from Hacker News

Bay Bridge: the cheapest H100 training clusters

by flaque on 11/14/23, 12:01 AM with 8 comments

• by dang on 11/14/23, 4:43 AM

  It's a super cool project but this post is not close to being a valid Show HN. Please read the rules: https://news.ycombinator.com/showhn.html. (The submitted title was "Show HN: The World's Cheapest Supercomputer", which also broke the HN guidelines by being baity.)

  Not only that but you had a huge Show HN just 3 months ago: https://news.ycombinator.com/item?id=36933603, making this repost a dupe by HN's rules (https://news.ycombinator.com/newsfaq.html) - and not only that but the earlier post, being a signup list, wasn't a valid Show HN either. Evidently we missed that.

• by neilv on 11/14/23, 3:30 AM

  > We'll sell you just for a month, for about $2.5m. It's more per-hour than what most folks will do, which doesn't always make it a great choice for inference. But for training very large models, it's a company-defining purchase.

  I loved the "company-defining purchase" bit of this copy. It's like they're persuading me how many months of salary I should spend on an engagement ring.

• by solarpunk on 11/14/23, 3:25 AM

  >We were trying to make "midjourney for music", everyone told us the minimum purchase was all the money in our bank account plus at least two of our limbs.

  i hate to say this gives off a 'failed prospectors pivoting to shovel sales' vibe that i find concerning.

• by markhahn on 11/14/23, 5:45 AM

  kind of weird crack about gpu and ib failure rates. can anyone expand on that?

  (I'm familiar with several large HPC clusters that have low gpu and ib failures)

• by latchkey on 11/14/23, 4:27 AM

  "No availability for 2 nodes for 1 week."

  For the 1 week, clicking "Purchase" just gives console errors and doesn't do anything.