Meta's GB{2,3}00 Cluster
engineering.fb.comIt seems that Meta didn't mention one aspect of their "next-generation" AI training system: their incompetent web crawler hammers websites while completely ignoring robots.txt. Meta's 57.141.0.0/24 made up about 90% of all requests to my server this past week, and it didn't stop after I blocked it. They continue to make a request every 2 seconds despite me responding with a 403 every time. Truly an engineering marvel that Meta has bestowed upon us.
edit: I just checked and they are back to hammering my server at 105 requests per minute. These are the Meta-owned ranges I'm blocking now, in case it helps anyone: 57.141.0.0/16, 57.142.0.0/15, 57.144.0.0/14, 57.148.0.0/15
I've blocked Meta, OpenAI and Anthropic and yet they seem to find ways to get back in. It's pretty annoying.
nft add table inet filter
nft add set inet filter meta_crawl '{ type ipv4_addr; flags interval; }'
nft add element inet filter meta_crawl '{ 57.141.0.0/24 }'
nft add chain inet filter input '{ type filter hook input priority 0; }'
nft add rule inet filter input ip saddr @meta_crawl tcp dport {80,443} reject with tcp reset
I did finally block them at the firewall level last night. They are constantly trying to connect anyway.
When you're a celebrity, they let you do it.
Poison the responses with massive ai slop dumps.
Make them block you.
maybe just drop them rather than reject with tcp reset
> incompetent
I think that's not incompetence, I think that's by design.
How? What's the incentive structure for that design? There are no extra ads to be displayed here, so how does Meta get value from that design? Seems way more likely to be incompetence in my world view. Explain why that design is superior!
Someone puts a high number of pages crawled per minute in their PSC. Doesn't matter if they're 403s.
There are no extra ads to be had even by scanning sites with a robots.txt that allows them, because the value of pages could never have been in displaying ads but must have resided in the information on the pages itself.
With Muse, people can enter their bank account information and medical information so that it can work hard for them. For this, they need a lot of training material.
LLM training data
Reverse Hanlon's Razor
(Rhetorical question) Who would you trust scraping your website? Google, for who among all their issues I haven't seen complaints of "Google is ddosing the web", both pre and post LLM
Or Meta, run by someone of Zuckerberg's character that wants to dominate everything and it shows by how many requests per minute? He's one of the reasons sites now have to implement bot blocking technologies
bet it was written in PHP
I didn't want to have to put everything behind cloudflare, but this sort of nonsense made it the least bad path forward for us.
Posting this two days before earnings seems to indicate that there is going to be some bad capex news.
-10B on VR, -(The Compny's Future) on AI
Ya dont say.
Negative Free Cash Flow are the words every one on HN needs to learn.
Every thing before now, you will look back upon as the happy time. When every year storage, ram and compute grew and the prices went down. Your cloud provider made so much money they paid cash for hardware.
Now they borrow at 8% and net margins on ram/storage were around $0.10/GB ... today they sit at over a dollar. (SEE micron's s10's if you want to check the math on that).
Forget circular financing, its a circular firing squad because your going to have to figure out where the hell your going to run any of your gear.
4x to 10x for ram +8% for financing +30percent AWS margins. Good luck everyone its gonna get rough out there.
Or hyperscalers revenue will continue to disappoint, there is no $60 trillion market for ai and there will be a glut of capacity for years to come!
The capital (heh) mistake in the whole assumption that these companies are worth that much is that they'll crash the economy in their zeal to get there. And when the economy goes down the drain so do the expenditures because in the end the companies that could spend money on AI are taking that money out of the economy to put it in the pockets of a tiny fraction of society. You can only do that if you are willing to leave the rest without food on the table and that sort of move is simply not very wise.
There is definitely a $60 trillion market for AI, just not by next year.
I am very lucky to have stocked up on RAM before this happened (256G) but I wished I had doubled or quadrupled that, an upgrade simply isn't going to happen now.
> The Ariel variant of GB200 NVL72 pairs each NVIDIA Blackwell GPU with an NVIDIA Grace CPU and 2x NVIDIA ConnectX-7 400G NICs at a 1:1 CPU-to-GPU ratio, letting us deploy 800G of I/O per GPU earlier than the rest of the industry
If the 1:1 ratio was purely for I/O (big assumption, very likely wrong), I am curious why they did not just go with ConnectX-8? Other hyperscalers have been shipping NVL72 clusters with ConnectX-8 on roughly the same timeline, and it offers the same 800Gb/s of throughput.
Managing a custom Ariel variant (which seems to be a Meta-only build) is not insignificant complexity, I wonder if they had another need for that 2x CPU-GPU ratio in their workloads relative to the stock GB200s.
EDIT: I looked, the gb200s NVL72s never have shipped with CX-8s, I made this up. This decision could have totally been due to compute availability and timing
IIRC one of the analysts said 1:1 CPU:GPU was for recommendation algorithms and 1:2 is better for LLMs.
Actual title: ”Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster”
Is this one of the pilot models they claim R&D tax credits for?
shell glob rather than RE? what's the world coming to!
Imagine your job is just replacing millions of Dollars worth of hardware every day. I mean there must be a reason the hold 8 GPUs per rack spare.
Yes the reason is they're pushing the GPUs hard so that they fail, and having hot spares saves the job until helping hands can go replace the bad component.
Theres a ton of people working at (neo)clouds doing exactly that