 0.96 seconds
GPU, NVIDIA RTX 4090, 16k threads: 0.21 seconds
The README mentions "fast" in 2 places, none of which is comparing to other languages.
The actually interesting claim is that somebody has found a practical use for Interaction Combinators as a computing model for putting GPUs to work.
"Hey, I'm accessing the 0th element here, just want to make that clear"
Without the +0, that statement looks disconnected from the +1 even though conceptually its the same.
Say somebody adds some special marker/tombstone/whatever into element 0 and now all those additions need to be bumped up by one. Someone else may go and see the +1, +2, +3 and just change them to +2, +3 +4, etc while completely missing the lone variable by itself as its visually dissimilar.
Ive usually seen it used in longer lists of statements. It also keeps everything lined up formatting wise.
A Pythonic implementation would use loops and mutation.
I appreciate this is early days, but it's hard to get excited about what seems to be incredibly slow performance from a really simple example you give. If the simple stuff is slow, what does that mean for the complicated stuff?
If I get a chance tonight, I'll re-run it with `-s` argument, see if I get anything helpful.