?
@unknown
@wholyv C++ is already used. Tensorflow and Pytorch are written in C++ and called by Python. Python just acts as glue
827
0
0
Community Sentiment Analysis
Real-time analysis of public opinion and engagement
Sentiment Distribution
73% Engaged
37% Positive
36% Negative
Positive
37%
Negative
36%
Neutral
27%
Key Takeaways
What the community is saying — both sides
Supporting
1
Python is mostly a glue layer
almost every reply insisted that the heavy math and GPU kernels live in C/C++/CUDA (and sometimes Fortran), so Python usually just orchestrates calls into those optimized libraries.
2
Switching entirely to C++ would bring limited raw speedups
many people noted that because the GPU and underlying kernels are the bottleneck, end-to-end training/inference would see only modest gains in most cases.
3
Big trade-off
iteration speed vs runtime speed: contributors argued that C++ brings compilation overhead, more boilerplate and slower experimentation, which would likely slow down research progress even if some runtime metrics improve.
4
Production and inference already use C++
examples like libtorch, ONNX, llama.cpp and various deployment stacks show C++ (or Rust/Go) is common in production for latency, memory and I/O improvements.
5
Where C++ helps most
replies highlighted lower-latency CPU inference, improved I/O, tighter memory control and niche workloads where GPU isn’t fully utilized — cases where native implementations can outperform Python-wrapped binaries.
6
Ecosystem and tooling matter more than language
the dominant theme was that years of optimization, libraries and community tooling (PyTorch, TensorFlow, BLAS/CUDA stacks) drive performance as much or more than the frontend language.
7
Community tone was mixed but helpful
many replies patiently explained technical details and linked projects/resources, while a minority were defensive or snarky; the original poster said they learned from the exchange despite some rough responses.
8
Practical takeaway
hybrid approach wins: use Python for rapid prototyping and research, and C++/ONNX/Rust for production inference and performance-critical components — the consensus being “best of both worlds.”
Opposing
1
Many replies insist Python is just glue — most heavy computation runs in C/C++/CUDA inside libraries like PyTorch and TensorFlow, so swapping the high-level language wouldn’t move the needle for GPU-bound workloads
Many replies insist Python is just glue — most heavy computation runs in C/C++/CUDA inside libraries like PyTorch and TensorFlow, so swapping the high-level language wouldn’t move the needle for GPU-bound workloads.
2
Researchers value rapid iteration — commenters argue that forcing C++ would slow prototyping and hurt progress, with several saying productivity losses would far outweigh tiny runtime gains
Researchers value rapid iteration — commenters argue that forcing C++ would slow prototyping and hurt progress, with several saying productivity losses would far outweigh tiny runtime gains.
3
Data pipelines and preprocessing get singled out as the real pain points (GIL, memory overhead, serialization), so optimizations there could yield measurable wins even if the model kernels remain unchanged
Data pipelines and preprocessing get singled out as the real pain points (GIL, memory overhead, serialization), so optimizations there could yield measurable wins even if the model kernels remain unchanged.
4
Practical nuance acknowledged — a few notes that C/C++/CUDA and projects like llama
cpp or ONNXRuntime are already used for CPU/embedded inference and startup/warmup times can sometimes be improved with native code.
5
Alternatives and trade-offs come up frequently — some recommend Rust, Julia, Go or even Lisp for different safety or ergonomics goals, but the community is split on whether replacing Python is worth the cost
Alternatives and trade-offs come up frequently — some recommend Rust, Julia, Go or even Lisp for different safety or ergonomics goals, but the community is split on whether replacing Python is worth the cost.
6
Tone is combative and skeptical toward the original claim — many replies call the post clickbait or “ragebait,” with blunt pushback, mockery, and personal anecdotes about working in C++ for ML
Tone is combative and skeptical toward the original claim — many replies call the post clickbait or “ragebait,” with blunt pushback, mockery, and personal anecdotes about working in C++ for ML.
7
Technical consensus
any per-call Python overhead is typically negligible versus GPU compute, though metadata, startup, and inefficient Python-level preprocessing can add nontrivial overhead at scale.
8
Final point
the conversation frames this as a trade-off between marginal runtime wins and major developer ergonomics losses, with most responders siding strongly with keeping Python for orchestration.
Top Reactions
Most popular replies, ranked by engagement
?
@unknown
@wholyv Python is popular for ml because it makes it reasonably easy to develop cuda kernels. Interference/ training is not running at the speed of python, it’s running at the speed of cuda. C++ wouldn’t change the speed of that.
783
0
0
?
@unknown
@wholyv please think harder before posting next time
614
0
0
?
@unknown
@devdiary0x yea but i meant entirely
59
0
0
?
@unknown
before any of you go ahead and leave another hate comment I would like to clarify something: I am still a uni student, who comes from infra background (DevOps and shi). I do low level programming, because I love it. and I don’t know much about ML/DL (learning) since I code in LL languages, and knew that ML heavily relies on python, I wondered what outcomes would be observed had we migrated 100% to some low level language like Cpp this was the only question. which many understood, and commented nicely, explaining things. while most were insecure in their own knowledge so they started attacking in the comments. you guys are repeatedly making the same argument that many libraries eventually solve down to C/Cpp. i know mate. eventually every instruction breaks down into binary. that was not the point of my question.
43
0
0
?
@unknown
@wholyv Isn't all the real computation being done using Numpy or other frameworks anyway? No one is multiplying matrices in Python.
23
0
0