Pretty cool! Which Google office do you work from?

I work in the Bangalore office. I lead the AI parts of the chip here

Ahh, that’s nice. I work closely with Google’s PyTorch XLA team

Noice!

This is pretty cool

@creamy-sundown-74328 I can't seem to find any details on the SoC itself. ARM? Cortex-AXX or custom micro-architecture? About time Google did this.

People who are really serious about software should make their own hardware - Alan Kay

🙂

Yeah the details will be made available pretty soon. It has a mix of both, a bunch of different ARM and various other IPs as well as completely custom IPs for various components.

Every pixel phone released so far had one glaring flaw, looks like this year would be an exception. 🤞

@creamy-sundown-74328 I'm guessing the GPU and media accelerators were licensed (or did Google buy up a company I missed?) while Google focused on the interconnects, NPU and possibly the micro-architecture for this first version. Looking forward to seeing details.

Graphcore IPU roast ☠️ https://semianalysis.com/graphcore-looks-like-a-complete-failure-in-machine-learning-training-performance/

@helpful-gigabyte-47939 All info pretty soon :p I’m not allowed to steal marketing’s thunder as yet. Though can talk in much more details once they send out more info publicly :)

https://twitter.com/geekyranjit/status/1422516136000643077 Looks like it ain't coming to India 🥺

Pixel 5 not releasing in India is still justified due to it using radar tech which isn't permitted for consumer goods in India. But 6 isn't coming (allegedly) is a bummer. 😬

Why do you guys end up calling everything tensor? 😅

Because tensors must flow 😛 In general though, most of our AI-primary products use this as a cohesive branding. The new chip is also trying to differentiate from incumbents in the same space, so that's the reason for marketing to tie it to the same. (PS: The AI connection of this development is huge compared to what is apparent from the mobile chip announcement. We have built it in a way that it scales from tiniest of IoT things to the largest of data centers)

built it in a way that it scales from tiniest of IoT things to the largest of data centers

aka. expect MCU-class SoCs with this technology to enable edge-AI use cases. Marketing thunder be damned. 😉

Not really. Those things have some other usecases for us internally. No consumer products as such

Please add support for PyTorch on Coral 👀

Yeah that's been one of the wishlist items internally. But unfortunately it's been on the low end of priority so far.

@creamy-sundown-74328 I already have a coral usb device but it didn't give us the performance we needed. We ended up with Nvidia Jetson which was far better bang for buck. But I've had my eye on the coral accelerator module for future HW designs though.

Yeah coral was our gen1 product in the lowest configuration. We may have boards for new ones as well but not finalised yet