How AlphaChip transformed computer chip design(deepmind.google) |
How AlphaChip transformed computer chip design(deepmind.google) |
Meanwhile, MediaTek built on AlphaChip and is using it widely, and announced that it was used to help design Dimensity 5G (4nm technology node size).
I can understand that, when this open-source method first came out, there were some who were skeptical, but we are way beyond that now -- the evidence is just overwhelming.
I'm going to paste here the quotes from the bottom of the blog post, as it seems like a lot of people have missed them:
“AlphaChip’s groundbreaking AI approach revolutionizes a key phase of chip design. At MediaTek, we’ve been pioneering chip design’s floorplanning and macro placement by extending this technique in combination with the industry’s best practices. This paradigm shift not only enhances design efficiency, but also sets new benchmarks for effectiveness, propelling the industry towards future breakthroughs.” --SR Tsai, Senior Vice President of MediaTek
“AlphaChip has inspired an entirely new line of research on reinforcement learning for chip design, cutting across the design flow from logic synthesis to floor planning, timing optimization and beyond. While the details vary, key ideas in the paper including pretrained agents that help guide online search and graph network based circuit representations continue to influence the field, including my own work on RL for logic synthesis. If not already, this work is poised to be one of the landmark papers in machine learning for hardware design.” --Siddharth Garg, Professor of Electrical and Computer Engineering, NYU
"AlphaChip demonstrates the remarkable transformative potential of Reinforcement Learning (RL) in tackling one of the most complex hardware optimization challenges: chip floorplanning. This research not only extends the application of RL beyond its established success in game-playing scenarios to practical, high-impact industrial challenges, but also establishes a robust baseline environment for benchmarking future advancements at the intersection of AI and full-stack chip design. The work's long-term implications are far-reaching, illustrating how hard engineering tasks can be reframed as new avenues for AI-driven optimization in semiconductor technology." --Vijay Janapa Reddi, John L. Loeb Associate Professor of Engineering and Applied Sciences, Harvard University
“Reinforcement learning has profoundly influenced electronic design automation (EDA), particularly by addressing the challenge of data scarcity in AI-driven methods. Despite obstacles including delayed rewards and limited generalization, research has proven reinforcement learning's capability in complex electronic design automation tasks such as floorplanning. This seminal paper has become a cornerstone in reinforcement learning-electronic design automation research and is frequently cited, including in my own work that received the Best Paper Award at the 2023 ACM Design Automation Conference.” --Professor Sung-Kyu Lim, Georgia Institute of Technology
"There are two major forces that are playing a pivotal role in the modern era: semiconductor chip design and AI. This research charted a new path and demonstrated ideas that enabled the electronic design automation (EDA) community to see the power of AI and reinforcement learning for IC design. It has had a seminal impact in the field of AI for chip design and has been critical in influencing our thinking and efforts around establishing a major research conference like IEEE LLM-Aided Design (LAD) for discussion of such impactful ideas." --Ruchir Puri, Chief Scientist, IBM Research; IBM Fellow
I think the next step is arrays of memory-based compute.
GPUs still treat memory as separate from compute, they just have wider bottlenecks than CPUs.
- A rebuttal by a researcher within Google who wrote this at the same time as the "AlphaChip" work was going on ("Stronger Baselines for Evaluating Deep Reinforcement Learning in Chip Placement"): http://47.190.89.225/pub/education/MLcontra.pdf
- The 2023 ISPD paper from a group at UCSD ("Assessment of Reinforcement Learning for Macro Placement"): https://vlsicad.ucsd.edu/Publications/Conferences/396/c396.p...
- A paper from Igor Markov which critically evaluates the "AlphaChip" algorithm ("The False Dawn: Reevaluating Google's Reinforcement Learning for Chip Macro Placement"): https://arxiv.org/pdf/2306.09633
In short, the Google authors did not fairly evaluate their RL macro placement algorithm against other SOTA algorithms: rather they claim to perform better than a human at macro placement, which is far short of what mixed-placement algorithms are capable of today. The RL technique also requires significantly more compute than other algorithms and ultimately is learning a surrogate function for placement iteration rather than learning any novel representation of the placement problem itself.
