Now that Brendan works for Intel, he can get a lot of this info from the much more open source Intel GPU driver, but that's only so useful since everyone is either Nvidia or AMD still. The more hopeful sign is that a lot of the major customers of Nvidia are going to start demanding this sort of access and there's a real chance that AMD's more accessible driver starts documenting what to actually look at, which will create the market competition to fill this space. In the meantime, take a look at the flamegraph capabilities in PyTorch and similar frameworks, up an abstraction level and eek what performance you can.
https://ieeexplore.ieee.org/document/9370339
It seems like the instruction sampler is there, and it also provides the stall reason.
A while ago, I read a paper on dissecting the Nvidia architecture using very specifically tuned microbenchmarking to understand things like cache structure on chip and the like [0]. Unfortunately, no one has done this for seriously in use, recent architectures, so it's hard to use this info today. Similarly, there isn't an eBPF VM running on the chip to summarize all of this and the Nvidia tools aren't intended to make this kind of info easy to get, probably specifically because of this paper...
Why would it be the case that reducing the costs of AI reduces power consumption as opposed to increase AI usage (or another application using electricity)? I would think with cheaper AI their usage would be come more ubiquitous: LLMs in fridges, toasters, smart alarms, etc.
For example, food got cheaper and consumption has increased to the extent that obesity is a major problem, but this is much less than you might conclude from the degree to which productivity has increased per farmer.
For image generation, the energy needed to create an image is rapidly approaching the energy cost of a human noticing that they've seen an image — once it gets cheap enough (and good enough) to have it replace game rendering engines, we can't really spend meaningfully more on it.
(Probably. By that point they may be good enough to be trainers for other AI, or we might not need any better AI — impossible to know at this point).
For text generation, difficult to tell because e.g. source code and legal code have a lot of text.
When it comes to converting electricity into images and text, there really is no upper bound in sight. People are happy to load the internet up with as much content as they can churn out.
Now maybe waste is a bigger issue with content than with food. I'm not sure. Both have some nonzero cost to waste. It might depend on how content is distributed
I'd would say that text is capable of being extremely useful even when no human reads it, because of source code, maths proofs, etc.
But I'm curious: 238 wpm * 0.75 words per token * 16 (waking) hours per day * 83 years * $10.00 / 1M output tokens (current API cost for 4o without batching) means the current cost of making as many tokens as a human can read in a lifetime is $92,300: https://www.wolframalpha.com/input?i=238+words+per+minute+%2...
With these numbers, a well-written project with even a billion lines of code would be a rounding error even if only a thousand people used any specific such software and none of that was ever shared with what other people wanted to get done.
Market dynamics should push people to stop generating that content if they don't enough value to justify the cost. In practice, though, it hasn't seemed to happen yet and we must be pass a threshold where there's more content created online than we could ever value.
It'd make for an interesting study, but short of having verifiable data I have to assume we'll continue increasing the rate at which content is created whether the value is there or not.
You'd most likely categorize all of the unseen textures or higher-than-needed resolution in your "waste" bucket, and I can't argue with that. But VR still clearly means that there is at least theoretical room for "realtime video generated custom for every viewer, which in turn is composed of even more content sources".
If everyone's doing that all day every day on each eye, that's a reasonable guess of an upper bound: you as a human cannot actually consume more even if you make it.
GANs can already do that speed, but any given GAN is a specialist AI and not a general model; diffusion models are general, but they're slower (best generation speed I've seen is 4-5 frames per second on unknown hardware). LLMs aren't really suited to doing images at all, but can control other models (this is what ChatGPT does when "it makes an image" — it calls out to DALL•E).
* how long I've been paying attention to that, not a detailed historical analysis
That said, if we got to such a massive scale I'd expect us to hit other limits first (electricity available, best produced, storage space, network transmission, etc.).
Or did I totally misunderstand your example here? I may have misread it completely, if so sorry about that!
Sure, absolutely. But I can say the same of food, which is why I drew the analogy between them previously.
> That said, if we got to such a massive scale I'd expect us to hit other limits first (electricity available, best produced, storage space, network transmission, etc.).
Difficult to guess when the quality isn't yet at the right threshold: GANs are already this speed on phone hardware*, so we're not bounded on that specific combination with available electrical energy; on the other hand, 2 years ago I was seeing images for about 3 kJ, which is in the region of hundreds of kilowatts for 2 eyes at 60 fps, which is absolutely going to be a problem… if they were limited to that hardware and with that model (though both are moving targets, I'd be very surprised if the unknown hardware that I've seen doing 4-5 fps was burning 12-15 kW, but it's not strictly speaking impossible it really was that power hungry).
* Specifically: on an iPhone 11, BlazeStyleGAN model was generating images in 12.14 ms, which is just over 82 fps — https://research.google/blog/mediapipe-facestylizer-on-devic...
I'm no Malthusian, but the paradox holds here pretty well.
The very specific point I'm claiming is that the increased consumption isn't always unbounded.
* More women work more and invest in their own education and fewer spend time alone at home as they might in poorer countries which would facilitate giving birth and investing time on childcare that way.
