> We all know the good practices. Don't be a dumbass
In theory yes, but in practice we are all dumbasses to some extent.
I used to have your attitude until I saw a friend die of a heart attack at an early age - and it appeared to me that he would have survived if he had an indication. So, now I have changed my attitude to one of more data does not hurt.
I dont get the point of helion as compared to its alternatives like gluon.
For best performance I would presume one needs low-level access to hardware knobs. And, these kernel primitives are written one-time and reused. So, what is the point of a DSL that dumbs things down as a wrapper around triton.
Funny, I feel the same way about Triton. Performant Triton looks like CUDA (but with tiles!) except it's ten times harder to debug since it doesn't have the tooling NVIDIA provides.
If I had to run on AMD I'd rather deal with their hipify tooling.
Performant Triton programs are usually simpler and shorter than their CUDA equivalents. This alone makes it easier to write, and I would argue that it helps with debugging too because the model provides a lot more guarantees on how your code executes. That said, some of the tooling is notably poor (such as cuda-gdb support).
Agree on shorter, disagree on simpler. The hard part of understanding GPU code is knowing the reasons why algorithms are the way they are. For example, why we do a split-k decomposition when doing a matrix multiplication, or why are we loading this particular data into shared memory at this particular time, with some overlapping subset into registers.
Getting rid of the for loop over an array index doesn't make it easier to understand the hard parts. Losing the developer perf and debug tooling is absolutely not worth the tradeoff.
For me I'd rather deal with Jax or Numba, and if that still wasn't enough, I would jump straight to CUDA.
It's possible I'm an old fogey with bias, though. It's true that I've spent a lot more time with CUDA than with the new DSLs on the block.
I don’t think it is possible to write high performance code without understanding how the hardware works. I just think staring at code that coalesces your loads or swizzles your layouts for the hundredth time is a waste of screen space, though. Just let the compiler do it and when it gets it wrong then you can bust out the explicit code you were going to write in CUDA, anyway.
What's the point of Triton compared to Gluon? What's the point of PyTorch compared to Triton?
One of the main values of Triton is that it significantly expanded the scope of folks who can write kernels - I think Helion could expand the scope even more.
The negative drumbeat has started, and the narrative is changing to become critical. Every new fad reaches a saturation point, and people's attention spans are short. Already, I see folks piling onto the quantum bandwagon as if LLMs are passe.
Here, the clouds have pulled a trick to inflate their revenues with their own cashflows, and have not been punished yet for it by shareholders - except meta which is getting asked some difficult questions.
These deals happen all the time. The case for a bubble is the following.
When Microsoft offers cloud-credits in exchange for openai equity, what it has effectively done is to purchase its own azure revenues. ie, a company uses its own cash to purchase its own revenues. This produces an illusion of revenue growth which is not economically sustainable. This is happening for all clouds right now wherein their revenues are inflated by uneconomic ai purchases. This is also happening for the gpu chip vendors as well, wherein they are offering cash or warrants to fund their own chip sales.
But nobody is falling for the "illusion of revenue growth". This is out in the open. This isn't a scam. Investors know this and are pricing accordingly. They see the revenue growth but also see the decrease in cash.
What Microsoft is actually doing is taking the large profits it would have otherwise made on its cloud compute with retail customers, losing much/all of those profits as it sells the compute more cheaply to OpenAI, and converting those lost profits into ownership of OpenAI because Microsoft's goal is to own more of OpenAI.
There is nothing "bubble" about this. Microsoft isn't some opaque startup investors don't understand. All of this is incredibly transparent.
There will be increased transparency since microsoft will now have to report on the performance of its openai equity [1]. The concern is that while chatgpt is a great app, the economic benefits of the current investments are being questioned. There is starting to be skepticism of ai as the public starts to get jaded. This happens in all fads. That explains why the media is buzzing with articles like these which are becoming increasingly critical while earlier they were all aboard the ai-train.
my prediction is that the ai-bubble pops no later than when openai files for its ipo. openai has an incentive to keep its finances outside of the prying eyes of the public shareholders. This way their mystique is kept in tact, and nobody can question their capex and datacenter investment claims. But, it is becoming difficult since msft now reports on their openai holdings.
I am surprised and glad that journalists are reporting on this. OpenAI is the hinge in the whole AI bubble, and it has every incentive to keep its financials private if it is not very flattering. So, thats the place to look for funny business - specially given their outlandish and provocative announcements of 16GW of datacenter buildouts which defies economic sense and demands more scrutiny.
In theory yes, but in practice we are all dumbasses to some extent.
I used to have your attitude until I saw a friend die of a heart attack at an early age - and it appeared to me that he would have survived if he had an indication. So, now I have changed my attitude to one of more data does not hurt.
reply