level1
patreon
Ask Questions about the Unified Architecture Video
🕑 Added 2024-06-26 16:02:32 +0000 UTCWe have an idea for a new series and we want to test it out with you guys!
Our idea is for our more discussion-heavy, future-thought, and non-review-style videos to be followed up by a live stream where you can ask Wendell more clarifying questions about the topic.
We'll start by getting your questions here, and then while the live stream is happening you can ask more questions if more clarification is needed. For now, these live streams will only be for our Floatplane and Patreon Subscribers. In the future, we may cut those live streams into short videos for our YouTube subscribers, but only you guys will be able to ask
Wendell questions.
So, for our first video, we want to answer your questions about the Unified Architecture and Developers Video. How did you like it?
Were there any parts that you would like clarified or expanded upon? If so, please ask here!
Comments
MyCookie
A disclaimer: I don't have direct experience with this part of the industry, but I try to keep up with the literature and papers when I get time. Almost a decade ago, similar solutions by AMD and Nvidia were proposed to solve slightly different set of problems. AMD designed and built components from the bottom to the top of the stack. For example, for those who wanted a similar environment in low-level HPC, AMD designed GCN's virtual memory model to be "compatible" with x86. For teams who wanted access to parallelized GPGPU hardware, but for various reasons didn't care about x86 compatibility, AMD also designed OpenCL (among others) as a software layer of abstraction away from GCN. Whereas, Nvidia went the full software route, designing CUDA to obfuscate away all hardware patterns across architectures. Sometimes, Nvidia went so far as to refuse access to hardware-level features that could help speed up GPGPU tasks, forcing everything to go through CUDA. Developers chose CUDA, even though they had: a similar landscape [to x86] in AMD's vision, and were forced to play in Nvidia's playground, by their rules. I believe a part of that is due to Nvidia's investments in GPGPU R&D, building a AI team that grew to become an industry leader, and ground level developer outreach programs in colleges and partners. AMD simply did not have the resources to compete with Nvidia in those areas. Do you agree with this assessment, and do you see similar patterns today?
