-
Posts
3,672 -
Joined
-
Last visited
Reputation Activity
-
igormp reacted to RollinLower in Should I buy Macbook pro M2 (24 GB Memory) or Macbook pro M3 (8 GB Memory) ?
get the 24GB model. The small bump in compute performance will mean nothing if your RAM is full 90% of the time.
-
igormp got a reaction from leadeater in SLURM and Servers Configuration
How many scientists do you have?
SLURM is nice, but also pretty annoying. As an example, if your devs are using VSCode jupyter notebooks to do stuff on the GPU server, it wouldn't be possible with SLURM anymore since there are no interactive sessions (AFAIK).
If you're ok with having your devs submit a bash job and do some workarounds to start a jupyter server and then connect VSCode to it, then it could work, and you wouldn't have the issue of multiple folks trying to use the same GPU anymore. You could also look into MIG instances for that A100 to properly share it.
-
igormp got a reaction from BoomerDutch in SLURM and Servers Configuration
How many scientists do you have?
SLURM is nice, but also pretty annoying. As an example, if your devs are using VSCode jupyter notebooks to do stuff on the GPU server, it wouldn't be possible with SLURM anymore since there are no interactive sessions (AFAIK).
If you're ok with having your devs submit a bash job and do some workarounds to start a jupyter server and then connect VSCode to it, then it could work, and you wouldn't have the issue of multiple folks trying to use the same GPU anymore. You could also look into MIG instances for that A100 to properly share it.
-
igormp got a reaction from wanderingfool2 in Is C++ necessary for some libraries over C?
AFAIK the WinAPI itself is C-only (with wrappers to other langs). Aren't you talking about other APIs that are also on windows?
This implies that C is kind of a subset of C++, and that's not the case. Most of the stuff under C is able to compile in a Cpp compiler, but not everything.
No, but I bet there's a C FFI/wrapper for those APIs.
-
igormp got a reaction from filpo in Revisiting an eight month old machine learning dual 4090 build
No, lanes are a physical thing. If the CPU has a max of 16 lanes available for those slots, and you populate two slots, each will be x8/x8, no matter the version of the device you plug into those (this will limit the version). If you plug one 3.0 device and another 4.0 device, the former will run at 3.0 x8 while the latter will be at 4.0 x8.
The 4090 is a 4.0 device, so you'll be capped bw wise. With that said, it shouldn't be that relevant for most cases.
What excatly are you planning to do? As in, what models and model sizes are you working with? Without this info I can't say if that'll be relevant or not.
Around this price range? No
For this kind of thing on PCIe 5 you're limited to Xeon WS and the newest Threadrippers.
AFAIK TR 7000 only accepts RDIMM sticks, so this build would not work with regular UDIMMs.
Always a good idea to avoid sag with those heavy af GPUs. Also get support for the first one even if the slot is reinforced.
If you're talking about the TR build, all of the PCIe slots go through the CPU, there are no chipsets for it (the chipset is only for stuff like USB and ethernet iirc).
You could take a look into some Xeon WS offerings. The 2400 line has 64 PCIe 5.0 lanes and some "cheap" entry offerings (like 6~16 cores).
Are you planning on fine tuning >24gb LLM models across those different GPUs by distributing the model itself? If so, data transfers become a bottleneck indeed, but without NVLink there's nothing you can do, and that third slot doesn't go through the chipset as I said above. Also, mixing a 3090 and 4090 in this case (LLM fine tuning/training) wouldn't make much of a difference since you'd be bottlenecked by memory speeds, and both GPUs have pretty similar memory speeds anyway.
As for power, just power limit the GPUs, doing 300W each should give you a negligible perf difference. I run 2x3090s at 275W each, only lost ~10% perf tops while saving 100W of power for each GPU (my electricity bill prefers it this way).
Also note that you won't be fitting a 3rd GPU in that case without a riser and some jankiness, the last slot is pretty crammed.
Actually, no regular ATX case/mobo will allow you to easily do 3 GPUs without some jankiness.
-
igormp got a reaction from CatHerder in Large language models (LLM) Local Running GPU, CPU: Benchmarks and Next-Gen Prospects
That doesn't make things faster, just make them require more less memory. (typo, sorry)
In fact, it's likely it actually makes things slower since you'll have some overhead on converting the data to a suitable unit of data for processing (int8/fp8 or whatever your hw has support for)
-
igormp got a reaction from Sushant Shah in 4060 Ti 16GB vs Intel Arc A770 16GB for AI/ML workloads (Keras)? Price to performance...
If you know what you're doing, and will keep yourself solely to using tensorflow and pytorch, then you can go by with an A770, and it actually manages to provide nice results. Take a look at this blog series:
https://christianjmills.com/series/notes/arc-a770-testing.html
However, I'd still recommend going with nvidia for the sake of out of the box compatibility. A used 3090 is a great option too, if you can afford it in your region.
