Jon Kennedy

Tomorrow at 9 AM PT, get your questions answered about NVIDIA #RTX Path Tracing. Join our #AMA to discover how you can take advantage of NVIDIA technologies to accurately recreate the physics of all light sources in a scene. Start asking questions now.

Definitely worth attending if you are interested in real-time pathtracing!

Isambard 3 announced at @ISChpc today! 384 NVIDIA Grace Superchips, 55296 @Arm cores, delivered by @HPE_Cray. Will be one of the most powerful and energy efficient systems in the UK. Hear more at the Arm BoF and AHUG workshop in Hamburg @GW4Alliance hpcwire.com/2023/05/21/nvi…

@CapFrameX @Dachsjaeger @D_S_O_Gaming Am I reading this right that disabling e-cores has 3% more frames within 2ms of each other?

@jon_c_kennedy @Dachsjaeger @D_S_O_Gaming Not big differences between configs.

*cue pokémon music* Name that engine!

@Dachsjaeger @D_S_O_Gaming I would think there *is* a difference in fluidity between configs. SMT/e-cores increase variance, depending on where different threads are run. If a thread is limiting perf and it suddenly runs on an e-core, or has to share resources on an SMT core, then it will cause variance.

@D_S_O_Gaming Those slightly different frame-rates though are not really an interesting thing for real performance. The real performance is found in fluidity, right? There is not a difference in frame-time spikes between different E, P, HT, etc. configurations.

@matiasgoldberg IIRC it can speculatively get the neighbouring cachelines. Not a good thing if you are already hitting false sharing!

Being a seasoned C++ developer means you know you need a local container variable for better performance if you're going to iterate a lot, due to how C++ rules work.

@matiasgoldberg don't forget the prefetcher which can catch you out with false sharing - might want padding[256] to be sure... 🫤

When you swap the container into a local variable, you make the problem go away. Technically it still exists, but the false sharing only happens at the beginning & end of the function. But not per iteration. Another solution would be to do:

@CapFrameX @antdavis1986 @Dachsjaeger @matiasgoldberg but isn't that due, in part, to the massive benefit of the huge L3 in the 5800X3D? Combine that cache with a fast CPU and its a killer combo! The faster CPUs get, the more other resources become the bottleneck, like L3 and DRAM.

@antdavis1986 @Dachsjaeger @matiasgoldberg @jon_c_kennedy Single threaded performance is completely overrated. An example, a 12900KS has up to 40% more single threaded performance than a 5800X3D but on average the gaming performance is very similar.

If single core CPU usage is low and GPU usage is also low then the bottleneck is neither CPU nor GPU. Could be something else like PCIe or RAM BW bottleneck, or too many thread sync points per frame, or a buggy sleep(), or the game bugs out the more cores you have.

@matiasgoldberg @Dachsjaeger HT doesn't always deliver - always worth trying to disable it to see if you get extra perf 🤔 Unless the SW threads are pinned, I wouldn't expect a core to be 100% utilised in a game.

@Dachsjaeger @jon_c_kennedy What a weird spread: The game is ignoring Hyperthreading (questionable choice, but understandable), no CPU core ever reaches 100% (?!) and 3 E cores at 18%. Maybe the missing 20% in those two 80% threads can be explained by CPU-hopping, waiting for GPU, for PCIe or E Cores

@toncijukic @matiasgoldberg @Dachsjaeger I've also seen issues a unity title with negative scaling related to core count. Too many threads trying to get work in a badly designed job system that only has a single job queue lock - results in job queue starvation.

There might be other issues at hand, but I'm not exactly certain about the culprit. I've seen a Unity title (stock engine and HDRP) go so low on CPU and GPU that GPU doesn't even clock to 3D mode and neither of the cores reached above 30%, not even E-cores. PCIE was not loaded either. Smells like some bug other than usual suspects. API overhead? Scheduler? Cache issues? Driver? It "fixed itself" some weeks later apparently as I wasn't on that project for a while before trying again and finding it fixed.

@hoytehablode @Dachsjaeger @matiasgoldberg I'd be interested in seeing if there is a perf gain by disabling the E cores

@jon_c_kennedy @Dachsjaeger @matiasgoldberg it does not help at all I tested that. and nope nothing. On a 12600k you cant get 60 fps no matter what with a gpu utilization of only 60%

@CapFrameX @Dachsjaeger @matiasgoldberg I think single thread perf is still pretty important, but cinebench MT is not representative at all of games, despite certain IHV's insistance that it is 😉 DRAM and L3 cache can have massive impacts. Ultimately though, it is all game dependent!

@Dachsjaeger @matiasgoldberg @jon_c_kennedy You cannot adapt Cinebench numbers to gaming performance. Gaming is a completely different workload. Singlethread performance isn't that important for gaming performance. Memory performance is more important.

@Dachsjaeger @matiasgoldberg HT off won't help as its not using the HT cores by the looks of it (core 1, 3, 5 etc are 0%. If you can capture an ETW/ETL file, that should tell us something. You can view by thread, then by CPU to see if it is bouncing around but fully saturating 'a' single core.

@matiasgoldberg @jon_c_kennedy I will test it with HT off and E-Cores off as a control for that behaviour you mention but I cannot atm as the game's DRM has literally locked me out of it for 24 hours since I tried the game on 2 different PCs back to back.

@Dachsjaeger @matiasgoldberg Yes - could still be CPU bound though. Either for 70% of the workload (waiting on the GPU for the rest), or the bound thread could just hop CPUs, so it is 70% on core 15 and 30% on core N == 100%

@matiasgoldberg @jon_c_kennedy Here is what the load spread looks like when set to "low" on a Core i9 12900K Notice those 2 threads (CPU 13 and 15).

@matiasgoldberg @Dachsjaeger Note that CPU utilisation shown will be for the full CPU, so could easily be CPU bound on a single thread. What CPU is he using and what freq is it clocked at? Just open up task manager to see whats going on (per logical core)...

@Dachsjaeger has anyone tried disabling CPU cores? If the game sees less cores, it may spawn fewer threads and behave differently. Also monitor PCIe Load in GPU-Z

@TheAnimator @KostasAAA Looks awesome! Let me know if I can help you with the UK stuff :D

@KostasAAA Yup! Was gunna do rendering equation stickers too :D For the UK, I’m just trying to figure out if I need to do anything about VAT 🤔

RenderThreads.com is open for business! 😱 Ever wanted graphics geek swag that wasn’t conference or business branded? Me too! It’s currently setup USA only as I figure out shipping etc, but if you send me a message I’ll see what I can do. #rendering #SmallBusiness

In the space of just four years we've somehow moved from a path-traced Quake 2 to a path-traced Cyberpunk 2077... but how? Here's @Dachsjaeger with a new Tech Focus on the hardware and software advances that made this remarkable achievement possible: youtu.be/vigxRma2EPA

Whoot - Launching on the 11th - looking forward to seeing this out there for everyone to see 😁: nvidia.com/en-gb/geforce/…

Today, we lost a visionary. Gordon Moore, thank you for everything.