Breakfast Bytes Blogs

Because of the Thanksgiving break (and me taking the whole week off), I skipped the regular November Update post, but suddenly there is so much semiconductor news that I'm doing it out of band today even though it is already December.

Samsung Fab

There are only two or three leading-edge semiconductor manufacturers in the world, depending on just where you put Intel in the list. One of those is Korean-based Samsung. It has a fab in Austin TX (I believe the biggest one in the US) and it just announced that it will build a new one in Taylor TX, which is basically Austin.

Here's the article in the WSJ Samsung to Choose Taylor, Texas, for $17 Billion Chip-Making Factory. One slightly weird paragraph:

The Taylor factory is expected to serve as an advanced chip-making facility for Samsung’s contract-manufacturing operations that make semiconductors designed by other firms. Such high-end manufacturing is attracting the bulk of semiconductor industry investment. The types of chips with the longest backlogs tend to be lower-priced and haven’t been the focus for massive expansion.

Translation: it is expected to be primarily a foundry for leading-edge (at this point, 3nm and below) processes requiring EUV. For more about that, see my post Samsung Foundry Fab Roadmap.

Silicon Laser

We have our photonics workshop coming up. See my post 6th Annual Photonics Summit Coming in December. One of the issues in photonics is that silicon is great for detecting and modulating light, but no good at producing it. So it always requires an off-chip laser. Well, McMaster University (that's in Hamilton, Canada) just announced Seeing the light: Research breakthrough by McMaster PhD student creates a simple, cost-effective laser on silicon. This seems significant, although it is hard to tell with early announcements, especially reading the PR rather than the paper. And even if it is good, it is years away from production at scale. The lead of the project is Khadijeh Miarabbas Kiani.

Her PhD supervisor is Jonathan Bradley, an assistant professor in the department of engineering physics. He gave his opinion:

This is the holy grail of photonics, Fabricating a laser on silicon has been a longstanding challenge.

Ford and GlobalFoundries

I've been saying for years that more automotive manufacturers will design their own chips. Just as pretty much every automotive manufacturer designs their own engines, since that is the heart of the car, in the future cars will be going electric and semiconductors will be the heart of the car. Ford recently announced a partnership with GlobalFoundries. it is not quite clear how much this is about designing future chips, and how much is locking up some production volume in the current chip shortage. Here is the WSJ Ford, GM Step Into Chip Business.

Money quote:

Ford on Thursday morning outlined a strategic agreement with U.S.-based semiconductor manufacturer GlobalFoundries Inc. to develop chips, a pact that could eventually lead to joint U.S. production.
...
“We feel like we can really boost our product performance and our tech independence at the same time,” said Chuck Gray, Ford’s vice president of vehicle embedded software and controls.

That's pretty vague, and the GM part is vaguer still:

GM later said it was forging ties with some of the biggest names in semiconductors—including Qualcomm Inc. and NXP Semiconductors NV—and has agreements in place to co-develop and manufacture computer chips.

The WSJ thinks that this is pandemic-related, but I think it is a longer-term trend that will last a lot longer than the current automotive chip shortage. Automotive manufacturers are not great customers for semiconductor foundries: low volume, very price sensitive, require extra costs for reliability, and so on. There are about 100M vehicles built annually, but there are about 15 major companies (and something like 50 brands). That compares to 1.5B smartphones per year. So automotive manufacturers have to create deeper partnerships if they are going to get help and capacity for their own chips.

Graviton 3

Amazon AWS (through Annapurna) builds its own Arm-based server chips. See my posts EDA on AWS Graviton and Liberate Trio on AWS/Graviton2 Instances. Well, earlier this week, AWS announced Join the Preview – Amazon EC2 C7g Instances Powered by New AWS Graviton3 Processors. Money quote from the blog post:

I am thrilled to tell you about our upcoming C7g instances. Powered by new Graviton3 processors, these instances are going to be a great match for your compute-intensive workloads: HPC, batch processing, electronic design automation (EDA), media encoding, scientific modeling, ad serving, distributed analytics, and CPU-based machine learning inferencing.

"EDA" in there is not just fluff. Graviton 3 was designed on AWS/Graviton2. See my post Climbing Annapurna to the Clouds. What's the performance:

While we are still optimizing these instances, it is clear that the Graviton3 is going to deliver amazing performance. In comparison to the Graviton2, the Graviton3 will deliver up to 25% more compute performance and up to twice as much floating point & cryptographic performance. On the machine learning side, Graviton3 includes support for bfloat16 data and will be able to deliver up to 3x better performance.

It also consumes 60% less power than Graviton 2. I'm guessing it is built in 5nm or 4nm.

Adversarial Image Attacks

I've written before about adversarial image attacks. See my post Fooling Neural Networks. This has been considered somewhat of a joke for several years, but a recent article Why Adversarial Image Attacks Are No Joke points to a major security weakness. Many neural networks are trained on the same image datasets (in particular Imagenet, see my post ImageNet: The Benchmark that Changed Everything), but also traffic sign databases, and cityscapes for training autonomous driving. If you know the database in advance, it is easy to craft examples that fool the image recognition (changing Reese Witherspoon into Russel Crowe with a carefully crafted pair of glasses, or a 30mph speed limit into a stop sign).

The reason that this is a serious issue is that it is expensive to fix and will become more expensive as these neural networks are deployed at scale:

It’s something that would be difficult (and very expensive) to fix right now, and which would be unconscionably costly to remedy once the current trends in image recognition research have been fully developed into commercialized and industrialized deployments in 5-10 years’ time.
...
Adversarial image attacks are being made possible not only by open source machine learning practices, but also by a corporate AI development culture that is motivated to reuse well-established computer vision datasets for several reasons: they’ve already proved effective; they’re far cheaper than ‘starting from scratch’; and they’re maintained and updated by vanguard minds and organizations across academia and industry, at levels of funding and staffing that would be difficult for a single company to replicate.

UPDATED UPDATE: Arm and NVIDIA

In my post yesterday, Today Is the 40th Anniversary of the BBC Micro...and the Ancestry of Arm, I wrote:

NVIDIA is in the process of acquiring the company from them, although it seems to be tied up in various regulatory issues for now. So watch this space.

You didn't need to watch for long. Today, the FTC has challenged the NVIDIA/Arm acquisition on the basis:

the acquisition would give Nvidia unlawful control over computing technology and designs that rivals need to develop their own competing chips.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.

.