In full disclosure, I am quite skeptical of their work and wrote a detailed post on my website: https://vighneshiyer.com/misc/ml-for-placement/
The AlphaChip authors address criticism in their addendum, and in a prior statement from the co-lead authors: https://www.nature.com/articles/s41586-024-08032-5 , https://www.annagoldie.com/home/statement
- The 2023 ISPD paper didn't pre-train at all. This means no learning from experience, for a learning-based algorithm. I feel like you can stop reading there.
- The ISPD paper and the MLcontra paper both used much larger older technology node sizes, which have pretty different physical properties. TPU has a sub 10nm technology node size, whereas ISPD uses 45nm and 12nm. These are really different from a physical design perspective. Even worse, MLcontra uses a truly ancient benchmark with >100nm technology node size.
Markov's paper just summarizes the other two.
(Incidentally, none of ISPD / MLcontra / Markov were peer reviewed - ISPD 2023 was an invited paper.)
There's a lot of other stuff wrong with the ISPD paper and the MLcontra paper - happy to go into it - and a ton of weird financial incentives lurking in the background. Commercial EDA companies do NOT want a free open-source tool like AlphaChip to take over.
Reading your post, I appreciate the thoroughness, but it seems like you are too quick to let ISPD 2023 off the hook for failing to pre-train and using less compute. The code for pre-training is just the code for training --- you train on some chips, and you save and reuse the weights between runs. There's really no excuse for failing to do this, and the original Nature paper described at length how valuable pre-training was. Given how different TPU is from the chips they were evaluating on, they should have done their own pre-training, regardless of whether the AlphaChip team released a pre-trained checkpoint on TPU.
(Using less compute isn't just about making it take longer - ISPD 2023 used half as many GPUs and 1/20th as many RL experience collectors, which may screw with the dynamics of the RL job. And... why not just match the original authors' compute, anyway? Isn't this supposed to be a reproduction attempt? I really do not understand their decisions here.)
Kahng's ISPD 2023 paper is not in dispute - no established experts objected to it. The Nature paper is in dispute. Dozens of experts objected to it: Kahng, Cheng, Markov, Madden, Lienig, Swartz objected publically.
The fact that Kahng's paper was invited doesn't mean it wasn't peer reviewed. I checked with ISPD chairs in 2023 - Kahng's paper was thoroughly reviewed and went through multiple rounds of comments. Do you accept it now? Would you accept peer-reviewed versions of other papers?
Kahng is the most prominent active researcher in this field. If anyone knows this stuff, it's Kahng. There were also five other authors in that paper, including another celebrated professor, Cheng.
The pre-training thing was disclaimed in the Google release. No code, data or instructions for pretraining were given by Google for years. The instructions said clearly: you can get results comparable to Nature without pre-training.
The "much older technology" is also a bogus issue because the HPWL scales linearly and is reported by all commercial tools. Rectangles are rectangles. This is textbook material. But Kahng etc al prepared some very fresh examples, including NVDLA, with two recent technologies. Guess what, RL did poorly on those. Are you accepting this result?
The bit about financial incentives and open-source is blatantly bogus, as Kahng leads OpenROAD - the main open-source EDA framework. He is not employed by any EDA companies. It is Google who has huge incentives here, see Demis Hassabis tweet "our chips are so good...".
The "Stronger Baselines" matched compute resources exactly. Kahng and his coauthors performed fair comparisons between annealing and RL, giving the same resources to each. Giving greater resources is unlikely to change results. This was thoroughly addressed in Kahng's FAQ - if you only could read that.
The resources used by Google were huge. Cadence tools in Kahng's paper ran hundreds times faster and produced better results. That is as conclusive as it gets.
