virtuallynathan

How does a DGX Pod w/ the new 3.2Tbps per machine NVLINK switch compare to Tesla Dojo?

RIPE Atlas tries to do this using their IP Map. It sorta works. https://ipmap.ripe.net/

> The surprising part was that it took the science guy 10 hours to calculate the calories for 5 food items ? Would be nice if we could get that down to minutes and accessible to personal consumers.

Bomb Calorimetery is a PITA, I almost did a startup around this. The researcher in the video has a fairly manual one, there are automated setups that reduce this time, but they still cost $40k+ and require oxygen tanks, etc.

The even more difficult case for calories is something like a steak: how marbled is it? I could see some computer vision approach working.

I'm a nerd with more money than sense, what can I say! I trust the output of the device I have, and I do truly believe I am burning 3300kcal/day just standing at my desk, walking to the bathroom, etc. I eat 3500kcal/day pretty regularly, and maintain weight.

If you can target enough of a deficit and not be hungry, good for you. It's hard for most people. The input data is wrong, the output data is wrong/unknown, and one can only be hungry for so long.

Do you know how many calories you burn? Do you actually know how many calories are in your food?

I have a metabolic cart, and the answer surprises me sometimes on the former - just standing at my desk all day I burn >3300kcal. I actually went to talk to the guy in the Casey Neistat video about food labels, and almost bought a bomb calorimeter. Calorie counting, while not useless, is not very helpful.

https://youtu.be/hE2lna5Wxuo

Yes, the dish tends to change satellites every 15 seconds, targeting would probably be a nightmare.

RIPE doesn't really have the power to withdraw assigned IPs, the Russian ISPs can just keep announcing them. Perhaps they could break their RPKI data so their prefixes would be dropped as invalid by those checking signatures (I would consider this a misuse of RPKI), but they would still mostly work.

Hamilton Morris has commented on this paradox of drugs...

People who take Amphetamines because they want more energy end up tired, etc.

There's no other way to know what they'll do.

Now there's a real source! "Amazon Web Services owns 11.9 million square feet of property, leases 14.1 million square feet"

That's 26 million sq ft vs Facebook's 40.5 million.

Personal intel, I maintain a map and spreadsheet, combined with publicly available information and press releases.

I follow Starlink's IP space and connectivity closely, if anyone is interested, see this: https://old.reddit.com/r/StarlinkEngineering/comments/ofly6k...

Not anymore, customers are being migrated to AS14593. I'm one of em.. I'm in 170.203.91.0/24.

Based on my calculations, Facebook has more DC space/power than Google/Amazon/Microsoft. They have around 40 million sq ft of DC space, and likely ~5GW of power. They have 48 buildings in 18 locations, with 47 more under construction.

AWS Data centers are fairly small, ~180,000 sq ft or so generally, us-east-2 is about 900,000 sq ft. us-east-1 perhaps 8 million sq ft.

Google likely has more network capacity, but FB is catching up.

100GbE NICs are <$500, and 100GbE switch ports are $100-150.

I’ve never had this as a problem, and I use 50-100ft of Cat5e + 10GBaseT SFP+’s… it’s well within the thermal limits of the SFP, which probably uses 2-5W.

The actual machines aren’t $60k anymore - I have a portable one, it was $5500. This is a very cool project to bring that down even more.

Yep! All TLS.

The production boxes are 2U, but it can be done in a 1U box.

At Netflix we’re doing close to 400Gbps on 1U commodity hardware, and pretty inexpensive.

Arctic routes have been proposed for a long time, I hope this finally happens.

2014: https://spectrum.ieee.org/arctic-fibre-project-to-link-japan...

I’m not sure where he got the idea that these don’t exist… I see them all the time: https://containertech.com/container-sales/53ft-high-cube-sto...

It came up on Joe Rogan #1742 last week, probably why it made it here.

Hard to track for other people’s VMs, but they probably have (or can sample) that data for every AWS-operated service (dynamo, S3, redshift, etc..)

No, this isn’t for video data, those tend to just be stored on S3. EVCache is like memcached for small data.

http://isomerdesign.com/PiHKAL/ Has a copy of the book, basically.

I would say very few in the medical community would even agree to that

There are plenty of risky activities one can legally partake in… are we just going to let those people die when they are injured?

They aren’t all public globally — there seems to be at least one in Austria based on this data, the location of which is presently unknown.

I think its fine for this to occur, and we should expect it, but how long should it take to reverse a bad practice? Is a decade or more acceptable?

Sure, in an ideal world we’d have some kind of review, but what we have now seems largely ineffective.