For kernel-level performance tuning you can use the occupancy calculator as pointed out by jplusqualt or you can profile your kernel with Nsight compute which will give you a ton of info.

But for model-wide performance, you basically have to come up with your own calculation to estimate the FLOPs required by your model and based on that figure out how well your model is maxing out the GPU capabilities (MFU/HFU).

Here is a more in-depth example on how you might do this: https://github.com/stas00/ml-engineering/tree/master/trainin...

It's harder than measuring CPU load, and depends a lot on context. For example, often 90% of a GPU's available flops are exclusively for low-precision matrix multiply-add operations. If you're doing full precision multiply-add operations at full speed, do you count that as 10% or 100% load? If you're doing lots of small operations and your warps are only 50% full, do you count that as 50% or 100% load? Unfortunately, there isn't really a shortcut to understanding how a GPU works and knowing how you're using it.

CUDA toolkit comes with an occupancy calculator that can help you determine based on your kernel launch parameters how busy your GPU will potentially be.

For more information: https://docs.nvidia.com/cuda/cuda-c-programming-guide/#multi...

you need to profile them, nsight is one even torch does flamegraphs

FYI, adding @ before a user name does nothing besides looking terrible and AFAIK dang does not get a notification when he’s mentioned. If you want to contact him, the best way is to send an email to hn@ycombinator.com .

I think I was shadow-banned because my very first comment on the site was slightly snarky, and have now been unbanned.

"Utilization" tells you the percentage of your GPU's SM that currently have at least one thread assigned to them.

It does not at all take into count how much that thread is actually using the core to it's capacity.

So if e.g. your thread is locked waiting on some data from another GPU (NCCL) and actually doing nothing, it will still show 100% utilisation. A good way to realize that is when a NCCL call timeout after 30 minutes for some reason, but you can see all your GPUs (except the one that cause the failure) were at 100% util, even though they clearly did nothing but wait.

Another example are operation with low compute intensity: Say you want to add 1 to every element in a very large tensor, you effectively have to transfer every element (let's say FP8, so 1 byte) from the HBM to the l2 memory, which is very slow operation, to then simply do an add, which is extremely fast. It takes about ~1000x more time to move that byte to L2 than it takes to actually do the add, so in effect your "true" utilization is ~0.2%, but nvidia-smi (and this tool) will show 100% for the entire duration of that add.

Sadly there isn't a great general way to monitor "true" utilization during training, generally you have to come up with an estimate of how many flops your model requires per pass, look at the time it takes to do said pass, and compare the flops/sec you get to Nvidia's spec sheet. If you get around 60% of theoretical flops for a typical transformer LLM training you are basically at max utilization.

Utilisation is counted by the OS, it's not exposed as a performance counter by the hardware. Thus, it's limited by the level of abstraction presented by the hardware.

It's useless on CPUs as well, just to a much much lesser extent to the point of it actually being useful.

Basically, the OS sees the CPU as being composed of multiple cores, that's the level of abstraction. Thus, the OS calculates "portion of last second where atleast one instruction was sent to this core" on each core and then reports it. The single number version is an average of each core's value.

On the other hand, the OS cannot calculate stuff inside each core - the CPU hides as part of its abstraction. That is, you cannot know "I$ utilisation", "FPU utilisation", etc,.

In the GPU, the OS doesn't even see each SM (streaming multiprocessor, loosely analogous to a cpu core). It just sees the whole GPU as one black box abstraction. Thus, it calculates utilisation as "portion of last second where atleast one kernel was executing on the whole GPU". It cannot calculate intra-GPU util at all. So one kernel executing on one SM looks the same to the OS, as that kernel executing on tens of SMs!

This is the crux of the issue.

With performance counters (perf for CPU, or nsight compute for GPU), lots of stuff visible only inside the hardware abstraction can be calculated (SM util, warp occupancy, tensor util, etc)

The question then, is why doesn't the GPU schedule stuff on each SM in the OS/driver? Instead of doing it in a microcontroller in the hardware itself on the other side of the interface?

Well, I think it's due to efficiency reasons and also for nvidia to have more freedom to change it without having compat issues due to being tied to the OS, and similar reasons. If that were the case however, then the OS could calculate util for each SM, and then average it, giving you more accurate values - the case with the kernel running on 1 SM will report a smaller util than the case with the kernel executing on 15 SMs.

IME, measuring on nsight compute causes anywhere from a 5% to 30% performance overhead, so if that's ok for you, you can enable it and get more useful measurements.