-
igormp got a reaction from Zando_ in 4060 Ti 16GB vs Intel Arc A770 16GB for AI/ML workloads (Keras)? Price to performance...
If you know what you're doing, and will keep yourself solely to using tensorflow and pytorch, then you can go by with an A770, and it actually manages to provide nice results. Take a look at this blog series:
https://christianjmills.com/series/notes/arc-a770-testing.html
However, I'd still recommend going with nvidia for the sake of out of the box compatibility. A used 3090 is a great option too, if you can afford it in your region.
-
igormp got a reaction from Somerandomtechyboi in 4060 Ti 16GB vs Intel Arc A770 16GB for AI/ML workloads (Keras)? Price to performance...
If you know what you're doing, and will keep yourself solely to using tensorflow and pytorch, then you can go by with an A770, and it actually manages to provide nice results. Take a look at this blog series:
https://christianjmills.com/series/notes/arc-a770-testing.html
However, I'd still recommend going with nvidia for the sake of out of the box compatibility. A used 3090 is a great option too, if you can afford it in your region.
-
igormp got a reaction from 907rider in Show off Your Setup! (Rev.2)
I see, nice workflow.
Most of my work is with large-ish CNNs/ViTs with 40~100GB of data (so I try to keep my GPUs well feed by properly batching my pipeline), and LLMs (from 7 to ~70B params).
-
igormp got a reaction from 907rider in Show off Your Setup! (Rev.2)
Sweet rig.
Couple questions that sparked my curiosity:
- Are you a full time windows user, or is it a dual boot machine? Asking because from my personal experience the performance hit on windows for ML stuff is pretty significant.
- Are those 128gb of RAM enough for you? Often I feel like my 128gb can't keep my 2x3090 well fed when I have a somewhat larger pipeline, forcing me to do all data processing before jumping into the actual training.
-
igormp reacted to 907rider in Show off Your Setup! (Rev.2)
Thank you!
1: It is a dual boot machine, in the photos where you can see the monitors, the laptop that is on the desk is actually what is running the monitors. There's a Razor Core X Chroma on the upper shelf with an RTX 3080 in it. But I almost exclusively use the tensorflow python library which will not run on GPU's if it is being run from Windows. Workflows typically look like: code up something on the laptop, get it running and working well, then the laptop and workstation are connected via a 2.5Gig switch, use "FreeFileSync" to sync the files/folders between the two, remote desktop into Ubuntu on the workstation, fire up VScode, and start the models training.
2: So far I have not had any issues with the 128gb of RAM. I was definitely running into some issues with the "old" workstation which "only" had 64gb of ram. I would be really curious to know what your workflows are like and see where you are running into issues! I am coming at this machine learning stuff from an Electrical Engineering perspective (not a computer science guy) so this is a whole new world to me and the learning curve has been steep 😬
-
igormp reacted to OddOod in question for the professional developer's
Honestly, 90% of what I use in my day to day job I learned *on* the job. Book learnin' and lecture can only take you so far, everything else is just time behind the keyboard.
If you're really gonna make a go at it, I'd commit and go for an internship after nailing the basics. I learned more SQL in a week of doing support tickets than I did in two semesters of database classes. I haven't taken a C# class in my life but that's >80% of my professional output.
So, can you get work in a language you don't know? Yes, but mostly at the entry level. I did get dang far in an interview process for a senior level Go position after spending 3 hours learning stuff, but that's a rarity.
Oh, and for internships, you really wanna hone your soft skills. Get good at talking to people. Active listening and empathy are learnable skills that will pay dividends for you every day for the rest of your life. In fact, I would say that the most valuable class I took in my 7 years of college was Improv for Engineers.
-
igormp reacted to OddOod in question for the professional developer's
I've got about a decade of professional coding experience.
I find that Java and C# are pretty interchangeable. They are both full fat languages with massive sets of libraries behind them. Takes a bit of effort to switch between them but it's not a ton of effort. I primarily professionally code in C#. If you want to work in the web dev sphere, either is a great place to focus effort
As for Python, dang, it's POWERFUL, but ultimately a lot of that power comes from precompiled binaries often written in some flavor of C. But it's most useful in building pipelines. If you want to work in data sciences (computational biology, any physics/astronomy, most geology/oil/water, etc.) it's a really good tool, but know that there is *very* little money in those fields without a PhD. That being said, if I'm writing something personal and durable that can't be easily done in bash/cmd/ps, I always reach for python.
Ultimately, the best advice is to stay language agnostic. Learn the fundamental concepts of programming: boolean logic, objects/interfaces/inheritance, lambda functions, etc. Then apply those ideas with whatever syntax the language requires. Heck, there are some super cool things being done with Rust and LUA and Go. They all have their strengths and they all have their weaknesses, but never let anyone tell you that one language is better than all others, even PERL has its place... probably.
