HAPPY HOLIDAYS to all.
Our fun holiday project here at Numerion was to build and benchmark a Threadripper 3970X based machine.
It's been 10+ years since we've considered AMD cpus worthy of testing, and of course with no effective competition in this space,
as we all know Intel has not exactly been pushing the performance forward. Yes, they have been giving us extra cores (for some crazy prices),
but cloth simulation also benefits from increased clock speed.
Well finally, along comes AMD and knocks it out of the park. A single CPU with lots of cores, sustained all core high clock speed AND large cache -
but without the rip-off price tag we've come to expect. What more can we want for simulation?
So, what did we find?
Well it's a monster improvement - the stock speed 3970X, with its 32 cores and 150Mb cache, ran typically around 4.2GHz with 100% Carbon simulation loading (vs typically 3 - 3.2GHz with our Xeons CPUs). Right there that clock speed increase, combined with the large caches to support the throughput, was turning in a 30% performance gain - vs comparable number of Xeon* cores.
Some components of Carbon ran 60% faster and we are investigating how that can be and will look to further leverage what we find.
The best part of course is that Intel will have to respond as AMD are taking most of the consumer desktop cpu sales right now. Meanwhile AMD is not sitting back, they are promising a 15% IPC gain and reduced cache latency with next gen, all great for simulation.
So, for the first time in maybe 7+ years, we are likely to see several back to back years of significant CPU driven cloth simulation performance improvement.
*Note we are aware that there is a Xeon W-3175X that's quite quick, but without serious overlocking it's still not as quick as the stock 3970X and then the Xeon also costs 60+% more as well!