Tbh, Modular getting acquired happened sooner than I would have expected, if ever. Don't know how to feel about this one.
Also so many mixed feelings about Mojo, the programming language powering Modular. Of course Chris Lattner is free to pursue whatever he wants, his many contributions to tech will always be highly regarded, but to me it feels as if he "wasted" lots of his precious mental capacity on making Mojo a python-like language instead of trying to come up with something better from first principles. I know, the promise of Mojo eventually being a Python superset has been taken back, which I think is the right move, and I understand why Mojo's initial motivation for being close to Python was to attract ML folks, but I'm getting counterfactual regret just by thinking about what Chris Lattner could have achieved by making a new programming language truly from scratch and not letting some undesireable pythonisms muddy the language.
Anyway, sorry for rambling. Congrats to the team at Modular!
I'm actually mostly worried about the future of Mojo at this time.
Though hopefully it will be fully released open source still, but I feel there are question marks around whether it will be a priority to continue to develop by Qualcomm, or if they are mainly interested in the AI compute stack?
Time will tell I guess, but a lot feels to be up in the air.
Maybe Chris was a little unhappy about where Mojo ended up, and sees this as an opportunity to start anew on a properly designed language from scratch :D
No, this was pure speculation based on what seems like a popular view on where Mojo ended up, where the initial Python-focus don't seem to help it that much anymore.
But they changed their goal from being a python superset to pythonish language with great python interoperability. The only other thing they could've done differently is making the language not look like python superficially.
I think chris achieved his goal of creating a language which takes full advantage of MLIR and also not repeating some of the mistakes made with swift's development.
"first principles" and "from scratch" are predictable failure modes... he had very good reason to pursue a Python-like language given the circumstances and objectives
I think I get what you mean and I should have been more precise in my wording. I didn't mean that an alien language that looks nothing like we have ever seen but for the sake of doing it "right" from scratch would have been a good idea. A new programming language definitely should steal the ideas of other languages that turned out to be good. But Mojo also adopted some of the arguably bad ideas from Python just because there was too much design pressure to appeal to Python programmers. I wonder what Mojo could have looked like without this particular pressure. Basically, with what kind of programming language would a person with as much experience and good taste as Chris Lattner come up with if there were no such external pressures?
Modular now joins SYCL, OpenCL, and One API on the list of cross platform languages which never really became cross platform.
After so long and so much investment in AI, the best cross-platorm API we've got for high performance Kernels is vulcan, a graphics API. That is sad.
Still, this is pretty good for Modular's employees, probably good for Qualcomm. It's just terribly disappointing for anyone who invested time learning mojo in the hope it might actually become cross platform.
Yesterday, LineShine a supercomputer in China emerges as #1 in the Top500 using ARM v9 based chips and no GPUs. Today, Qualcomm a premier designer of ARMv9 licensed chips in the United States acquires Modular, who has been creating a compiler stack that provides an alternative to NVIDIA's CUDA stack.
Are you ready for Qualcomm ARMv9 powered inference running Mojo/MAX written kernels doing low-cost inference at scale for AI?
The reason to move away from arm has nothing to do with performance, but rather avoiding licensing snafus like happened with their laptop chips. So far no one has delivered a risc-v core with class leading performance outside of the really low end. Not saying it can’t be done, but it will likely be a step back at first.
It's kind of funny that Modular is getting acquired by a hardware company considering what it's founder has said repeatedly in interviews and articles about how those companies fail to make AI stacks.
Qualcomm has acquired excellent engineering talent here, the infrastructure I've seen Modular build in the 3 years I've followed the company is insane.
> I don't get it.
>
> Qualcomm has almost no products in the high-end inference/training market. The industry standard is the NVIDIA Hopper H100/H200.
>
> What could they possibly get from acquiring Modular?
Don't ask what they will gain from owning it, ask what they will gain from others not owning it...
It's now focusing on inferencing, both for data centers and edge. They already have an older AI100 NPU card and have other products in the pipeline including server class CPU that they are targeting for "Agentic" applications.
It's interesting that acquire.fyi data shows tech M&A deal volume is down 11% year to date, but total deal value is up 40%. So, fewer deals are closing in tech, but the deals that are closing are much larger. I wish we had the deal value for this one.
I honestly think Mojo would be better served if it is just a high-level language for GPU programming that compiles down to PTX with clear Python/Rust interop boundaries instead of trying for the "one language, multiple computational model" thing that they seem to be going for. The programming model between CPU and GPU programming is very different: code that runs best on CPU with heavy branching behaviors should not be written the same way as massively parallel matrix multiplication oriented GPU code, which I think they will be forced to do in the MLIR level anyway.
So, you end up with a language that looks like Python, but doesn't behave like Python, and companies that adopt Mojo early with the promise of Python compatibility may find themselves running into edge cases with difficult to trace compiler error messages that would be nearly impossible to debug, especially with the addition of Zig style `comptime` as their metaprogramming model.