It doesn't take a Ph.D. to understand fair comparisons.
Other commenters already addressed the pre-training issue. Please kindly include a link to Kahng's 2023 discussion addressing your complaints. Otherwise, you are unfairly supporting those people you know.
Kahng's placer is open-source and was used in the Nature paper. It does not make sense to accuse Kahng of colluding with companies against open-source.
https://en.wikipedia.org/wiki/Top_Chess_Engine_Championship
So perhaps the critics had a point there.
This is what you get if you make academic researchers compete for citation counts.
Pretraining seems to be an important aspect here, and it makes sense that such pretraining requires good examples, which unfortunately for the free lunch people, is not available to the public.
That's what you get when you let big companies do fundamental research. Would it be better if the companies did not publish anything about their research at all?
It all feels a bit unproductive to attack one another.
Whichever approach ends up winning is improved by careful evaluation and replication of results
When you see a chip that has the datapath identified and laid out properly by a computer algorithm, you've got something. If not, it's vapor.
So, if your layout still looks like a random rat's nest? Nope.
If even a random person can see that your layout actually follows the obvious symmetric patterns from bit 0 to bit 63, maybe you've got something worth looking at.
Analog/RF is a little tougher to evaluate because the smaller number of building blocks means you can use Moore's Law to brute force things much more exhaustively, but if things "looks pretty" then you've got something. If it looks weird, you don't.
They must feel vindicated by their work turning out to be so fruitful now.
[1] https://www.theregister.com/AMP/2023/03/27/google_ai_chip_pa...
[2] https://regmedia.co.uk/2023/03/26/satrajit_vs_google.pdf
I think it is time for you to take a deep breath and think about what you are doing and why.
You seem to be obsessed with the idea that this work is overrated. MediaTek and Google don't think so, and use it in production for their chips, including TPU, Dimensity, Axion, and others. If you're right and they're wrong, using this method loses them money. If it's the other way around, then using this method makes them gain money.
Please read PG's post and ask yourself if it applies to you: https://www.paulgraham.com/fh.html
Chatterjee settled his case. He has moved on. This is not some product being sold -- it is a free, open-source tool. People who see value in it use it; others don't, and so they don't. This is how it always works, and it's fine.
They don’t produce but they are tailored for them just the same. “We have” doesn’t have to mean “we made”. They don’t say it as such here but elsewhere they refer to the IP they can make available, which can also be made in house or cross licensed and still count as “we have”.
Without knowing much, my guess is that “quality” of a chip design is multifaceted and heavily dependent on the use case. That is the ideal chip for a data center would look very different from those for a mobile phone camera or automobile.
So again what does “better” mean in the context of this particular problem / task.
[1] https://en.wikipedia.org/wiki/Eurisko
What's more, Eurisco was then used in designing Traveler TCS' game fleet of battle spaceships. And Eurisco used symmetry-based placement learned from VLSI design in the design of the spaceships' fleet.
Can AlphaChip's heuistics be used anywhere else?
Instead they could have demonstrated their amazing method on any number of standard NP hard optimization problems e.g. traveling salesman, bin packing, ILP, etc. where we can generate tons of examples and verify easily whether it produces better results than other solvers or not.
This is why many in the chip design and optimization community felt that the paper was suspicious. Even with this addendum they adamantly refuse to share any results that can be independently verified.
It is not obscure (in chip design). If anything it is one of the most easily reachable problems. Almost every other PhD student in the field has implemented a macro placer, even if just for fun, and there are frequent academic competitions. A lot of design houses also roll their own macro placers since it's not a difficult problem and generally adding a bit of knowledge of your design style can help you gain an extra % over the generic commercial tools.
It does not surprise me at all that they decided to start with this for their foray into chip EDA. It's the minimum effort route.
Chips designed with the help of AlphaChip are in datacenters and Samsung phones, right now. That's pretty neat!
Also: when is this coming to KiCad? :)
PS: It would also be nice to apply a similar algorithm to graph drawing (e.g. trying to optimize for human readability instead of electrical performance).
