Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Inference cost at scale with napkin math

injuly.in

Post reply on HN
Print view View original

11–20 of 20 posts

Re: Inference cost at scale with napkin math

#11

Post by martinald »

BadBadJellyBean wrote: post #9
JBAnderson5 wrote: post #7

> Realistically, serving 300 users per GPU you'll spend a lifetime cost of about $133 per user, plus the datacenter/upkeep bill. What is the operational cost and when does it become more expensive than the upfront capex? The B200 tops out at 1000W and idles around 140W. It averages around 600W. https://www.lightly.ai/blog/nvidia-b200-vs-h100 U.S. average electricity cost is $.14 per kWh in March. https://www.eia.gov/…

I wonder what the power costs are when you put jet turbines in front of your DC to power it.

In general, less for fuel cost alone. But you obviously need to buy the turbines.

Re: Inference cost at scale with napkin math

#13

Post by martinald »

breput wrote: post #8

> We'll assume a 32B dense model, as they've have gotten quite good for production use and a B200 can comfortably serve them. This could be a Gemma, Qwen, DeepSeek, whatever. That seems like a very consequential point to include halfway through the post. They aren't wrong that Qwen 3.6 26B or Gemma 4 31B are quite good, depending on the use case, but if we're doing napkin math, I'd want some more headroom in the assu…

Yes 32B dense is a weird one to choose.

But in reality, 32B dense is very similar* to 32B activated on MoE in terms of inference costs. And I highly suspect eg Opus is around that level of active params.

A 284ba13b model at scale, is almost certainly cheaper to serve than a 32b dense model.

*as you can shard the model across multiple GPUs at scale. but in reality you have some loss of efficiency from GPU coordination and expert routing

Re: Inference cost at scale with napkin math

#14

Post by breput »

martinald wrote: post #13
breput wrote: post #8

> We'll assume a 32B dense model, as they've have gotten quite good for production use and a B200 can comfortably serve them. This could be a Gemma, Qwen, DeepSeek, whatever. That seems like a very consequential point to include halfway through the post. They aren't wrong that Qwen 3.6 26B or Gemma 4 31B are quite good, depending on the use case, but if we're doing napkin math, I'd want some more headroom in the assu…

Yes 32B dense is a weird one to choose. But in reality, 32B dense is very similar* to 32B activated on MoE in terms of inference costs. And I highly suspect eg Opus is around that level of active params. A 284ba13b model at scale, is almost certainly cheaper to serve than a 32b dense model. *as you can shard the model across multiple GPUs at scale. but in reality you have some loss of efficiency from GPU coordination…

That's good information. I couldn't possibly even start to run even DeepSeek Flash on my system, but also if you're assuming multiple GPUs, that is going to affect the napkin math.

Re: Inference cost at scale with napkin math

#17

Post by 3eb7988a1663 »

JBAnderson5 wrote: post #7

> Realistically, serving 300 users per GPU you'll spend a lifetime cost of about $133 per user, plus the datacenter/upkeep bill. What is the operational cost and when does it become more expensive than the upfront capex? The B200 tops out at 1000W and idles around 140W. It averages around 600W. https://www.lightly.ai/blog/nvidia-b200-vs-h100 U.S. average electricity cost is $.14 per kWh in March. https://www.eia.gov/…

The EIA says $.14 for "Commercial" but $.086 for "Industrial". I assume that the big data centers have such high electricity costs that they would be able to cut better deals that would put them in line with the lower "Industrial" rates. So, even better potential margin.

Re: Inference cost at scale with napkin math

#18

Post by injuly »

smalltorch wrote: post #2

>This largely depends on whether you own or rent your hardware. At $40,000 per B200, your lifetime cost per user is 40_000/num_users. In the 100% duty cycle case (worst for cost), that's 6k$ per user. Realistically, serving 300 users per GPU you'll spend a lifetime cost of about $133 per user, plus the datacenter/upkeep bill . If you rent the GPU, the cost is more straightforward. At an hourly rate of $43, your hourl…

Yes. You can either rent an entire blade, or purchase a dis-assembled box (majority share of the price will be GPU), and place it at a datacenter.

This cannot be done on most premises because of power, noise, and cooling.

Re: Inference cost at scale with napkin math

#19

Post by martinald »

breput wrote: post #14
martinald wrote: post #13

Earlier quoted context omitted.

Yes 32B dense is a weird one to choose. But in reality, 32B dense is very similar* to 32B activated on MoE in terms of inference costs. And I highly suspect eg Opus is around that level of active params. A 284ba13b model at scale, is almost certainly cheaper to serve than a 32b dense model. *as you can shard the model across multiple GPUs at scale. but in reality you have some loss of efficiency from GPU coordination…

That's good information. I couldn't possibly even start to run even DeepSeek Flash on my system, but also if you're assuming multiple GPUs, that is going to affect the napkin math.

The point is that tok/s/GPU stays ~roughly stable. So you need say 4 GB200s minimum to fit the modules, but this provides 4x the tok/s as 1 GPU.

Re: Inference cost at scale with napkin math

#20

Post by smalltorch »

namibj wrote: post #5
smalltorch wrote: post #4

Earlier quoted context omitted.

So what's the cost separating them from placing this box at their premise? Network throughout?

Plus power and cooling.

The napkin math should figure in the base amount of GPU's i need to buy in order to be able to place it in a data center otherwise this math is not relevant. Because surely any place can cool, power, and protect a single gpu easily.

Not having physical access to my assets doesn't sound secure at all, and even a residential internet connection could handle this throughout.

Post reply on HN