Yes, ARM sued Qualcomm, Qualcomm won, and separately Nuvia has shipped, 2, 3? times now? I don't know how it's a failure or if the delay were "huge" and "dragged out". It's not like it launched an old product or took years and years and years. 39 months between acquisition and Snapdragon X Elite being available for purchase.
> You’re equivalent of saying the Intel delays were a success too.
If Trump nuked TSMC's production lines the day before M1 went to production, and the production lines came back 3 years later, would Apple ship the M1 on it? Or, the M3?
As you point out, it makes 0 sense to ship the M1.
If it makes 0 sense, why project that idea onto me?
When faced with a contradiction, first, check your premises. (and read your interlocutor's, "It's not like it launched an old product" obviates your claim that I'd also applaud Intel's delays)
Nvidia wasn't going to buy them. Unless Mojo intended to compete toe-to-toe in the hardware space, they were destined to get bought out by a hardware underdog at some point or another.
This is where an industry-spanning consortium would have helped out, but Mojo never really built those inroads with the hardware space. They just expected everyone else to opt-in to their mercurial middleware, which is a fundamental misunderstanding of how and why CUDA is successful.
Also so many mixed feelings about Mojo, the programming language powering Modular. Of course Chris Lattner is free to pursue whatever he wants, his many contributions to tech will always be highly regarded, but to me it feels as if he "wasted" lots of his precious mental capacity on making Mojo a python-like language instead of trying to come up with something better from first principles. I know, the promise of Mojo eventually being a Python superset has been taken back, which I think is the right move, and I understand why Mojo's initial motivation for being close to Python was to attract ML folks, but I'm getting counterfactual regret just by thinking about what Chris Lattner could have achieved by making a new programming language truly from scratch and not letting some undesireable pythonisms muddy the language.
Anyway, sorry for rambling. Congrats to the team at Modular!
Though hopefully it will be fully released open source still, but I feel there are question marks around whether it will be a priority to continue to develop by Qualcomm, or if they are mainly interested in the AI compute stack?
Time will tell I guess, but a lot feels to be up in the air.
To each their own!
After so long and so much investment in AI, the best cross-platorm API we've got for high performance Kernels is vulcan, a graphics API. That is sad.
Still, this is pretty good for Modular's employees, probably good for Qualcomm. It's just terribly disappointing for anyone who invested time learning mojo in the hope it might actually become cross platform.
Are you ready for Qualcomm ARMv9 powered inference running Mojo/MAX written kernels doing low-cost inference at scale for AI?
1. Moving beyond ARM to RISC-V
2. Being competitive for AI/could needs instai of just chips for phones and other edge devices.
Interesting to see bold and high-conviction moves in this direction. Tenstorrent, Modular, Ventana, Alphawave, etc.
The reason to move away from arm has nothing to do with performance, but rather avoiding licensing snafus like happened with their laptop chips. So far no one has delivered a risc-v core with class leading performance outside of the really low end. Not saying it can’t be done, but it will likely be a step back at first.
* https://www.modular.com/blog/democratizing-ai-compute-part-9...
Qualcomm has almost no products in the high-end inference/training market. The industry standard is the NVIDIA Hopper H100/H200.
What could they possibly get from acquiring Modular?
There's actually a lot of ML deployed on phones. Both Google's and Apple's photo software uses it heavily for example.
> The industry standard is the NVIDIA Hopper H100/H200.
B200/B300/GB300 actually...
Don't ask what they will gain from owning it, ask what they will gain from others not owning it...
It's now focusing on inferencing, both for data centers and edge. They already have an older AI100 NPU card and have other products in the pipeline including server class CPU that they are targeting for "Agentic" applications.
You're allowed to get a new job. Qualcomm is allowed to enter new markets.
> They must see value in the technology.
What value? Mojo doesn't currently support any of Qualcomm's GPUs.
So, you end up with a language that looks like Python, but doesn't behave like Python, and companies that adopt Mojo early with the promise of Python compatibility may find themselves running into edge cases with difficult to trace compiler error messages that would be nearly impossible to debug, especially with the addition of Zig style `comptime` as their metaprogramming model.
Not true. Nuvia has had huge delays as part of the acquisition. It resulted in ARM licensing lawsuits and many more and things dragged out.
Yes a world of a difference. That’s competing against an Apple M2 vs M4. You’ve given yourself 2 generations of disadvantage.
You’re equivalent of saying the Intel delays were a success too.
If Trump nuked TSMC's production lines the day before M1 went to production, and the production lines came back 3 years later, would Apple ship the M1 on it? Or, the M3?
As you point out, it makes 0 sense to ship the M1.
If it makes 0 sense, why project that idea onto me?
When faced with a contradiction, first, check your premises. (and read your interlocutor's, "It's not like it launched an old product" obviates your claim that I'd also applaud Intel's delays)
But that's what happened. Go check the benchmarks. Clearly you haven't. That Snapdragon X that got released (1st gen) was way off the mark.
> When faced with a contradiction
When faced with a hallucination...
This is where an industry-spanning consortium would have helped out, but Mojo never really built those inroads with the hardware space. They just expected everyone else to opt-in to their mercurial middleware, which is a fundamental misunderstanding of how and why CUDA is successful.