perfobotto

So difficult to read ..

Maybe OpenAI or Anthropic

It’s very simple. Apple doesn’t pay as well as other places and they stock price increases and their flashiness are not enough anymore to keep people around. It’s time they realize they are not the darling they used to be

No way an offchip HBM has same or better bandwidth then onchip

is this approach viable on training? wat about kernels that require different grids?

What an amazing write up! Thank you very much!

To be clear a cerebras chip is consuming a whole wafer and has only 44 GB of SRAM on it. To fit a 405B model in bf16 precision (excluding kv cache and activation memory usage) you need 19 of these “chips” (and the requirement will grow as the sequence length increases for the kvcache). Looking online it seems on one wafer one can fit between 60 to 80 H100 chips, so it’s equivalent to using >1500 H100 using wafer manufacturing cost as a metric

This is supposed to be a product , not a research artifact.

What stops nvidia from making sure this stops working in future driver releases?

To be honest restricting funding to the kind of symbolic based AI research that is criticized in this discussion might have helped AI more than it hurt , by eventually pivoting the research toward neural networks and backpropagation. I don’t know how much of a good thing would have been if this kind of research continued to be funded fully.

I mean, he was right … for what we knew at that time. He predicted correctly that the only way to achieve a general intelligence it would require to mimic the extremely complex neural networks in our brain that the hardware of the time was very far away from achieving. He could not predict that things would move so fast on the hardware side (nobody could have)that made this somewhat possible. We are atill I would argue a bit out in having the appropriate computer power to make this a reality still, but it now is much more obvious that it is possible if we continue on this path

Snaaaake

Whenever I see a product that advertises itself based on the technology it is built on I’m skeptical. Rust doesn’t make a game engine good.

Am I the only one not understanding what’s the point of all these efforts to try to run Linux on completely closed with no documentation devices (like iPhones and even mace for that matter). Seems like a lot of effort, for a crappy solution that maybe 0.0001% is going to use before realizing it is crap. Even only effort to keep up with the yearly hw updates that apple does and the complete disregard for backward compatibility and writing hw documentation makes these effort just a technocrat exercise (very cool for sure), but I can’t see anything more than that

This seems specifically designed to make the us look bad. If you try to compare us and Libya it almost seems like it would be so much better to move there which is stupid. Statistics are calculated in such a way that is just unfavorable for the us