Hacker News Reader

Comment by nuz | Hacker News Reader

Meta has their own data centers so they definitely didn't pay the equivalent to what AWS costs

Good point, although it's possible that with the extreme price of GPUs that it cost more to train by buying hardware than it would to rent. For example it might take two to three years before the GPUs are paid for by customers.

I think AWS prices scale with hardware price

Linux reserved cost of p3.16xlarge is $146,362.0800 annually. On-demand cost is $214,444.8000 annually

I am pretty damn sure I could build a 8 GPU Intel Xeon E5-2686 v4 (Broadwell) (that's what Amazon uses - it's $30 to $75 on eBay) server for less than that and come out ahead on electricity even at full throttle. RTX 4090 are just under $2000 each on eBay.

8 GPU × $2000 (RTX 4090) + $1000 (for the rest of the computer) = $17,000

If pulling 2kW continuously at 15 cents per kW*hr for 1 year that's 2000 watts × 365 days × (0.15/(kW×hr)) or $2,628

In total the computer will cost $19,628 if you throw it in the dumpster at the end of each calendar year of using it.

If you stack internet cost of $200 a month on top, that's $2400 a year, which raises your annual cost to: $22,028

This is still $124,334 cheaper per year than one AWS 8-GPU server if you fully depreciate your own hardware at the end of year 1 to $0.

I could hire an engineer in America to babysit it with the money left over.

An A100-80 GPU goes for about $20K each.

The instances in question use Tesla V100-SXM2-16GB

Are consumer grade RTX 4090 cards going to be suitable for running full tilt 24/7 for a year? Those things are fine to stress on the latest game for a few hours at a time, but would probably cause some defects from significant heat stress after just a few days at 100%.

This is inconsequential when you're playing Overwatch for a few hours a night and a frame drops now and again. If you're training an iteratively developed LLM though, physical defects could propagate into huge deficiencies in the final model.

I don't think they'd become a fire hazard, but it is true that one would likely pick something else for this application.

Having said that, switching to something like the Tesla V100-SXM2-16GB wouldn't cost that much more.

TBH, I'm shocked at how many people treat Amazon as the first choice for this stuff. Much of it isn't even what most would consider a "production" workload. You are paying for a lot of enterprise-readiness that you don't need for training.

If you wanted to finetune a Mixtral 8x7B, what would you use?

Given the relative availability, I'd probably try to do it with a couple of rtx4090s on tensordock.

> TBH, I'm shocked at how many people treat Amazon as the first choice for this stuff

You can thank Amazon's legions of salespeople for that, particularly the end of year junket in Las Vegas where attendees are so pampered that about the only thing they won't do is suck your dick

Oh, yeah, they'll also yell at you on stage if you complain about their UI

Yep absolutely, crypto miners have been doing it for years.

I still think it would be impractical at scale because they are so much more hot and power hungry than the datacenter cards, and you would be lucky to score one or two if you’re on a wait list.

Except you can absolutely obtain 4090s today, while enterprise hardware is (was? haven't looked at the data) recently, which is the exact opposite scenario you mentioned.

I'm actually really surprised that you can still buy 4090s for under $2,000 (cheapest available I saw was $1,800 new and I only took 30 seconds to look), but you can usually sell certain models for quite a bit more. For example, my used 4090 FE is currently worth more than I paid for it.

I've played with AI, and while admittedly I've not done anything super serious, I can tell you that both the 3090 and 4090 are more than capable of performing. Tie them with a power efficient AMD CPU and you have something that can be competitive with enterprise (somewhat).

I've seen the pricing of "cloud" offerings and I've toyed with the idea of creating an "AI Cloud" because I have access to really fast internet and super cheap electricity, but I haven't executed because I'm most certainly not a salesperson. I do, however, know enough about marketing that one should not target price, so there is that...

You could under-volt or watt-limit a bit and lose just a fraction of FLOPS for much less heat/power though, depending on the workload

You can build old Xeon based but only has 40 lane PCIe. For training 8 GPUs how do you push data fast? I’m using 7000 series Epyc for this to get 128 lanes. Have you built this kind of machine? You see good speed with 40 lane? Curious because then I can use old Tyan motherboard which comes in full case with good layout for multi GPU. Epyc based I have to use riser and custom frame which is painful.

New Tyan more costly but great case layout.

Though this comparison is really only relevant for a couple of machines. Beyond that, at this cost, if you pay AWS list prices "at scale" you're doing something very wrong.

Don't get me wrong - I've frequently argued that AWS is price gouging and relying on peoples lack of understanding of how the devops costs of running your own works out, but it doesn't take a huge budget before this calculation will look very different (still cheaper to own your own, though).

Since the GPUs can be rented out afterwards, they amortize very quickly with prices in the order of $1/h.

Meta also doesn't pay AWS anywhere near retail price for instances.

Why is this the case? Even AWS internal pays the same AWS prices as everyone else

They have their own data centers, they don't use AWS

They don't use AWS to run the big three apps (FB, Insta, WhatsApp). They very much use them for other things. FB data center machines tend to be highly-specialized and optimized for running the apps, not general-purpose compute.

I'm less surprised if AWS internally pays AWS list prices, because that's just internal accounting. From the even relatively small AWS customers I know, none of them needed to get very far into the 6 digits per year spend before a couple of quiet mentions to their account manager that they were reviewing other options was enough to get steep discounts.

Add in lots of credits, and if you pay list price, you're being taken to the cleaners..

I've done contract work for clients to be ready to migrate both as part of maximising credits and as part of negotiating posture, and the savings can be enormous (though it'd still usually be cheaper to use managed servers).

Couple of reasons: Meta's annual spend with AWS is large enough that they'll have a negotiated blanket discount that takes a fixed percentage off the top of their monthly spend. This is very common for larger AWS customers, not just Meta.

For instances specifically, any planned usage will be using either reserved instances or at minimum a compute savings plan (CSP) that drops the hourly rate dramatically in exchange for a committed number of instance hours, with or without an upfront payment.

Finally, there may be a negotiated rate for specific instance types built into the contract. Again, common for very large customers.

source: I was on one of the cloud-related infrastructure teams (left in early 2022). I have no idea about their spend (or discounts) today, but two years ago it was enough that Andy Jassy would meet 1:1 with Mark to "discuss the relationship".

Reply on news.ycombinator.com