Hacker News Reader

Comment by Twirrim | Hacker News Reader

Titans are $1.5k each, so that's $6k down before you even account for the rest of the hardware to run it. Ouch.

The whole machine off the lot is going to be less than a drywall contractor's Ford F150, yet potential payback is many times higher while the operating costs are several orders of magnitude lower. Throw in the three year depreciation and the box is bargain so long as it can be put to work.

The whole machine off the lot is going to be less than a drywall contractor's Ford F150...so long as it can be put to work

Yes. You're a drywall contractor you go to Craigslist and start humping some jobs. Make some bucks and buy you a nice F150.

Conversely, there aren't a lot of "need deep learning in my dentistry" ads on CL -- and not as easy or clearcut path from point A to point B.

So the investment makes sense, if the right conditions are true. To the average developer, those conditions look very murky and uncertain to gauge. So while it's obviously some sweet tech, but it takes a lot more than tech.

I don't disagree. The price is the price because the price makes sense for businesses. Prosumers may buy one for the same reason pump and valve salespeople buy 911 GT3's...to drive something fast slowly. It's on a continuum the drywall feller ordering the 6" lift and 34" tires, leather seats and dual climate control. All that he really needed was the four wheel drive and the diesel V8.

Which BTW, the drywall impresario who's buying a new truck for his business isn't finding jobs on Craigslist. He's got business relationships with serious people who call him when a bid needs bidding and drywall needs hanging. It's deal flow just as it is for the sort of person who needs a 4 GPU box and CUDA code for their business. It's only a lot of money if it sits idle.

If there isn't a business case, there isn't a business case and buying it is an inefficient allocation of resources.

On the other hand, there are lots of researchers twiddling their thumbs waiting for code to run right now. The bet is that this activity will continue to grow quickly.

Of course this isn't for the average developer. Which is why they are marketing it to researchers and maybe AI startups.

The boxes will be $15,000 USD. Lead time is 8-10 weeks.

Poster above says the GPUs cost $6k total, and cases that can fit 4GPUs aren't that expensive nor elusive. So I guess the rest of the system is made from unobtainium, or they figure deep learning is the kind of red hot topic that attracts enough suckers with money that this will fly.

Or people for whom the money value of their time exceeds the margin over a custom build that nVidia is asking.

I can spec out a system like that in less than 1h, but let's say it takes 2h, ending up with a total hardware cost of ~$9k plus $200 to have it assembled. Are you really valuing your time at more than $2000 an hour?

If someone's willing to pay, sure! But more realistically, a ML researcher doesn't want to know the differences between Haswell-E and Skylake, or which DIMMs are rated at which speed, or what's the optimal Linux driver situation, still less to know how to fix any of the myriad things that can go wrong. When Caffe starts screwing up, it's really nice to be able to call nVidia and make your overheating Titan their problem.

I'm down with you about calling support when your hardware craps out. But an ML researcher who doesn't know the difference between Haswell-E and Skylake, etc., will likely make suboptimal choices when using the hardware. Using GPUs efficiently requires in-depth understanding of how they function and how they interact with the rest of the system.

This is true, and also, I think it's perfectly likely that nVidia is trying to soak people buying with other people's money.

Reply on news.ycombinator.com