What's so special about Nvidia architecture
Be careful not to post any IP related information, just looking for publicly available high level information about what makes GPU architecture so special. My limited understanding is that GPUs are simpler than CPUs and have lesser architectural complexity. How is Nvidia running so much ahead of the "competition"? Same can be asked about apple processors. Publicly leaked info or Tom's hardware kind of analysis is welcome.
Ease of programming.
How long can that be a leverage, if the market is big enough, I assume companies are likely to invest in optimizations and better hw/sw mapping to exploit the hardware.
Intel has been trying that since mid-2000. OpenCL didn’t pan out as most would have liked. In general software is big part and other HW vendors are not there yet. Things of course might change but NV can adapt too.
Google "Dissecting the NVIDIA Volta GPU Architecture via Microbenchmarking"
Sure, will check that, anything in particular?
No, but you get a lot of details on the architecture. Their value though is in the general ecosystem.
cuda and the focus and belief on GPU are the key ingredients of nvidia success
I agree, single minded execution does stand out. Semiconductor leadership should learn from this. I guess move fast and break things may not work in hardware, hardware is 'hard'
I believe this is due to nvidia’s heavy emphasis on making the quality of their software stack world class, which has led to many performance improvements and optimizations in the overall product. With this in mind they focus a lot on hardware acceleration (thereby making new innovations in their hardware architecture) and improving the overall hardware-software system. I have not seen any other hardware company investing so much to improve the performance of their software stack (except maybe apple).
@All fellow Intelers - Good point! Why can't Intel do the same and improve their software??? Is OneAPI going to change anything? Does Intel even have the capacity to pull this off with their lower payed (and thus less motivated or skilled) SWEs?
I don't think it has to do with the lack of talent but more to do with management's vision and direction. Sometimes it's too little too late. Only a handful teams or sometimes people doing all the heavy lifting and not much recognition. The actual engineer might have no idea of the real impact. Sometimes even subtle things also can be of great help to customers. Like more telemetry or better methods, better APIs.
Nvidia success is due to CUDA
What's special about CUDA, its been around for a while