We're in the timeline that took the wrong path. The other world has isolinear memory, which can be used for compute, or as memory, down to the LUT level. Everything runs at a consistent speed, and hardware faults LUTs can be routed around easily.
Better architectures without the yearly investment train will no longer be better quite quickly.
You would need to be 100x to 1000x better in order to pull the investment train onto your tracks.
Don’t has been impossible for decades.
Even so, I think we will see such a change in my lifetime.
AI could be that use case that has a strong enough demand pull to make it happen.
We will see.
But if you do pay attention to the programming model, they're unusable. You'll see that dozens of these approaches have come and gone, because it's impossible to write software for them.
I don't think it's necessarily demand or any particular calculation that makes things happen. I think people including investors are just herd animals. They aren't enthusiastic until they see the herd moving and then they want in.
I don’t want art that wasn’t made by a human, no matter how visually stunning or indistinguishable it is.
Imagine your favorite movie, the most moving book. You read it, it changed you, then you found out it was an AI that generated it in a mere 10 seconds.
Artificial sentimentality is useless in the face of reality. That human endeavor is simply data points along an multi-dimensional best fit curve.
TPU v5e [1]: not available for purchase, only through GCP, storage=5B, LLM-Model=7B, efficiency=393TFLOP.
Forget LLM's. What DeepMind is doing seems more like how an AI will rule, in the world. Building real world models, and applying game logic like winning.
LLM's will just be the text/voice interface to what DeepMind is building.
Protein Folding? That was against a defined data set and other organizations.
Nobody can re-produce? Isn't that the definition of a competitive advantage?
They are building something others can't, and that is bad? That is what companies do.
For example, how much better are these latest gen TPU's when compared to NVidia's equivalent offering ?
Re:using RL and other types AI assistance for chip design, Nvidia and others are doing this too
I'd even dare to claim we are already at the point where the growth has stopped, but even then you will only see the effect in a decade or so as there are still many small low-hanging fruits you can fix, but no big improvements.
Practically speaking, though, maintaining Moore's law would have been economically prohibitive if circuit design and layout had not been automated.
> Synopsys DSO.ai autonomously explores multiple design spaces to optimize PPA metrics while minimizing tradeoffs for the target application. It uses AI to navigate the design-technology solution space by automatically adjusting or fine-tuning the inputs to the design (e.g., settings, constraints, process, flow, hierarchy, and library) to find the best PPA targets.
Still, the fact that Google uses it for TPU is pretty telling - this is a multi-billion dollar, mission-critical chip design effort, and there's no way they'd make TPU worse just to prop up a research paper. MediaTek's production use is also a good indicator.
Floorplanning/placement/synthesis is a billion dollar industry, so if their approach were really revolutionary they would be selling the technology, not wasting their time writing blog posts about it.
https://research.google/pubs/spanner-googles-globally-distri...
or Bigtable?
https://research.google/pubs/bigtable-a-distributed-storage-...
or GFS?
or MapReduce?
or Borg?
or...I think you get the idea.
Maybe all together, but I don't think automatic placement algorithms are a billion dollar industry. There's so much more to it than that.
(no paywall): https://www.cl.cam.ac.uk/~ey204/teaching/ACS/R244_2021_2022/...
I assume that the human benchmark is a human using existing EDA tools, not a guy with a pocket protector and a roll of tape.
To quote certain popular TV series .... Sorry, are you from the past? Do your "production" chips only have a couple dozen macros or what?
What nonsense! XD
I want to emphasize the biggest barrier for IC design to the outsiders: prohibitively expensive software licenses. IC design software costs are the much higher than conpute and the production costs, and often similar order of magnitude but definitely higher than engineer salaries. This is because of the monopoly of the 3 big companies (Synopsys, Cadence and Mentor Graphics). What wxcites me the most about stuff like OP isn't AI, everyone is doing that. It's the premise of more competition and even open source tool options. In the good old days companies used to have their im-house tools. They are all sacrificed (and pretty much none made open source) because investors thought it's not a core business, so it's inefficient. Now even Nvidia or Apple have no alternative.
