nothing super special to be honest. It's just a quick way for me to take a look at a couple of GPU boxes from the browser. Sometimes I check it from the ipad too
The AI/vibe coded "purple" color scheme is a meme at this point - might want to tweak the look and feel to not be so on the nose, but it's otherwise a good dashboard.
Hi everyone, I just built a GPU dashboard to check the utilization on NVIDIA cards directly in your browser. It also works with multiple GPUs. The idea is to have real-time metrics from a remote GPU server instead of running nvidia-smi. Let me know if you try it out!
This looks neat and would probably be cool to run on one of my passive info screens. But until it supports more than just Nvidia I'll have to stick with nvtop. Might be a good idea to pull the theme out to a file so it's all swappable too (assuming you haven't, I can't look at the code right now.)
I’m skeptical of “no ssh” being a benefit. I’d rather have one port opened to the battle tested ssh proc (which I probably have already anyway), than open a port to some random application.
I suppose it’s trivial to proxy a http port over ssh though so that would seem like a good solution
This is awesome! Tested it out while running some plex encoding and everything worked as expected!
I did notice that nvidia-smi shows the process name as plex-transcoding but gpu-hot is showing [Not Found]. Not sure if that is where the process name is supposed to go
Properly measuring "GPU load" is something I've been wondering about, as an architect who's had to deploy ML/DL models but is still relatively new at it. With CPU workloads you can generally tell from %CPU, %Mem and IOs how much load your system is under. But with GPU I'm not sure how you can tell, other than by just measuring your model execution times. I find it makes it hard to get an idea whether upgrading to a stronger GPU would help and by how much. Are there established ways of doing this?
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).
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.
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 .
"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.
Definitely a better high level metric than nvidia-smi, and probably fine if you just want to get a very coarse idea of whether or not your are using the GPUs reasonably at all.
But when you get to the point where you care about a few percentage points of utilisation it's just not reliable enough as many things can impact energy consumption both ways. E.g. had a case were the GPU cluster we were using wasn't being cooled well enough, so you would gradually see power draw getting lower and lower as the GPUs were throttling themselves to not overheat.
You can also find cases were energy consumption is high but MFU/HFU isn't, like memory intensive workloads
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.
half readable color scheme.. random python and javascript mixed in, ships with 2 python CVEs out of the box out of 5 total dependencies... yep it checks out bois...certified infested slop
python-socketio==5.8.0: 1 CVE (CVE-2025-61765); Remote Code Execution via malicious pickle deserialization in multi-server setups.
eventlet==0.33.3: 1 CVE (CVE-2025-58068); HTTP request smuggling from improper trailer handling.
And then economists wonder why are none of these people getting jobs...
I mean the python-socketio is from a few days ago and likely doesn't affect this package (it's not using message queues, right?)
Eventlet .33 is ancient, no idea why they would use that.
With this said, most people should have some kind of SCA to ensure they're not using ancient packages. Conversely picking up a package the day it's released has bit a lot of people when the repository in question gets pwned.
TBH this seems useful only for a very select niche.
If you're a company and you have several GPU machines in a cluster, then this is kinda useless b/c you'd have to go on each container or node to view the dashboard.
Sure, there's a cost to using opentelemetry + whatever storage+viz backend, but once it's set up you can actually do alerting, historical views, analysis, etc. easily.
I think AI vibe codes that because it’s probably seen that default so much.
I suppose it’s trivial to proxy a http port over ssh though so that would seem like a good solution
I think it’s super cool. Clean design. Great for your local self-hosted system or one of your local company systems in the office.
If you have a fleet of GPUs then maybe use your SSH CLI. This is fun and cool looking though.
I did notice that nvidia-smi shows the process name as plex-transcoding but gpu-hot is showing [Not Found]. Not sure if that is where the process name is supposed to go
Does not change the usefulness of this dashboard, just wanted to point it out.
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...
For more information: https://docs.nvidia.com/cuda/cuda-c-programming-guide/#multi...
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.
But when you get to the point where you care about a few percentage points of utilisation it's just not reliable enough as many things can impact energy consumption both ways. E.g. had a case were the GPU cluster we were using wasn't being cooled well enough, so you would gradually see power draw getting lower and lower as the GPUs were throttling themselves to not overheat.
You can also find cases were energy consumption is high but MFU/HFU isn't, like memory intensive workloads
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.
Eventlet .33 is ancient, no idea why they would use that.
With this said, most people should have some kind of SCA to ensure they're not using ancient packages. Conversely picking up a package the day it's released has bit a lot of people when the repository in question gets pwned.
If you're a company and you have several GPU machines in a cluster, then this is kinda useless b/c you'd have to go on each container or node to view the dashboard.
Sure, there's a cost to using opentelemetry + whatever storage+viz backend, but once it's set up you can actually do alerting, historical views, analysis, etc. easily.
// solves everything at the above container claims to do lol
check also netdata amazing project