Microbenchmarking Ada and RDNA 3

[trinibwoy]

I was looking for VOPD info and came across a few excellent articles from the usual suspects.

Microbenchmarking AMD’s RDNA 3 Graphics Architecture

Editor’s Note (6/14/2023): We have a new article that reevaluates the cache latency of Navi 31, so please refer to that article for some new latency data. RDNA 3 represents the third iteratio…

chipsandcheese.com

Microbenchmarking Nvidia’s RTX 4090

Nvidia’s RTX 4090 features Nvidia’s newest architecture, named Ada Lovelace after a pioneer in early computing. Compared to their previous architecture, Ampere, Ada Lovelace enjoys a pr…

chipsandcheese.com

Bonus: RDNA 2 deep dive

AMD’s RDNA 2: Shooting For the Top

In 2019, AMD moved off their long-serving GCN architecture in favor of RDNA. We’ll cover the first generation of RDNA some other time. RDNA 2 takes that foundation and scales it up while addi…

chipsandcheese.com

 

[Arun]

Chips and Cheese are doing a great job. Is there any other website left that does any of this analysis for GPUs, and/or good papers, and/or any useful open source microbenchmark suites?

Last time I looked at this, it was all pretty depressing. There's some fairly good analysis of NVIDIA GPUs focusing on Deep Learning & HPC (e.g.: the "Dissecting" Volta/Turing/etc. papers from Citadel) but very little that's graphics-related, and there's loads of missing pieces in the compute puzzle as well.

I've been seriously considering spending a few weeks making a new open source Vulkan microbenchmark suite, assuming my employment status allows me to do so going forward (currently in the process of hopefully moving from full-time architect at Imagination to being a part-time contractor/consultant for them).

Any thoughts or interest in that? It's a pretty niche field these days, especially for anything outside of Deep Learning (which would be my next focus after this as well), and there's no way I'd have the time to do long-term maintenance of those tests unless it becomes a business in itself, but it frustrates me how much low hanging fruit there is here.

 

[trinibwoy]

There really isn’t anyone doing even basic micro benchmarks any more. Most of the good info comes from developer presentations (GDC, Siggraph etc) and the occasional academic paper or patent. Chipsandcheese is an oasis in the desert.