-
igormp reacted to Eigenvektor in Getting 9 year old to play games other than Roblox?
I get it. You want to play with your kid, but you don't want to play the games she enjoys. I guess the adult thing to do would be to suck it up and play the game anyway.
"I want her to enjoy what I enjoy" or "I want here to play what I enjoy" feels kind of selfish. Maybe you can find games that are enjoyable for you both, but I'm not sure how we're supposed to recommend something if we don't really know either of you.
-
igormp got a reaction from Eigenvektor in Why can't strings have some pre-allocated bytes in C for storing the length?
You can create a struct and do that.
Many other languages do have such abstractions over strings, such a Cpp.
-
igormp got a reaction from shadow_ray in Why can't strings have some pre-allocated bytes in C for storing the length?
You can create a struct and do that.
Many other languages do have such abstractions over strings, such a Cpp.
-
igormp got a reaction from PDifolco in Why can't strings have some pre-allocated bytes in C for storing the length?
You can create a struct and do that.
Many other languages do have such abstractions over strings, such a Cpp.
-
igormp got a reaction from YoungBlade in NVIDIA GeForce RTX 2080 Ti gets a 44GB memory mod
Yes, also the VRAM is slower and GDDR6, there are no professional GPUs using GDDR6X on the green side. Those chips also have ECC.
Apart from the x100 and its derivatives, yes.
-
igormp got a reaction from WhitetailAni in NVIDIA GeForce RTX 2080 Ti gets a 44GB memory mod
Fun fact, the last quadro that had FP64 was the V100 Quadro, which was based on the... V100 chip. The last non x100 chips to have FP64 were the kepler ones, every non-100 chip after that lacks full FP64 support, be it quadro, tesla or geforce.
-
igormp got a reaction from Eigenvektor in What is the point of these useless typedefs?
Those are more often than not abstractions related to other things, like different CPU architectures (a int and long change on size depending on which architecture you are on), or higher level stuff (DWORD and QWORD are common naming through the computing world, and iirc those names are also used in Window's registry editor).
Undersocres are often used to indicate private functions/variables. Since C doesn't have the proper concept of private functions, this is a workaround that's used, which helps to avoid naming collisions.
If you're planning on supporting windows in the long term and across different versions, using their abstractions is likely the best way, otherwise you'll be doing your own IFDEFs for each version in case any of the underlying types change.
-
igormp got a reaction from Sauron in What is the point of these useless typedefs?
Those are more often than not abstractions related to other things, like different CPU architectures (a int and long change on size depending on which architecture you are on), or higher level stuff (DWORD and QWORD are common naming through the computing world, and iirc those names are also used in Window's registry editor).
Undersocres are often used to indicate private functions/variables. Since C doesn't have the proper concept of private functions, this is a workaround that's used, which helps to avoid naming collisions.
If you're planning on supporting windows in the long term and across different versions, using their abstractions is likely the best way, otherwise you'll be doing your own IFDEFs for each version in case any of the underlying types change.
-
igormp got a reaction from NinJake in Linux AV
There's no real reason to use a real time AV on linux, not many viruses out there target desktop-linux users.
Just be sure to not execute random scripts from the internet without properly checking what they do beforehand. Just double clicking .deb files is also not a great idea for a beginner, use your distro's package manager whenever you can.
You can use clamav to occasionally scan any new file that you download and that'd be it.
-
igormp got a reaction from AAAAAAAAAAAAAA in Framework 16 vs Framework 13 + Immersed Visor
Why not go for an eGPU solution, or, as you mentioned later, just have a desktop at home?
Do you actually use 4k at that screen size without any scaling? If so, wew, I envy your eyesight
Otherwise, if you do use scaling, going for a smaller res and just going down with the scaling factor should give you as much screen real estate. You'd be losing a bit on crispiness, but would have better battery life, so that's a tradeoff you need to consider.
Does it work with regular linux/windows devices? TIL, nice to know. Not sure how comfortable it is to use while on the go, but I never tried it out anyway.
I'm personally not a fan of large laptops, I have a powerful desktop at home that I can always remote it or just use when I'm there, otherwise I prefer 13~14" laptops with good battery life. A dedicated GPU is pretty much useless for me on a laptop, I don't play games at all in those, and for any GPU-demanding tasks I can just remote back into my machine at home, so my vote goes for the FW 13.
What about FW 13 + one of those portable external monitors? This could help with your screen real estate issue, which seems to be your biggest gripe (which I understand way too well, since I use a 42" 4k TV as a monitor lol)
-
igormp got a reaction from Avaviel in GPU docks without PSU?
Aren't you talking about some of those thunderbolt eGPU solutions that have the PSU built-in? They still need a PSU anyway, the only difference is that it's built into the product already.
I can remember gigabyte's one out the top of my head: https://www.gigabyte.com/Graphics-Card/GV-N3080IXEB-10GD-rev-10#kf