Hacker News Reader

Comment by Kon-Peki | Hacker News Reader

Energy efficiency and improved packaging are things I can readily agree with. The last thing though - “hybrid algorithms leveraging AI efficiency” sounds an awful lot like a buzzword sales pitch.

This article reminds of this other one [1] posted about 1 month ago.

It’s an interview with some guys that just got done building an exascale supercomputer, in which it was originally estimated to need 1000 megawatts but ultimately only needs 60. The reporter asks about zettascale and the power requirements; they wave it off and say that the big question about whether it will even be possible in the next 10 years is getting the chip lithography small enough so that you can physically build a working zettascale supercomputer.

[1] https://news.ycombinator.com/item?id=34604319

View on HN

Re. The “hybrid algorithms” bit: I was at this talk. The example she gave was a physics sim like CFD, iterating between a fast/approximate ML-based algorithm and slow/accurate classical physics algorithm, with the output of each feeding in as the starting point of the next round. But this was just an example, clearly there are lots of area you could apply a similar approach.

The main thing AMD has in their accelerators that enables this this is unified memory between CPU and GPU. Thats really interesting.

This has been something I've been incredibly pleased with on the apple silicon SOC's. Albeit slowly, being able to load large datasets or blender scenes on a portable, efficient laptop and still being able to use the GPU is a nice touch.

Of course performance wise it doesn't touch the $1k+ graphics cards with crazy amounts of ram, but for students and if I need to do something quick on the go, its a really useful tool.

Don't forget that AMD also bought Xilinx

@WithinReason

I expect them to use Xilinix's AI engines primarily in their CPUs, APUs and GPUs - not so much FPGA.

I expect Xilinx's AI engines to never be integrated into anything AMD. Because Xilinx AI engines are VLIW - SIMD machines running their own instruction set.

----------

AMD is doing the right thing with Xilinx tech: they're integrating it into ROCm, so that Xilinx AI engines / FPGAs can interact with CPUs and GPUs. But there's no reason why these "internal core bits" should be shared between CPU, GPU, and FPGA.

How do you expect FPGAs to be useful here?

I think its also leaning into their new product MI300, which has 24 cpu cores with 8 compute dies. Both CPU and GPU (and memory) on a single package.

Conventional processing + AI together. Hybrid approaches.

Afaik, having worked in HPC, one area where this can be employed is error bias correction of CFD models. E.g. weather models have various biases that need to be corrected for - so far this is just done with some relatively simple statistics afaik.

Yes they were really crowing about MI250X adoption and touting the upcoming MI300

So differentiable programming with nns as approximators? Cool!

I imagine something like AlphaZero would also benefit: It's basically a hybrid of tree search and neural network.

“hybrid algorithms leveraging AI efficiency”: An example I came across recently:

https://arxiv.org/abs/2202.11214

A neural network-based solution for weather forecasting: "FourCastNet is about 45,000 times faster than traditional NWP models on a node-hour basis."

I assume it means using ML to choose between multiple algorithms instead of more traditional heuristics.

That or the algorithms use a combination battery and gasoline.

AMD already uses a Perceptron for their branch prediction. I would say they're talking about support for ML speed ups in hardware but maybe the plans also include a more complex complete neural net in hardware for data prediction.

What are hybrid algorithms supposed to be anyways? Half algorithm & half dictionary?

Half algorithm & half user manual?

Half algorithm & half class?

Generally the concept of hybrid algorithms is:

- Half "hard maths" algorithms. i.e. cominbatorics, geometry, etc.

- Half "fuzzy maths" algorithms. i.e. heuristics, approximation, machine learning.

The idea being to solve the parts that can be easily solved by hard maths with those hard maths so that you can reduce the problem space for when you apply the fuzzy maths to solve the rest of the problem.

In other words, it's taking the problem, breaking out discrete pieces to solve with well established hard maths, using heuristics & numerical solutions to tackle the remaining known problems without "easy" analytical solutions, then using ML to fill in all the gaps and glue the whole thing together.

Half conventional HPC simulation that runs best on CPU, half Neural Networks that need GPUs. As proposed for example here

https://www.nature.com/articles/s41586-019-0912-1

and here

https://www.nature.com/articles/s42256-021-00374-3

> The next step will be a hybrid modelling approach, coupling physical process models with the versatility of data-driven machine learning.

The Frontier HPC system that AMD just delivered is aimed fully at that problem.

https://en.wikipedia.org/wiki/Frontier_(supercomputer)

The article gives a concrete example in the same paragraph: "For example, AI algorithms could get close to a solution quickly and efficiently, and then the gap between the AI answer and the true solution can be filled by high-precision computing."

Interestingly the example is backwards (statistical reasoning first, hard reasoning second) compared to traditional usage of "hybrid" in AI and control contexts.

That's not concrete at all.

From the article it looks like half AI guessing the solution, half some static algorithm fixing the result to be better. Not sure how it's supposed to really work.

one variant of this is sciml approaches where you use an ODE solver wrapped around a NN. the ODE solver guarantees you get the right conversation laws which NNs don't do well and the NN is more accurate than the hand written model since it doesn't ignore the higher order effects.

Half SQL statements and half array.sort() methods I assume.

    (solve (this :by strong-ai) (that :by weak-ai))

> The last thing though - “hybrid algorithms leveraging AI efficiency” sounds an awful lot like a buzzword sales pitch.

It's supposed to. Investors want to hear this, not some crap about efficiency, when the entire world is talking about AI.

Edit: TBC, I care about efficiency and it's not crap, but that's likely the view of investors.

> The last thing though - “hybrid algorithms leveraging AI efficiency” sounds an awful lot like a buzzword sales pitch.

Oddly, to me this just sounds like efficiency gains by potentially introducing massive security holes i.e. the vector of the Meltdown/Spectres of the future. It also seems like they're trying to sell AI as some sort of secular qubit that they'll be error-correcting.

Not a cpu designer, but I remember AMD using AI for branch prediction since long ago. Hopefully they still mention AI in that sense and not only branding.

Reply on news.ycombinator.com