> Generally considered to be the strongest GPU engine, it continues to provide open data which is essential for training our NNUE networks. They released version 0.31.1 of their engine a few weeks ago, check it out!
[1]
I’d say the impact AlphaZero has had on chess and go can’t be understated considering it’s a general algorithm that at worst is highly competitive with purpose built engines. And that’s ignoring the actual point of why DeepMind is doing any of this which is for GAI (that’s why they’re not constantly trying to compete with existing engines)
[1] https://lichess.org/@/StockfishNews/blog/stockfish-17-is-her...
This makes sense given that both authors of the paper left Google in 2022. And one no longer seems to work in the chip design space, plausibly because of the bullying by entrenched folks.
Then again, since rejoining Google the other author has produced around one patent per month in chip design with RL in 2023 and 2024, so perhaps they feel there is a marketable tool here that they don't want to share.
Did they tested their ML solution ? With real world chips ? Are there any "benchmarks" that show that their chip performs better ?
I think it would feel hollowed out, disingenuous.
It feels too close to being a rat with a dopamine button, meaningless hedonism.
I haven’t thought it through particularly thoroughly though, I’d been interested in hearing other opinions. These philosophical questions quickly approach unanswerable.
With the current trendline of AI progress in the last decade the question has a high possibility of being answered by being actualized in reality.
It's not a random question either. With AI quickly entrenching itself into every aspect of human creation from art, music, to chip design, this is all I can think about.
We are going to relearn this lesson with ambulation and grasping as all the large companies are trying to make useful robots from human shadowing to reduce the gigantic sample size requirements burden with self play. Likely after the initial years computers will just get a couple more doublings in compute per watt and we will see the full self training models take over those domains as the old human data biased models get thrown out.
>Results >Ariane RISC-V CPU >View the full details of the Ariane experiment on our details page. With this code we are able to get comparable or better results training from scratch as fine-tuning a pre-trained model.
The paper includes a graph showing that it takes longer for Ariane to train without pre-training however the results in the end are the same.
Sometimes training from scratch is able to match the results of pre-training, given ~5X more time to converge. Other times, though, it never does as well as a pre-trained model, converging to a worse final result.
This isn't too surprising -- the whole point of the method is to be able to learn from experience.
Most design houses don’t write their own macro placers but customize commercial flows for their designs.
The problem with macro placement as an RL technology demonstrator is that to evaluate quality you need to go through large parts of the design flow which involves using other commercial tools. This makes it incredibly hard to evaluate superiority since all those steps and tools add noise.
Easier problems would have been to use RL to minimize the number of gates in a logic circuit or just focus on placement with half perimeter wirelength (I think this is what you mean with your grad student example). Essentially solving point problems in the design flow and evaluating quality improvements locally.
They evaluated quality globally and only globally and that destroys credibility in this business due to the noise involved unless you have lots of examples, can show statistical significance, and (unfortunately for the authors) also local improvements.
That’s what the follow on studies did and that’s why the community has lost faith in this particular algorithm.
Most I don't know, but all the mid-to-large ones have automated macro placers. Obviously, the output is introduced into the commercial flow, generally by setting placement constraints. The larger houses go much further and may even override specific parts of the flow, but not basing it on an commercial flow is out of the question right now.
> The problem with macro placement as an RL technology demonstrator is that to evaluate quality you need to go through large parts of the design flow which involves using other commercial tools.
Not really, not any more than any other optimization such as e.g. frontend which I'm more familiar with. If you don't want to go through the full design flow (which I agree introduces noise more than anything else), then benchmark your floorplans in some easily calculable metric (e.g., HPWL). Likewise, if you want to test the quality of some logic simplification _in theory_ you'd have to also go through the entire flow (backend included), but no one does that and you just evaluate some easily calculable metric e.g. number of gates. These distinctions are traditional more than anything else.
