[atleastoptimal]

Actual inference costs without considering subsidies and loss leaders are going down, due to algorithmic improvements, hardware improvements, and quantized/smaller models getting the same performance as larger ones. Companies are making huge breakthroughs making chips specifically for LLM inference

[menaerus]

In August 2023, llama2 34B was released and at that time, without employing model quantization, in order to fit this model one needed to have a GPU, or set of GPUs, with total of ~34x2.5=85G of VRAM.

That said, can you be more specific what are those "algorithmic" and "hardware" improvements that has driven this cost and hardware requirements down? AFAIK I still need the same hardware to run this very same model.

[stocknoob]

Take a look at the latest Llama and Phi models. They get comparable MMLU performance for ~10% of the parameters. Not to mention the cost/flop and cost/gb for GPUs has dropped.

You aren’t trying to run an old 2023 model as is, you’re trying to match its capabilities. The old models just show what capabilities are possible.

[menaerus]

Sure, let's say that 8B llama3.1 gets comparable performance of it's 70B llama2 predecessor. Not quite true but let's say that hypothetically it is. That still leave us with 70B llama3.1.

How much VRAM and inference compute is required to run 3.1-70B vs 2-70B?

The argument is that the inference cost is dropping down significantly each year but how exactly if those two models require about the ~same, give or take, amount of VRAM and compute?

One way to drive the cost down is to innovate in inference algorithms such that the HW requirements are loosened up.

In the context of inference optimizations one such is flash-decode, similar to its training counter-part flash-attention, from the same authors. However, that particular optimization concerns only by improving the inference runtime by dropping down the number of memory accesses needed to compute the self-attention. Amount of total VRAM you need in order to just load the model still remains the same so although it is true that you might get a tad more from the same HW, the initial requirement of total HW you need remains to be the same. Flash-decode is also nowhere near the impact of flash-attention. Latter enabled much faster training iteration runtimes while the former has had quite limited impact, mostly because scale of inference is so much smaller than the training so the improvements do not always see the large gains.

> Not to mention the cost/flop and cost/gb for GPUs has dropped.

For training. Not for inference. GPU prices remained about the same, give or take.

[stocknoob]

> How much VRAM and inference compute is required to run 3.1-70B vs 2-70B?

We aren’t trying to mindlessly consume the same VRAM as last year and hope costs magically drop. We are noticing that we can get last year’s mid-level performance on this year’s low-end model, leading to cost savings at that perf level. The same thing happens next year, leading to a drop in cost at any given perf level over time.

> For training. Not for inference. GPU prices remained about the same, give or take.

See:

https://epoch.ai/blog/trends-in-gpu-price-performance

We don’t care about the absolute price, is the cost per flop or cost per GB decreasing over time with each new GPU?

—-

If it isn’t clear why inference costs at any given performance level will drop given the points above, unfortunately I can’t help you further.

[menaerus]

We absolutely care about absolute costs. 70B model this year will cost as much as it will next year, unless Nvidia decides to lose their profits. The question is whether an inference cost is dropping down. And the answer is obviously no. I see that you're out of your depth so let's just stop here.