Hello Jared and thanks for the review, I understand that MSFS review was very challenging.
That said, I would appreciate if you could further deepen the CPU analysis, specifically regarding multithreading, cores load, ideal core count, and memory bandwidth. Also the lack of Ultra 285 and 9950, is very disappointing.
Thanks in advance.
I've done some testing using Lasso, trying to pinpoint where the stuttering in VR came from...
Of course I only had a 7950X3D to play with and the RTX 4090 on 4k with up to 144Hz.
I always fly the same route from Frankfurt airport over my home to Wiesbaden airbase for testing. And I typically use the AeroElvira Optica because it offers the best view of the outside and is super easy to fly. It was the ultra light airplane on earlier FS models for the same reason, I'm not training to become a pilot, but have fun exploring the planet.
I could not see any significant difference between exclusively using CCD0 (V-cache) or CCD1 (higher clocks) or just both, neither in 4k ultra with DLSS nor with VR and the primary screen dialed down to essentially THD low settings, while the headset is set to highest quality. Overall a Ryzen 9 is just bored to death on all cores.
Astonishingly enough for the VR stuttering it also made very little difference what the primary monitor was running at: Setting that to 4k and ultra and using the top VR quality settings also made very little difference: running essentially two 4k screens instead of one didn't bother the 4090 much, but also didn't solve the stuttering outside world, while the rendering of the cockpit inside VR is always extremely fluid.
Changing VR Hz from 72 to 120Hz with all the intermeditates supported made no difference to the world render stutter, but evidently there is a small difference between 72 and 90 Hz in how the cockpit renders on rapid head movement.
Basically the GPU is just as bored as the CPU, never goes near any limit, utilization, memory bandwidth, memory size, whatever resources HWinfo can list, there is plenty left over.
Changing the world update rate from 30 to something higher, mostly seems to make stuttering somewhat more erratic and causes some extra CPU consumption. I don't see it going anywhere near the maximum of 90 Hz I tried, I guess stuttering is mostly the world update rate going below the 30 Hz default when VR is active. Without VR the world is usually rather smooth, but still probably not near 90 Hz if that's what I set. I have no way of measuring that, unfortunately: there is no effective rate or skipped frame rate for the internal world updates.
Something responsible for world rendering seems to be single threaded and if because of VR more than one angle of the world needs to be rendered, it evidently can't just simply use another CPU for that or has to go around semaphores or locks with code fragments so big, it causes those stutters. When that single thread is just spread around all those cores, the CPU looks very bored, even if a single program thread uses 100% CPU. There is no thread consistently using 100%, just because I can hear you asking: I looked for it and didn't find it.
My personal impression is that that's because of a single thread limit inside the time critical part of the flight simulator, same old, same old M$ flight sim has never made good use of extra cores and that architectural limitation still hasn't been removed.
My current error theory is that modern CPUs may have moved it to a point where it's not a concern most of the time, because even a single CPU core isn't overloaded by peak simulator load, but with VR and perhaps extra angles on extra monitors or windows active, that single thread will peak above what a single physical core can deliver and then cause world render stutter, that's otherwise less noticable. You can't see it on the CPU graphs, because that loaded thread gets migrated between cores via the Windows scheduler much faster than any sampling graph could show.
So far I only tried to test just how important extra cores were by using Lasso to keep FS2024 on different core sets. As I already noted above, 3D or not didn't make much of a difference, I also tried disabling SMT with CCD0 and CCD1, because that's a "Game Mode" BIOS option today.
Nothing really happened and then I just took more cores away, again with little impact. At two cores (no SMT) I started noticing some extra lags, FS2024 runs almost as good (or bad) even on a single core as with 16.
It just cannot spread the crucial bits and unless you use VR, a single fast enough core will carry most of everything else, too. Even at 4k Ultra.
I feel safe to say that FS2024 would run just as well on a 4 core machine, even with an RTX 5090 at ultra, and especially if that at least one core on that CPU ran at 6 GHz or beyond with good IPC.
FS2024 doesn't consume significant amounts of VRAM, with dual 4k screens and the VR headset somehow involved I never saw it use more than around 11 of 24GB of GPU memory.
I didn't see it use any significant part of the 96GB of DDR5-6000, nor does HWinfo show any significant RAM bandwidth being used. Basically I used HWinfo and Lasso on the 2nd 4k screen to search for anything that might be a bottleneck and the only thing I could identify is that the single core it uses for everything critical, would need to run at 10GHz or more.
In summary: FS2024 pretty much like FS2020 is pretty near only limited by the speed of the fastest CPU core. Most of the time that's good for better than 60Hz, unless you do VR: then the simulated world stutters in ways it never would in a real plane, which kills immersion faster than a crash would kill the pilot. No currently available CPU will give you 240 Hz, perhaps not even an RTX 4090 at 720p. Throwing extra hardware at it, just doesn't help because the issue is software.
As far as I can tell, under the hood the FS2024 core isn't that much changed from the 2002 edition which M$ sells every few years as something completeley new.
The code that invents the outside world, streets, vegetation, houses, cars etc. when there is only 2D data in Microsoft's card data has not changed much in terms of quality and realism since Flight Simulator X.
It has not changed at all since FS2020 even if it has moved to the cloud, because the rather stately 1830 villa where I live in the servants quarters under the roof, on FS2024 looks like the exact same generic cottage it is on FS2020. And the low barn on the other side of the field, has been a rather substantial multi-floor Hamburg harbour style storage house, the likes of which FS likes to sprinkle generously across all of Germany for some reason. Everything up to the color of each individual house on my street is exactly the same on FS2020 and FS2024 and apart from a rough outline has zero similarity to what's actually on the ground.
You paint a house in that red in Germany, you get put into prison before you're half finished, but Microsoft doesn't know that, not even with OpenAI.
And real cars don't drive through rivers, across fields or straight into each other on four lane highways. Again, it should only take a few ten thousand weights to teach that to a machine and that model would easily run on the tiny CPU inside the TPM, perhaps even on the old 80486 on PCHs, much less powerful than all those idling E-cores.
M$ Flight simulator is so pathetic it's must be inspiring: so many words to prove it, sorry!
Facebook
Twitter
Instagram