Academic macro placers generally have limited access to commercial flows (either due to licensing issues or computing resource availability) so it is rather common to benchmark them in other metrics. Google paper tried to be too smart for its own good and therefore incomparable to anything academic.
As for "nobody can re-produce", no, that's not the definition. Imaginary things are not competitive advantage. They are exaggerating, and that's bad. But yeah, that's what companies do, you are right.
I get the impression you just aren't keeping up with DeepMind.
They have made huge break throughs in science, and they publish their results in Nature. Just because the parent company Google had some bad demo's doesn't mean it is all bunk.
So guess if you are of the ilk that just doesn't trust anything anymore, that there is no peer reviews, all science is a fraud. I really can't help that.
The team that did RL for chips work was at GoogleBrain, and you already pointed out that Google had bad demos. The fact that this team was absorbed into DeepMind does not magically rub the successes of DeepMind onto them.
The RL for chips results were nothing like AlphaGo. Imagine if AlphaGo claimed to beat unknown go players, you would laugh. But the Nature paper on RL for chips claims to outperform unknown chip engineers. Also, imagine if AlphaFold claimed to fold only proprietary proteins. The Nature paper on RL for chips reports results on a small set of proprietary chip blocks (they released one design, and the results are not great on that one). That's where imaginary results come up. One of these things is not like the others.
https://www.cl.cam.ac.uk/~ey204/teaching/ACS/R244_2021_2022/...
Of course they do. I'm waiting for their products.
As for external usage of the method - MediaTek is one of the largest chip design companies in the world, and they built on AlphaChip. There's a quote from a MediaTek SVP at the bottom of the GDM blog post:
"AlphaChip's groundbreaking AI approach revolutionizes a key phase of chip design. At MediaTek, we've been pioneering chip design's floorplanning and macro placement by extending this technique in combination with the industry's best practices. This paradigm shift not only enhances design efficiency, but also sets new benchmarks for effectiveness, propelling the industry towards future breakthroughs."
The more marketing claims we see, the less compelling the Google story is.
Your perseverance is as admirable as it is suspicious. You are the lonely voice here defending the Google announcement.
Even if the AlphaChip authors redid Kahng's study properly, this still wouldn't give us useful information -- what matters is AlphaChip's ability to optimize chips in a real-life, production setting, for modern chips, where millions of dollars are on the line.
Really, I wasn't arguing about the chips so much. I mentioned DeepMind and you said I must like Sci-Fi, so I assumed you were inferring that DeepMind results were not that extraordinary.
And, I can't keep up with the internal re-orgs now that DeepMind was merged with the other groups at Google. So Maybe I am assuming too much, if this wasn't the same DeepMind group. -- Though I think when companies merge groups like this, they are definitely hoping some 'magic success rubs off on them'.
I guess for the Chip design, is your argument about it was compared against generic human engineer. So if they would set up some competition with some humans, that would satisfy your issue with the results?
So my original more flippant post was Sci-Fi, it's just things are changing fast enough that the lines have blurred, and DeepMind has real results that aren't Sci-Fi:
Take Games as a simplified world models, DeepMind has made a lot more progress in winning games than other companies, then take some of the other companies that have had break throughs in Video-to-Real-World-Models, where the video can be broken down into categories, and those can be fed into a 'Game' function. Now put that on a loop(default mode network in brain) and in a robot body (so it has embodied subjective experience of the world, there are consequences of actions). And I am making a bit of a Sci-Fi leap that you can get human behavior. And, if they can then make leap to designing chips, then they can reach that hockey stick of increasing intelligence.
So, guess I am making Sci-Fi leap. But I think the actual results from DeepMind already seem Sci-Fi like but are real. So are we really that far away, given things we thought would take 100's of years are falling by the wayside.
ok. take back ad-hominem. but it is hard to tell on the internet. it seemed like you were questioning verified results, and you must know that there is a large contingent on internet that casts doubt on all science. So once someone goes down that path it is easier to ignore them.
