ismailmaj

Feels AI-generated, so much text real estate is spent explaining the basics in sufficient detail then just two paragraphs on their scheduler with a lot of hand-waving.

Given the size of the model, that's very aggressive pricing.

I wonder if it is serving at a loss to get training feedback (maybe using the margins they get from GLM 5.3 serving which is a slightly smaller model but double the price).

They control the supply chain though, so they can't be price-gouged by Nvidia

they can pivot to foods: Origan, Carbonara, ....

After all they had Nano Banana for years.

> For ultrafast latency, only activations need to be passed between the wafers, and this can be done via RDMA over Ethernet.

I believe it for decode but the article says nothing about how they make prefill fast on that interconnect story :/.

Still 9 days before the release of weights apparently.

Amazing news, I remember the days when you had to fight XLA to make PyTorch run on TPUs, it seems we finally got first-party support!

Since HW specifics live mostly in the kernel layer (Pallas or CUDA), I imagine the Jax vs PyTorch question is more of a style preference now?

Schools got better at teaching programming but smartphones replaced the PC culture.

I don't think I'd be currently doing software if YouTube and games were accessible and good on phones back then.

Not really, well depends on the price point but it would make more sense given enough users to multi-GPU setup with good interconnect between them to do expert parallelism.

Reminds me of the early 2000s era where we'd fight in my household if too many people used the internet at the same time, can see that happening here if it's not 2 to 4 users but 10.

1.5 TB/s 1.3 PFLOPS with no fast multi-GPU, it's a consumer card that pretends to be a server card.

It would work great on Qwen3.8-Flash-Next or GLM-5.3 as single user decode only hits a few experts, anything with concurrent users and you'd be sitting for a while.

Fun to see the LLMism on the product page: "A, not B"

I don't think it needs higher VRAM, the whole point is horizontally scaling.

Most models have moved to MoE and expert routing which is done over infiniband, so tensor parallelism where Tenstorrent was really weak at isn't used much, I can see it as a viable alternative where latency isn't a concern but cost is.

Also even if they wanted to increase the capacity on their GDDR, they wouldn't be able to since they are limited by the GPU memory channels.