IBM Research's Vela Is an AI Supercomputer in the Cloud | Hacker News Reader

IBM Research's Vela Is an AI Supercomputer in the Cloud

research.ibm.com

40 pointsrbanffy44 comments

All of this is just this month's iteration of large corporations using trendy buzzwords to stay relevant.

Anyone remember "IBM Blockchain Platform: Hyperledger Fabric Support Edition"? Or "Watson Assistant"? Or "IBM Quantum Systems"?

> Or "Watson Assistant"?

And then they started calling everything "Watson"? https://www.ibm.com/watson/products-services

Ì believe Watson was their umbrella for anything ML, AI, and NLP.

Watson Assistant has been around for a while now, at least since 2016 when I first used it for a project, and it's been used extensively by that company and its customers since. So maybe not quite fair to include it with the others.

> Or „IBM Quantum Systems“?

Hey hey now. Quantum Systems is actually a free to use (up to certain qubits) cloud platform which you can use qiskit, their open source python lib, to interact with. Its super cool to use it to get some insights into quantum computing and actually evaluate stuff.

So Id take that out of the equation if I were you

> super cool

Every quantum computer is.

Speaking of Watson, how well does ChatGPT perform at Jeopardy?

I gave it a whirl some time ago. The answer is: extremely well. Blog post incoming

On a personal note I was involved with a tiny NLP company that won a contract against IBM Watson around a decade ago just showing a proof of concept that worked using known NLP techniques.

The contract was not large but showed how easy was beating IBM in that technical side, even when you compete with a sales machine.

Here's a key quote: "We’ve recently been asking ourselves: what system would we design if we were exclusively focused on large-scale AI?"

Because, the existing players aren't thinking about scaling, right? OpenAI just runs its 700 gigabyte GPT model in a low-scale, cost-inefficient way?

Kinda sounds like what you'd try to market to clueless executives who are expected to follow the trend of AI. Just like Watson before.

Business is different from technology.

This is a business post masquerading as a post about technical innovation.

My take home from this is that IBM is probably going to have a competitive AI cloud offering.

> IBM is probably going to have a competitive AI cloud offering

Colour me sceptical.

"IBM is probably going to try to compete with an AI cloud offering"

me too. Literally, IBM is a me too company. They haven't innovated in years. Their marketing department is probably larger than their engineering department.

I actually interviewed at IBM 15 or so years ago. They had me attend this seminar for an hour before the interview. They spent the whole seminar bragging about IBM's past, including how the IBM brand was #2 in the world (after Coca-Cola).

Seriously. They didn't talk about tech, their future, or any cool problems they were working on. Just their brand. Ironically, their brand keeps becoming less and less significant each year.

> They haven't innovated in years

Erm.. They do quite a lot R&D and are still a quite impressive basic research powerhouse.

The cache architecture of the new z-series CPU is absolutely awesome, as well as their work in disaggregation of processing and memory on their POWER architecture. Also, their quantum computers are on the forefront of the field (even though we don't quite know what we can do with them).

I wish their cloud offerings had things like POWER and s390x Linux down to tiny machines and that they had a free tier (because both POWER and Z can be segmented into vanishingly thin slices of a computer)

The world is the process of moving from x86 to Arm. This is a huge opportunity for lots of tech companies. So what does IBM do? Innovate on POWER and s390x. Why? Because they need to avoid competition, because they'll lose.

Am I the only one who can't believe a word from IBM after Watson fiasco?

Came here to say this. Could be legit for all I know, but I’m not wasting my time clicking through.

This is my immediate thought when I saw “IBM.”

If they have a real product they should brand it separately. If they’re just trying to sell more snake oil to big corporate fools, then they’re probably on target.

This is their business model and their marketing targets exactly the group of people who are clueless enough to continue paying them. Why risk selling to someone who has a nose for bullshit by doing anything that implies you have a real product?

IBM has demonstrated a voice system that can perform as a debater in a lawmaking context. If it is not marketing smoke and mirrors for the C-Suite business lunch or breakfast let's see more of that in ops.

Can you please explain what the fiasco about Watson is? I know Watson, but not the fiasco part.

https://news.ycombinator.com/item?id=11751267 https://news.ycombinator.com/item?id=14979642 https://news.ycombinator.com/item?id=30046432

Watson won jeopardy by being an expert system with hand coded knowledge (their NLP parser was written in Prolog) and sophisticated QA text search. It wasn't a pure unsupervised/semi supervised deep learning model so it didn't generalize well to other domains like medicine despite it being hyped as a ML trailblazer.

This is complete ignorance of history.

Watson (the quiz winner) was in 2011 before the deep learning revolution (which kicked off in 2014 with AlexNet for image recognition). It was probably the pinnacle of old-style natural language processing.

The success of that led to a set of basically unrelated tools being released under the IBM Watson brand which were mostly failures.

It seems the market for highly customised quiz-winning applications was much smaller than expected.

