Mellow

Poor Writer. Frequent Purger

[Mellow @dbierce]

@TalosDreams @InsideEVs Indeed. What I used to do in 1 stop I now takes 2. 3 stop if its -20F out which makes 4 months of the year travel fun.

[Mellow @dbierce]

@AndrewH43114926 @Kadotttt What do you mean? Total return? I assume you mean only in direct payments. Since measuring all of the other costs doesn't translate easy to everyone.

[Mellow @dbierce]

@fugitivemama If you're down coolant, it means you have a leak. Leak means you need to get it fix as the amount of coolant ever leaving the system should be 0

[Mellow @dbierce]

@asymmetricinfo Difference between wealth and income?

[Mellow @dbierce]

@J_Von_Random He's just mad he bought FSD twice and was told his car won't support it a few years later.

[Mellow @dbierce]

@BellikOzan @EmperorBag What do you mean? Less than 1% cost to launch? Total TCO?

[Branden Flasch @brandenflasch]

This is the wrong way to look at EV range. Range gives you flexibility to stop where you want to rather than where the car can charge.

[Mellow @dbierce]

@peterrhague @speakjava Spinning magnet nuke or thermal decay? It's a different use case for rejecting heat. Few thousand degrees from a reactor is more of a problem with scale not precision

[Mellow @dbierce]

@ajWasThere @ManaByte Latency and data locality is a good reason to pay a premium for closer POP access. What do you think is low latency? Workload in the satellite count be cool but it's still 20ms away which is forever for most workloads.

[Mellow @dbierce]

@klaehnr @GravityAnalyti1 Current satellites are 2kW NASA studied the crap out of this and they get around it by paying the weight and bringing extra
NASA has lots of studies. TLDR more gates nm means bring 3 computers they gonna die. Current AI workloads need a shared clock 1ms is an eternity

[Mellow @dbierce]

@DrPhiltill Today I learned about perlite. The name for that clay stuff.

[Mellow @dbierce]

@klaehnr @GravityAnalyti1 If you can live with the high MTBF from heat and hi energy physics, non repairability and high latency. They better have a really cool use case for that kind of expense.

[Mellow @dbierce]

@jimstewartson The IPO requires you to be excited for this technically possible, but OMG the engineering and cost which includes finishing the launch platform. The cost has to fall.....and the workload has to live with the latency and small power footprint.

[Mellow @dbierce]

@TheWizardTower @RadioFreeTom 150kW is a rack not a datacenter.

[Mellow @dbierce]

@yatharthmann It's a question about cost (it's expensive) and size 150kW is a Rack of compute. There might be a workload for that, but it isn't a lot of compute. You know you can throw up 4000kW in a structure in 6 months. if you can get parts.

[Mellow @dbierce]

@VaibhavSisinty Oh man, this is gonna make so much IPO

[Mellow @dbierce]

@ghost_motley Yeah, H100 runs at 75 but gets a high MTBF for your troubles. Not just the H100, all the supporting stuff too :(

[Mellow @dbierce]

@Freedom_44_ @PTrubey Nah, HW5 clustered.

[Mellow @dbierce]

@ajWasThere @ManaByte Depends on your workload. You don't always care about >5ms, but for edge applications, seems like an expensive way to get closer.

[Mellow @dbierce]

@ot_giwa @steph4n What's the workload these are for? It can't be that demanding, thats only 200 H100s of budget. (as a mesaure) they have to have custom workloads and can't be using off the shelf silicon even in LEO.