HPWL (half-perimeter wirelength) is an approximation of wirelength, which is only one component of the chip floorplanning objective function. It is relatively easy to crunch all the components together and optimize HPWL --- minimizing actual wirelength while avoiding congestion issues is much harder.
Simulated annealing is good at quickly converging on a bad solution to the problem, with relatively little compute. So what? We aren't compute-limited here. Chip design is a lengthy, expensive process where even a few-percent wirelength reduction can be worth millions of dollars. What matters is the end result, and ML has SA beat.
(As for conflict of interest, my understanding is Cadence has been funding Kahng's lab for years, and Markov's LinkedIn says he works for Synopsis. Meanwhile, Google has released a free, open-source tool.)
Everything optimized in Nature RL is an approximation. HPWL is where you start, and RL uses it in the objective function too. As shown in "Stronger Baselines", RL loses a lot by HPWL - so much that nothing else can save it. If your wires are very long, you need routing tracks to route them, and you end up with congestion too.
SA consistently produces better solutions than RL for various time budgets. That's what matters. Both papers have shown that SA produces competent solutions. You give SA more time, you get better solutions. In a fair comparison, you give equal budgets to SA and RL. RL loses. This was confirmed using Google's RL code and two independent SA implementations, on many circuits. Very definitively. No, ML did not have SA beat - please read the papers.
Cadence hasn't funded Kahng for a long time. In fact, Google funded Kahng more recently, so he has all the incentives to support Google. Markov's LinkedIn page says he worked at Google before. Even Chatterjee, of all people, worked at Google.
Google's open-source tool is a head fake, it's practically unusable.
Update: I'll respond to the next comment here since there's no Reply button.
1. The Nature paper said one thing, the code did something else, as we've discovered. The RL method does some training as it goes. So, pre-training is not the same as training. Hence "pre". Another problem with pretraining in Google work is data contamination - we can't compare test and training data. The Google folks admitted to training and testing on different versions of the same design. That's bad. Rejection-level bad.
2. HPWL is indeed a nice simple objective. So nice that Jeff Dean's recent talks use it. It is chip design. All commercial circuit placers without exception optimize it and report it. All EDA publications report it. Google's RL optimized HPWL + density + congestion
3. This shows you aren't familiar with EDA. Simulated Annealing was the king of placement from mid 1980s to mid 1990s. Most chips were placed by SA. But you don't have to go far - as I recall, the Nature paper says they used SA to postprocess macro placements.
SA can indeed find mediocre solutions quickly, but keeps on improving them, just like RL. Perhaps, you aren't familiar with SA. I am. There are provable results showing SA finds optimal solution if given enough time. Not for RL.
I'm glad you agree that HPWL is a proxy metric. Optimizing HPWL is a fun applied math puzzle, but it's not chip design.
I am unaware of a single instance of someone using SA to generate real-world, usable macro layouts that were actually taped out, much less for modern chip design, in part due to SA's struggles to manage congestion, resulting in unusable layouts. SA converges quickly to a bad solution, but this is of little practical value.
This is written as a textbook example logical fallacy of appeal to authority.
2. HPWL is indeed a nice simple objective. So nice that Jeff Dean's recent talks use it. It is chip design. All commercial circuit placers without exception optimize it and report it. All EDA publications report it. Google's RL optimized HPWL + density + congestion
3. This shows you aren't familiar with EDA. Simulated Annealing was the king of placement from mid 1980s to mid 1990s. Most chips were placed by SA. But you don't have to go far - as I recall, the Nature paper says they used SA to postprocess macro placements.
SA can indeed find mediocre solutions quickly, but keeps on improving them, just like RL. Perhaps, you aren't familiar with SA. I am. There are provable results showing SA finds optimal solution if given enough time. Not for RL.