* More men and women derive their primary income from work that children cannot easily participate in. EG: office work, work from home computer work, vs farming or working with one's hands. In many poorer countries it is common practice to have more children at least partially to bolster the labor force around the house.
* Wealthier nations have better access to family planning: contraception, abortion, pasttimes that can meaningfully compete against getting laid in the first place.
Sources: Colleran, H., Snopkowski, K. Variation in wealth and educational drivers of fertility decline across 45 countries. Popul Ecol 60, 155–169 (2018). https://doi.org/10.1007/s10144-018-0626-5 https://link.springer.com/article/10.1007/s10144-018-0626-5
More Work, Fewer Babies: What Does Workism Have to Do with Falling Fertility? - Laurie DeRose and Lyman Stone https://ifstudies.org/ifs-admin/resources/reports/ifs-workis...
I'd assume environmental, but there's also more subtle answers than will fit in a comment box — whatever the cause, it has to be near-global.
China's building loads more houses, still has a fertility decline.
It is really a good example of what natural dimension reducers we are, even when we know it makes no sense. It is like we can't but help ourselves to reduce things to one explanatory variable.
My favorite is the news headline "The market went up today because of X".
They say: Tesla shares up as revealations surface that the wind is blowing east.
https://arxiv.org/abs/2408.14837
Also TIL this is generated at 20 frames per second, the best I've used myself was "only" 4-5; does anyone know the performance and power consumption of a Google TPU?
The only hope is to generate this power greenly.
Increasing fuel economy resulted in many more cars being replaced by SUVs.
AI usage will definitely increase to fill the space.
But you can be more charitable and imagine more productive uses of AI on the edge that are impossible today. Those uses would presumably create some value, so if by reducing AI energy costs by 90% we get all the AI usage we have today plus those new uses that aren't currently viable, it's a better bang for buck.
The next bottleneck will be datacenter power interconnects, but in that scenario as you say you can expect power usage to expand to fill the supply gap if there is a perf win.
[0] For inference, that is. Training is another matter, and energy consumption for hardware manufacturing yet another.
Google has done this: "In eighteen months, we reduced costs by more than 90% for these queries through hardware, engineering, and technical breakthroughs, while doubling the size of our custom Gemini model." https://blog.google/inside-google/message-ceo/alphabet-earni...
Tried the Gemini Advanced trial last week. For some reason their so called 1M context model is limited to 10 files at a time, so you can't upload a codebase for it to reference and even with the extra data the end result is somehow worse than both Sonnet or 4o without much given context at all. It's definitely not on the level as a coding assistant at least.
I wonder if this pushes a bit much towards flamegraphs specifically. They were an innovation when they were first invented and the alternatives were things like perf report, but now I think they’re more one tool among many. In particular, I think many people who are serious about performance often reach for things like pprof for statistical profiles and various traceing and trace-visualisation tools for more fine-grained information (things like bpftrace, systemtap, or custom instrumentation on the recording side and perfetto or the many game-development oriented tools on the visualisation (and sometimes instrumentation) side).
I was particularly surprised by the statement about intel’s engineers not knowing what to do with the flamegraphs. I read it as them already having tools that are better suited to their particular needs, because I think the alternative has to be that they are incompetent or, at best, not thinking about performance at all.
Lots of performance measuring on Linux is done through the perf subsystem and Intel have made a lot of contributions to make it good. Similarly, Intel have added hardware features that are useful for measuring and improving performance – an area where their chips have features that, at least on chips I’ve used, easily beat AMD’s offerings. This kind of plumbing is important and useful, and I guess the flamegraphs demonstrate that the plumbing was done.
It wasn't clearly defined but I think EU stall means Execution Unit stall which is when a GPU "becomes stalled when all of its threads are waiting for results from fixed function units" https://www.intel.com/content/www/us/en/docs/gpa/user-guide/...
https://ftp.gnu.org/old-gnu/Manuals/gprof-2.9.1/html_chapter...
For each function you know how much CPU is spent in the function itself, as opposed to child calls. All in a simple text file without the need for constantly scrolling, panning, and enlarging to get the information you need.
https://github.com/Stratus3D/eflambe/blob/master/README.adoc
I don't find the usecase presented here compelling. Cutting out the "yo we will save you $x billion in compute" costs the tools presented here seem to be…stacktraces for your kernels. Stacktraces that go from your Python code through the driver shim to the kernel and finally onto the GPU. Neat. I don't actually know very much about what Intel has in this area so perhaps this is a step forward for them? If so, I will always applaud people figuring out how to piece together symbols and whatnot to make profiling work.
However, I am still not very impressed. Sure, there are some workloads where it is nice to know that 70% of your time is spent in some GEMM. But I think the real optimization doesn't look like that all. For most "real" workloads, you already know the basics of how your kernels look and execute. Nobody is burning a million dollars an hour on a training run without knowing what each and every one of the important kernels are. Some of them were probably written by hand. Some might be written in higher-level PyTorch/Triton/JAX/whatever. Still others might be built on some general library. But the people who do this are not stupid, and they aren't going to be caught unawares that a random kernel has suddenly popped up on their flamegraph. They should already know what is there. And most of these tools have debugging facilities to dump intermediate state in forms that tools understand. Often this is incomplete and buggy, I know. But it's there and people do use them.
