Paywalled...

Loaded fine for me. Maybe substack has some limit of number of posts that people without accounts can read and I just haven’t visited all that many posts this month.

The article does end with "The rest of this edition is for paid subscribers" so I guess only part of it is readable via the above link.

At the very least the article does talk about one whole case study before offering more so there is some substantial free content, it’s an interesting way to partition an article for that.

On the other hand the length of the fire part of that article is enough that it is a stand-alone article. The whole thing is starting to look like one of those roundups or listicles that I’d like to teach YOShInOn to suppress (an article about everything is ‘t about something in particular) but the free part of the article is “just right” to me.

Agree completely about lists – it's a point I hammer home quite a bit: https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor....

In this case the article seemed more substantial than the usual roundup/list so we invited cpeterso to repost it. Maybe that was the right decision after all :)

It looks like he just dropped the paywall on the rest of the article, since the rest of the posts are shorter than that first one so I think the length of the article is just fine but really I think each of those cases is worth its own treatment as long as the first.

>YOShInOn

What is this? A cursory search didn't get me anywhere.

If you searched on Google you probably found an image of a girl in a strange raincoat holding a puppet. The girl is Yoshino 四よ糸し乃 (note a real Japanese name with Kanji in it) and the puppet is called Yoshinon よしのん (not a real name, spelled out phonetically.) I’m not going to say I am like that girl (I don’t have a magic green raincoat, never mind a green raincoat at all) but my YOShInOn is named after よしのん.

There is also YOSHINON which is the network name of the body that the YOShInOn system inhabits, seen here

https://mastodon.social/@UP8/110568695732952713

If you do a search on Hacker News, however, you will find what you are looking for

https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...

in fact probably a lot more than you are looking for. Sorry I have been too busy coding that and another side project to blog about it.

Author here. As noted elsewhere this originally had a paywall in the middle but I’ve removed it for the time being. Hope the extra projects are of interest.

In other news the US is looking to limit China’s access to RISC-V

https://www.reuters.com/technology/us-china-tech-war-risc-v-...

Discussed here

https://news.ycombinator.com/item?id=37798178

China isn’t going to stop using riscv no matter what the US says

It’ll just force them to be self sufficient faster

The US is clearly targeting Google releasing Android for RISC-V, or Qualcomm releasing RISC-V chips which can run RISC-V Android.

RISC-V has been an 'obvious' choice for Android for a long time - particularly for lower end devices or set-top-box type designs where the vendor has control of all the software (no third party binaries that need arm emulation).

The fact an official RISC-V port is soon to be released, I suspect is the reason behind this activity now.

There will still be US suppliers who want riscv android, so google will still do it. China could still design their own riscv chips and run AOSP without google, so it wouldn’t change much

But without Google making all the app ecosystem support RISC-V, I suspect any RISC-V phone would have a chicken and egg problem - with no apps available, no user would want to run it.

Google is in a unique position to solve that by making the toolchain automatically compile for RISC-V as well, and in a matter of months most apps will support it.

Google is in unique position, because they are a company with a track record delivering software in massive ecosystem scale.

Non-US companies like Samsung, Nokia, and every Chinese company, cannot write quality software, based on the 25 years of track record. This is why they all depend on Google for Android.

This has been discussed earlier on HN why is this and it comes down to corporate culture and such matters.

How would that change anything in regards to China having access to RISC-V?

AOSP already supports RISC-V.

Reverting course wouldn't prevent China from using the code they already have, which in the first place was largely contributed by companies based in China.

They don't have the technology or the expertise to produce high quality, high performance chips.

It's a moot point whether or not they have access to the specs, and the only reason to think otherwise is buying into chinese propaganda.

Like this blatant bullshit: https://www.scmp.com/comment/opinion/article/3236924/huaweis...

Saying "someone doesn't have the technology or expertise" is a point in time statement only true that moment. Progress marches ever forward, and China has quite a few exceptional engineers and scientists (and money) to pursue their objectives.

> China has quite a few exceptional engineers and scientists (and money) to pursue their objectives

They could, if they played the long game: "... the only way that China is going to break through the 5nm barrier is by going back not just a few years but a few decades: Shanghai Micro Electronics Equipment (SMEE), China’s answer to ASML, is building fully home-grown lithography machines using krypton fluoride lasers (Krf); this is previous generation DUV — first introduced around the turn of the century — using wavelengths of 248nm (current DUV processes use argon fluoride lasers (Arf) to achieve 193nm wavelengths). What SMEE specifically and China broadly needs to do is move down the learning curve for lithography. SMEE needs to learn how to use Arf, then immersion, and then EUV. It doesn’t need to take 20 years, either: while ASML needed to invent everything, SMEE has the luxury of already knowing what is possible; the same thing applies to every other part of the semiconductor value chain.

This, then, gets at why this PR cycle could be bad for China: the Mate 60 is a short-term propaganda victory that threatens to distract leaders who don’t understand semiconductors from facing up to and fully investing in the long road to semiconductor independence; it’s harder to accept that it might take 15 years when you have 7nm capability in hand right now."

https://stratechery.com/2023/the-huawei-mate-60-pro-7nm-back...

Unfortunately, Xi is a dictator. Dictators don't have the luxury of playing long games.

Well now that you’ve told them, they’re retooling. The Grand Emperor thanks you.

That's not enough to recreate EUV lithography, which is probably the most complex process in the world.

Yet recreating it today is going to be much easier than creating it for the first time was n years ago.

And China doesn't have to start from scratch; It has the resources to leverage insiders, spy and steal blueprints. And is not afraid of using them.

I actually do think it would be harder to recreate it, because you have to remake the whole supply chain, and it's several layers deep of sole suppliers that did already exist when it was originally developed.

> Like this blatant bullshit:

Which part(s) of linked article would you consider inaccurate?

If China can make fast RISC-V chips to compete, they can make fast ARM chips to compete too.

Artificially limiting software is only going to hold back everyone.

linked talk

> Accelerating ML Recommendation with over a Thousand RISC-V/Tensor Processors

Why are they running recommendation engines on nets? I thought this has been proven that this is region where Random Forests / XGBoost outperform nets.

Deep learning-based recommendation models are very popular in industry. Not sure what “proof” you’re referring to, but it’s definitely not the case that random forests outperform DL based approaches in general.

I did a bit of deeper dive here, and the evidence remains mixed, but definitely generally in favor of random forests for recommendation.

My two cents: Most recommendation engines are far from perfect, and the customer is too often left scratching their heads. Given this, why bear the orders of magnitude more expense to go with with a deep learning solution?

If you're interested in low-power embedded, the WCH chip there has two powersaving modes; at 3.3V (it is 5V tolerant). At the lowest-power mode, the whole SoC consumes 9uA, and has a wakeup time of 200us. I wonder how much extra oomph the chip would get if there was a lower-speed power-saving mode – although I am ignorant about whether this would be compatible with SRAM.

I interned last summer at Esperanto. Glad to see it on the list! It's definitely interesting. Some colleagues worked on porting OpenJourney (fine-tuned stable diffusion) to Et-Soc-1 and it was fast and more efficient than Nvidia gpus!