ADIOS2 VS AdaptiveCpp

I thought I'd share something with my experience with HPC that applies to many areas, especially in the rise of GPUs.

The main bottleneck isn't compute, it is memory. If you go to talks you're gonna see lots of figures like this one[0] (typically also showing disk speeds, which are crazy small).

Compute is increasing so fast that at this point we finish our operations long faster than it takes to save those simulations or even create the visualizations and put on disk. There's a lot of research going into this, with a lot of things like in situ computing (asynchronous operations, often pushing to a different machine, but needing many things like flash buffers. See ADIOS[1] as an example software).

What I'm getting at here is that we're at a point where we have to think about that IO bottleneck, even for non-high performance systems. I work in ML now, which we typically think of as compute bound, but being in the generative space there are still many things where the IO bottlenecks. This can be loading batches into memory, writing results to disk, or communication between distributed processes. It's one beg reason we typically want to maximize memory usage (large batches).

There's a lot of low hanging fruit in these areas that aren't going to be generally publishable works but are going to have lots of high impact. Just look at things like LLaMA CPP[2], where in the process they've really decreased the compute time and memory load. There's also projects like TinyLLaMa[3] who are exploring training a 1B model and doing so on limited compute, and are getting pretty good results. But I'll tell you from personal experience, small models and limited compute experience doesn't make for good papers (my most cited work did this and has never been published, gotten many rejections for not competing with models 100x it's size, but is also quite popular in the general scientific community who work with limited compute). Wfiw, companies that are working on applications do value these things, but it is also noise in the community that's hard to parse. Idk how we can do better as a community to not get trapped in these hype cycles, because real engineering has a lot of these aspects too, and they should be (but aren't) really good areas for academics to be working in. Scale isn't everything in research, and there's a lot of different problems out there that are extremely important but many are blind to.

And one final comment, there's lots of code that is used over and over that are not remotely optimized and can be >100x faster. Just gotta slow down and write good code. The move fast and break things method is great for getting moving but the debt compounds. It's just debt is less visible, but there's so much money being wasted from writing bad code (and LLMs are only going to amplify this. They were trained on bad code after all)

[0] https://drivenets.com/wp-content/uploads/2023/05/blog-networ...

[1] https://github.com/ornladios/ADIOS2

[2] https://github.com/ggerganov/llama.cpp

[3] https://github.com/jzhang38/TinyLlama

Sapphire Rapids is a CPU.

AMD's primary focus for a GPU software ecosystem these days seems to be implementing CUDA with s/cuda/hip, so AMD directly supports and encourages running GPU software written in CUDA on AMD GPUs.

The only implementation for sycl on AMD GPUs that I can find is a hobby project that apparently is not allowed to use either the 'hip' or 'sycl' names. https://github.com/AdaptiveCpp/AdaptiveCpp

Not natively, but AdaptiveCpp (previously hiSycl, then OpenSycl) has a single source single compiler pass, where they basically store LLVM IR as an intermediate representation.

https://github.com/AdaptiveCpp/AdaptiveCpp/blob/develop/doc/...

Performance penalty was within ew precents, at least according to the paper (figure 9 and 10)

AdaptiveCpp (formerly known as hipSYCL) is an independent, open source, clang-based heterogeneous C++ compiler project. I thought some of you might be interested in knowing that we recently added support to offload standard C++ parallel STL algorithms to GPUs from all major vendors. E.g.:

In fact SYCL was initially called hipSYCL because it is based on AMD's ROCm/HIP. AMD had hipSYCL code running on the Frontier supercomputer four years ago at least and continues to support it.

Apple Silicon support through Metal is something that is actively discussed in hipSYCL. See https://github.com/illuhad/hipSYCL/issues/864 https://github.com/illuhad/hipSYCL/issues/460 (loooong discussion)

But really, the DPC++ part of oneAPI (which is many APIs) is just SYCL + extensions, and there are several other SYCL implementations which have already featured CUDA and Hip (AMD) support for a long time. The most popular and widely-used is hipSYCL, which we've been using in an HPC context on NV hardware for over 4 years now.

Unfortunately, the AMD and Nvidia plugins are proprietary. AMD users are probably better served with hipSYCL, if they somehow find an application using SYCL...

Also, you might want to take a look at an implementation like hipSYCL :)

Yup. SYCL is the future: https://github.com/illuhad/hipSYCL

ROCm is completely independent from these. It's a compute stack containing an OpenCL implementation for Radeon GPUs, plus a CUDA-like language called HIP which can be compiled to either device code for Radeon GPUs or to PTX to work with Nvidia GPUs. However, some researchers also created hipSYCL that allows SYCL to run atop HIP; you can think of it like DXVK - the program contains the DirectX/SYCL API, and DXVK/hipSYCL converts it to Vulkan/HIP (with one difference - DXVK does the conversion at runtime, while hipSYCL does it at compile time).