What these people are optimizing are things that flamegraphs do not show. That's things like latency in kernel launches, or synchronization overhead with the host. It's global memory traffic and warp stalls. Sure, the tools to profile this are immature compared to what the hyperscalers have for CPUs. But they are still present and used heavily: I don't buy the argument that knowing that your python code calls a kernel through __cuda12_ioctl_whatever is actually helpful. This seems like a solution searching for a problem, or maybe a basic diagnostic tool at best.
What OP is showing is an example of what can be shown on flamegraphs. They are a generic visualisation tool so if you want to include latency or whatever (financial cost maybe?) you are free to do it.
As for the rest, Intel is here providing tools for developers who would like to optimize the sw stacks on their platform. Invaluable if you would like to efficiency support non-NVidia hardware.
The way this is phrased threw me off. It sounded to me like the author was comparing the power use of a more efficient LLM industry to US usage without LLMs and expecting it to be 10% lower.
Looking into the source linked with the claim, it doesn't even hold up when compared against how much power LLMs use today. The linked article raises an estimate that LLM power use could increase 15-23 times between 2023 and 2027, and that by 2030 LLMs could account for 20-25% of our total energy use.
Working that match backwards, the benefit the author is hailing as a success is that we would only increase energy use by say 7.5-11.5 times by 2027 and that in 2030 LLMs would only be 10% of the total energy use. That's not a win in my book, and doesn't account for the Jevan's Paradox problem where we would almost certainly just use all that efficiency gain to further grow LLM use compared to the 2030 prediction without the efficiency gains.
Is that implying that by 2030 they expect at least 20% of all US energy to be used by AI?
In the limit case where Prineville just gets 100k BH100 slammed into it? The absolute best you’re going to do is to have Brendan Gregg looking at the cost. He’s the acknowledged world expert on profiling and performance tuning on modern gear in the general case. There are experts in a vertical (SG14, you want to watch Carl Cook).
I’ve been around the block and my go-to on performance trouble is “What’s the Gregg book say here…” it your first stop.
But i don't think its too far fetched.
The compute needed for digital twins, simulating a whole army of robots than uploading it to the robots, who sitll need a ton of compute, is not unrealistic.
Cars like Tesla have A TON of compute build in too.
And we have seen what suddenly happens to an LLM when you switch the amount of parameters. We were in a investment hell were it was not clear in what to invest (crypto, blockchain and NFT bubble bursted) but AI opened up the sky again.
If we continue like this, it will not be far fetched that everyone has their own private agent running and paying for it (private / isolated for data security) + your work agent.
Are they saying it is hard to sample the stacks across the boundary? Are they saying it is hard to do so coherently because the accelerator engine is actually asynchronous so you need to do some sort of cross-boundary correlation?
However, they then talk about file systems and /proc representations which have nothing to do with the actual sampling process; only posing problems for the display of human-readable information. Many naive profiling, tracing, and logging implementations conflate these actions to their detriment; are they being conflated here or is it just a generic statement of the scope of problems?
Check us out: https://yeet.cx
Our current package index is a bit thin:
We have a ton in the pipeline and are going to add more in the coming weeks and release an SDK.
Basically, these APIs are set up to busyspin while waiting for a bus write from the GPU by default (!), rather than use interrupts like every other hardware device on your system.
You turn it off with
NVidia: `cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync)`
AMD: `hipSetDeviceFlags(hipDeviceScheduleBlockingSync)`
On Pytorch
NVidia: `import ctypes \ ctypes.CDLL('libcudart.so').cudaSetDeviceFlags(4)`
AMD: `import ctypes \ ctypes.CDLL('libamdhip64.so').hipSetDeviceFlags(4)`
This saves me 20W whenever my GPU is busy in ComfyUI.
Every single device using the default settings for CUDA/ROCM burns a CPU core per worker thread for no reason.
For AI/ML applications, perhaps no one will notice.
For gaming, yielding threads of execution to the OS can periodically incur minimum scheduler delays of 10-20ms. Many gamers will notice an ~extra frame of latency being randomly injected.
There is simply no excuse for an app that does 10 API calls a second to burn 100% CPU.
There isn't a magic bullet here, it's just people improving a relatively new technology. Even though the underlying neural nets are fairly old now, the newness of transformers and the newness of the massive scale means there's quite a lot of low hanging fruit still. Some of the best minds are on this problem and are reaching for the hardest to get fruit.
A lot of these advancements work well together improving efficiency a few percent here, a few percent there.
This is a good thing, but people are doing crazy comparisons by extrapolating older tech into future use cases.
This is like estimating the impact of cars by correctly guessing that there are 1.4 Billion cars in the world and multiplying that by the impact of a single model-T Ford.