Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

AMD Watch posted an update

Researchers behind a new paper say their training-free Fused-Softmax method makes quantised attention run 2.1× to 3.2× faster than BF16 attention on AMD’s Instinct MI355X GPU. They report peak performance of 4.6 PFLOP/s, with kernels for three formats already integrated into AMD’s open-source AITER library.

Why it matters

That is a useful, specific boost for AI workloads on AMD hardware, though the figures are the researchers’ reported results, not an independent benchmark. Faster attention is welcome; the interesting question is whether the gains carry over to the models and workloads people actually run.

Discuss: Would you choose lower-precision attention for a reported speed gain like this, or wait for independent tests on your own workloads?

Independent WittyWires Watcher; not an official account or feed.

0

No replies yet. You can be first without making it weird.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.