The market for news about quiz-winning applications was much larger than I expected though!

Whose to say which is more important.

It remains to be seen how large is the market for search that can subtly and confidently lie to you.

ChatGPT could certainly use the ability to come up with correct answers at some point.

Some other comments here are good but I'll point out that what most people don't understand is that "Watson" is the name of a marketing umbrella, not a technology. Essentially, almost anything at IBM that had AI/ML smell to it got tossed into the "Watson" market bucket. This allowed IBM to push a huge set of technologies that they could then customize and build services around, for a large fee of course.

The problem is that most of the AI/ML technologies IBM used were stock, off the shelf, often not even state of the art. You could literally go to a github page or use scikit_learn out of the box to get equivalent or better performing models without being tied to IBM's proprietary, consultant heavy, solutions.

The marketing of Watson, tied to the Jeopardy performance, made people think it was the first coming of AGI -- a kind of Skynet moment. It was really just a cobbled together basket of unintegrated, nothing special, industry bog standard AI/ML stuff.

Watson as a system was designed specifically to play Jeopardy. After the win it got a lot of global hype behind it and IBM's marketing and sales teams went into overdrive positioning it as some next generation intelligent AI that could help with generic business processes, analytics, medical diagnosis, legal briefs, customer service and every other business area under the sun. Instead, companies that bought into it got the same army of IBM offshore consultants to custom build half-assed solutions for them like before.

"Watson" is ultimately a marketing name for a sub-par IT consulting service, nothing more.

Took the words right out of my mouth

I would add the 'IBM Blockchain Platform: Hyperledger Fabric' as a second point.

How the mighty has fallen.

IBM Research used to be the best research group in the world. Now they are implementing a fairly reasonable NVIDIA cluster and the only "research publication" is a video talk given at NVIDIA's conference.

Compare that to the linked Microsoft Open AI Supercomputer, which talks about DeepSpeed (really significant breakthrough in distributed training), the ONNX runtime (which allows the same models to be run on things other than NVIDIA GPUs) etc etc.

The M in IBM standing for _Machines_, which have been fully commoditized and virtualized with the evolution of the Internet. IBM's biggest software advancements were proprietary and specific to their platforms (OS/360, etc.). Mainframes were called "big iron"; you could feel the mass of the computing engines when touching the beige powder-coated plates of their cabinets. The punch cards were never very far off, it was always just a faster version of the original mechanical processor.

Any smaller computer they made was always just to plug a market breach created by some young wazoos. IBM PC responding to Apple II. RS/6000 responding to Sun. None of these were truly "serious" machines in the eyes of Real Men in Suits. They knew that when the time came to handle big business, you needed big beige boxes sitting on raised floors in an atmosphere-controlled room that you could watch over through a window, their blinkenlights softly atoning for whatever sins The Man had committed in the struggle for civilisation.

Not really sure what this is supposed to be about, but I was talking about IBM Research[1].

As Wikipedia notes, 6 Nobel prizes and 6 Turing awards is just the start. This has very little to do with OS/360 or anything - instead they were inventing scanning tunneling microscope[2] or Fortran[3].

I'm sure I'd be able to give a better list, but the Watson research center website[4] is down for me....

[1] https://en.wikipedia.org/wiki/IBM_Research

[2] https://en.wikipedia.org/wiki/Scanning_tunneling_microscope

[3] https://en.wikipedia.org/wiki/Fortran

[4] https://researcher.watson.ibm.com/researcher/view_page.php?i...

> big beige boxes sitting on raised floors in an atmosphere-controlled room that you could watch over through a window, their blinkenlights softly atoning for whatever sins The Man had committed in the struggle for civilisation.

Since then they turned black, most likely because of such sins...

"virtualization overhead is less than 5%, which is the lowest overhead in the industry that we’re aware of"

Is it really lower than AWS bare metal instances? Amazon are fairly reticent on details of their nitro platform, but it's pretty clear they've implemented most of their hypervisor in FPGA/Silicon.

TL;DR: Traditional HPC architecture is already good enough for large-scale ML tasks, but using the more cost-effective, commodity cloud datacenter architecture with some tweaks in virtualization to improve performance we got the solution that works fast enough while also more user-friendly for AI researchers.

P.S. This TL;DR was provided by human agent with no AI-generated parts.

* can't build an Apple Silicon DB2 driver, busy building 'AI in the Cloud'.

I competed against IBM for many contracts over many years as an engineer at a small startup. The experience left me with zero respect for IBM. On occasions when a customer did choose IBM over us for a project, I would genuinely feel bad for the customer.

IBM is all smoke and mirrors; sales, marketing, and predatory business practices once they get their claws in you. Very little deep technical competence at that company.

'AI' now is just a completely meaningless buzzword.

Again. It's funny, I couldn't wait for the next breakthrough in AI and now I can't wait for the next AI winter.

Wait a second! I’ve heard this one before…

Reply on news.ycombinator.com