When working at major networking services company, I'd often joke that our wireless sucked because every network engineer worth a damn was pimped out for billable hours.

India has this problem. One could joke that any person worth employing (and even more who aren't) are contracted out in other countries and probably intend to stay there. As a result, there are far too few highly skilled people in India.

I've been told by at least a few Indian people that no sane person who has the option to leave, even to live somewhere people hate you would stay to work in India. They tell me that Indian investment houses don't want to "maybe make money by maybe building a successful Indian tech company when they can definitely make money by selling India's people".

If the market for India's people is staggering, now is a great time for India to invest in India.

The current political nightmare known as Dumb and Dumber (Trump and Biden, you decide who gets the "er") is proof that doesn't work.

I've travelled to China to work with Huawei as a representative of a western scientific HPC project recently and no matter where I went, the people were friendly and welcoming. The Chinese are generally very positive towards... pretty much everyone. This is almost certainly because the can choose what to consume of western media, and of course most western political media regarding negativity towards China is blocked. But I also didn't meet anyone who didn't have a VPN.

We on the other hand are restricted to consuming a limited selective subset of Chinese media and culture. We only see what someone decided was worth translating. That almost always means media that makes headlines, not media that we can gain from.

And, no, China doesn't win. We can only win by behaving rationally and working together.

We can't justify every developer requiring Adobe subscriptions. We of course have a few people running Adobe, but we buy software, not rent it.

Privacy is dead. Has been for a while.

You have to choose,

1) live in a country like the US where you never know which nutjob politician is going to pass laws to help their campaign sugar daddies make money watching you (think Obama and his handing of the DOJ over to the RIAA and MPAA, or GWB signing off on "habeus corpus is more of a recommendation" bit)

2) Move somewhere who passes laws restricting the use of the tech.

The US is unhealthy because the whole country is paranoid. Try to think of the last day someone in the US wasn't trying to make a buck by making you think someone is out to get you.

Who on planet earth would buy a new PC for AI?

Every exciting AI model is gigabytes in size, does one particular thing (llm, diffusion,...) and even with a bank of neural networking processors, each with 8GB or more of HBM3, the cloud will be faster.

I'd prefer less drama with my IT

Russia can definitely write software. They're a true world resource for developers.

But, Russia has never been able to do computing hardware and certainly not chips. China is the only country on earth capable of playing catch-up and while I believe China still openly trades with Russia, China who is still a generation behind on AI hardware doesn't sell their current generation externally.

China who already was working to keep pace and catch up used 3 years and far more money than Russia can afford and far more people than Russia has since destroying a whole generation of young people by treating them as cannon fodder is scraping by to catch up. And that's not an insult to China, that's high praise. Any fool in Russia who thinks "if China can do it, so can we" is as foolish as anyone who thinks "if we can do it, we're special and China can't possibly catch up for 10 years"

It is good this happened.

We needed to show regulators that companies like GM who actually have to recall their network connected cars to perform software updates shouldn't be allowed to make self driving or any other kind of cars.

They went 400Gb/s about 8 years ago and at the time, they were procuring 4 sets of optics for their trans-siberian link. Which means they've done 1.6Tb/sec for some time.

I have felt for some time that connected cars should have a website where vehicle diagnostics are displayed.

CANbus is a piece of shit without a tool capable of interpreting registers. Modern vehicles can achieve far more with modern LVDS interconnectivity. Of course, there is no point standardizing this if the car can provide a web page to access and even alter data. Maybe requiring a JSON API and Swagger support would be smart.

The real issue is aftermarket parts. In modern cars, this may not make sense

Car company recalls cars for a software update?

Don't the simply do it over the air? Did they really make self driving cars that lack the ability to download and install updates?

I don't see why we keep focusing on GM failures in cars, electronics, computers and such.

It's GM, they have never made anything that functions particularly well.

GM's best possible marketing slogan is "Buy GM, we work like a Lada, but we're still less shitty than Ford"

What does AI have to do with it?

AI runs just fine on larger node processes. Using chiplets 10nm can be produced with near 100% yield. Performance-wise, 10nm vs 2nm makes little difference since parallelism is everything. Massive cache or HBM is pretty important, but memory tech is often on older nodes for reliability.

If I were to design a neural networking core, I'd hope for smaller nodes but design for bigger as it would give me leverage negotiating. I'd start by choosing a memory architecture for 24GB. Micron has 24GB HBM3 but it only does 1.2TB/sec, so either 48GB or double 12GB to increase bandwidth. And that's DUV, not EUV.

Next step is to tensor ALUs, 8-bit pyramid multipliers are cheap in transistors and Multiply-Accumulate for a full tensor is cheap.

Then interconnect, a limited data cache per core makes sense, l1 coherency will be expensive. This won't be a major issue for inference, but will be a problem for training. It would make sense to design a bus line for whether coherency is important or not. Then, connect 16 or 64 cores in a full mesh.

A pipeline is needed and as tensors mostly don't profit from branch prediction or out of order execution, a DSP-like pipeline could work well.

A central memory arbitrator will be the last core component. A mesh is best, but to fit a lot more tensor chiplets, a series of meshed ring busses might work without costing too much for latency. This could be the one place a better node would shine as you'd want to run ring busses at considerably higher frequencies.

But, yeh, I stand by argument, I don't need better nodes, I just need reliable packaging and I can gain energy efficiency by using water cooling and harnessing waste heat like Huawei does.

We're evaluating Huawei Atlas 9000 because we need a reliable supplier and US government and companies buy everything up and make the waiting lists so long the tech is obsolete by the time it arrives. Huawei delivers in a month.

Unless you can become the default chatbot in a default browser on a major OS, it's a wasted investment.