▲ 316 points
back
89 comments
I actually looked at this in detail about a year ago for some automated driving compute work at my previous job, and I found that the detailed info you'd want from Nvidia was just 100% unavailable. There's pretty good proxies in some of the data you can get out of Nvidia tools, and there's some extra info you can glean from some of the function call stack in the open source Nvidia driver shim layer (because the actual main components are still binary blob, even with the "open source" driver), but over all you still can't get much useful info out.
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.
I just sent the link to a driver developer at Nvidia. If he shares the link with others at Nvidia, they should become aware of the idea tomorrow. That said, I have no idea if he will do that, but at least I tried.
Are they interested in you optimizing your workloads or just selling you more gpus to help you get to market faster...
I'm not sure, it seems to me like this should be doable in Nvidia as well. This is a paper that uses instruction sampling (called CUPTI) in Nvidia to provide optimization advice:
https://ieeexplore.ieee.org/document/9370339
It seems like the instruction sampler is there, and it also provides the stall reason.
The issue there is that that info is what Nvidia chooses to port out from the on-chip execution. Most of what we can do for observation is in the kernel driver space and not really on-chip or even low level transit to the chip. One of the other commenters pointed out that you can get huge benefits from avoiding busy waiting on the returned data from the chip, which makes total sense, but also increases latency, which didn't work for my near-realtime use case when I was investigating. Other than those types of low hanging fruit where you can accept a little latency for better power state management, it's hard to find low level optimizations specifically for Nvidia through the closed source parts of the CUDA stack or through the driver transit to chip when those are intentionally hidden.
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...
> Imagine halving the resource costs of AI and what that could mean for the planet and the industry -- based on extreme estimates such savings could reduce the total US power usage by over 10% by 20301.
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.
This is the https://en.wikipedia.org/wiki/Jevons_paradox and it's what always happens in these cases.
It does happen, but not always.
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.
It's possible to decrease costs faster than usage can rise.
Insulation, double glazed windows and other improvements in reducing heating and cooling costs in houses resulted in houses doubling in size.
Increasing fuel economy resulted in many more cars being replaced by SUVs.
AI usage will definitely increase to fill the space.
You specifically picked things like toasters and fridges which seem like frivolous if not entirely useless applications of LLMs.
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.
AI will be useful with toasters and fridges but of course that doesn't mean it will have to run on the devices itself
I actually think that fridges with image recognition would be a value add depending on the price. Could evaluate whether or not your food has spoiled, queue up a list of items to purchase, etc.
I had the same thought - power use will not be halved, usage will double instead.
The answer depends on what is rate-limiting growth; while we are supply-constrained on GPUs you can’t just increase AI usage.
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.
That depends on whether AI cost is dominated by power consumption cost [0]. I don’t think it is?
[0] For inference, that is. Training is another matter, and energy consumption for hardware manufacturing yet another.
> Imagine halving the resource costs of AI and what that could mean for the planet and the industry
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...
That would be notable... if anyone was actually using Gemini.
People who don't are missing out. I get perfect JSON formatted responses to my prompts for pennies.
rephrased as "We took compute from everything else.... and gave it to AI"
Being able to ‘connect’ call stacks between python, c++, and the gpu/accelerator seems useful.
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's a bit weird, very much a "software optimization" approach. But looking at the flame graph, you couldn't tell a model running in FP32 from one in INT8, taking 3x the time and energy.
> based on Intel EU stall profiling for hardware profiling
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/...
I've tried using flame graphs, but in my view nothing beats the simplicity and succinctness of gprof output for quickly analyzing program bottlenecks.
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.
This is so cool! Flame graphs are super helpful for analyzing bottlenecks. The eflambe library for elixir has let us catch some tricky issues.
https://github.com/Stratus3D/eflambe/blob/master/README.adoc
I never really liked flamegraphs much but I am going to put that aside for a bit and try to be as objective as possible.
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 these people are optimizing are things that flamegraphs do not show. That's things like latency in kernel launches, or synchronization overhead with the...
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.
> Imagine halving the resource costs of AI and what that could mean for the planet and the industry -- based on extreme estimates such savings could reduce the total US power usage by over 10% by 2030
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.
> Imagine halving the resource costs of AI ... based on extreme estimates such savings could reduce the total US power usage by over 10% by 2030
Is that implying that by 2030 they expect at least 20% of all US energy to be used by AI?
Data centers are big consumers of energy. Most modern data centers will have a mix of vector and scalar compute because ML/AI is a bunch of stuff, most of which was ubiquitous a decade ago.
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.
The data source is linked and is based on the ARM Datacenter Energy prediction.
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.
Seems pretty absurd
This is super interesting and useful. I tried reading the code to understand how GPU workloads worked last year and it was easy to get lost in all the options and pluggable layers.
I do not really understand the mentioned difficulties with instruction profiling.
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?
Trying to find out more about this EU stall thing Brendan talks about. Is it instruction sampling that gives you the reason for the stall? Sounds like a pretty advanced hw functionality.
I can imagine Nelson and other Anthropic engineers jumping for joy at this release.
Would love it if it was available and open source so people could use it in their own projects (or on their own hardware), instead of only being available on Intel's AI Cloud. But cool idea and execution nevertheless!
Yeah, would love to built-in support for this in PyTorch or TF
i am actually working on a platform that makes this sort of stuff easy. We use BPF under the hood and let you remotely deploy them across a cluster and visualize them.
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.
Wow. Nice.
Unrelated, but on the topic of reducing power consumption, I want to once again note that both AMD and NVidia max out a CPU core per blocking API call, preventing your CPU from entering low power states even when doing nothing but waiting on the GPU, for no reason other than to minimally rice benchmarks.
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 no reason other than to minimally rice benchmarks.
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.
totally looks like self promotion article lol
This guy invented flame graphs (among other things) so... I'm gonna allow it.
There has been a bit of hyperbole of late about energy saving AI.
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.