• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

[NXG] Matrix UE5 PC Tech Analysis

winjer

Gold Member
This demo favors the 6000 series. The RTX 6700xt is a 13 tflops card. only 9% more powerful than the XSX. And yet it is averaging 44 fps while the XSX and PS5 average around 25 fps around the city.

Those benchmarks were done with a 12900K. That CPU is a lot faster than Zen2, especially one with cut down clocks, cache and higher memory latency.
On a demo that is so heavy on the CPU, that 12900K is doing a lot to skew results.

More importantly, it is destroying the 2080 and 2080 Super which are usually much better at ray tracing performance and on par with it in standard rasterization so clearly the game was optimized for RDNA 2.0 GPUs found in consoles. And yet, the consoles are struggling to hit 30 fps at a mere 1080p although TSR might have some kind of performance hit compared to native 1080p results we see here.

AMD has already released a first set of drivers optimized for this release of UE5. Nvidia still hasn't done that.
Also consider that a lot of shader code is optimized for RDNA2, since this is the base for current gen consoles.

Still, things dont add up. The PS5 and XSX are underperforming. They are both roughly on par with the 2080 if not the 2080 Super, but they arent performing like they should here despite the same RDNA 2.0 family cards destroying even 30 series cards on PC.

It's a CPU bound tech demo. And the PS5 and Series S/X have a good CPU, but it's not high end stuff.
 

SlimySnake

Flashless at the Golden Globes
Those benchmarks were done with a 12900K. That CPU is a lot faster than Zen2, especially one with cut down clocks, cache and higher memory latency.
On a demo that is so heavy on the CPU, that 12900K is doing a lot to skew results.

AMD has already released a first set of drivers optimized for this release of UE5. Nvidia still hasn't done that.
Also consider that a lot of shader code is optimized for RDNA2, since this is the base for current gen consoles.

It's a CPU bound tech demo. And the PS5 and Series S/X have a good CPU, but it's not high end stuff.
yep. thats precisely my point. The console GPUs are underperforming because of the low 3.5 Ghz clocks and the IO isnt exactly helping. Otherwise, we would be seeing much higher resolutions on the PS5 at the very least. I would love to see GPU utilization on the XSX because i am pretty sure it isnt maxing out here like we saw in NX gamer's benchmarks.

I am very impressed by the performance on my rtx 3080 already even before the official UE5 drivers. It's supposed to be only 65% better than my 2080, but its offering way better than 100% more performance here. My 2080 was doing 23-25 fps in 1440p and yet my 3080 gets me 40-45 fps at native 4k. Thats 2x more pixels and almost 2x more framerate.
 

winjer

Gold Member
yep. thats precisely my point. The console GPUs are underperforming because of the low 3.5 Ghz clocks and the IO isnt exactly helping. Otherwise, we would be seeing much higher resolutions on the PS5 at the very least. I would love to see GPU utilization on the XSX because i am pretty sure it isnt maxing out here like we saw in NX gamer's benchmarks.

I am very impressed by the performance on my rtx 3080 already even before the official UE5 drivers. It's supposed to be only 65% better than my 2080, but its offering way better than 100% more performance here. My 2080 was doing 23-25 fps in 1440p and yet my 3080 gets me 40-45 fps at native 4k. Thats 2x more pixels and almost 2x more framerate.

Resolution is GPU. Fragment shading is done on the GPU. The CPU does create draw calls for the GPU, but on consoles, this is unlikely to be a bottleneck.
Consoles use dynamic resolution. Meaning that if there was room to spare, the GPU would try to increase resolution. Even if the CPU is limiting frame rate.

Now one thing that might be happening is that such a heavy load on the CPU is increasing memory bandwidth contention. This results in the GPU having less for itself, limiting performance.
But this is just speculation. We would need something like a devkit on the PS5, to see how resources are being allocated and hitting performance.
 
Last edited:

GHG

Gold Member
I thought about it but Let’s face it we will all lose to those fucking bots, miners and scalpers once again and even if there is no shortage by some miracle, nvidia won’t make the same mistake of pricing the xx80 cards at $700. They will all be $1,200 like the ti and 12gb ones they came out with.

Starfield and avatar are likely going to be out by November or early next year and i just don’t see myself beating out these scalpers and bots for at least the first six months.

Yeh it will be a struggle no doubt but I'm ready to fight. Will just need to set aside a day to get one when they are launching.

Really I agree with the context of what you guys are saying, the 2700x is probably really struggling here and 3700x is a much better cpu, and 5700x much better still so it's definitely time to upgrade a 2700x these days.

But I wouldn't ever call it horrible, esp. just because of a very unoptimized engine showing its weaknesses. I am still using a 6 core version of the 2700x.

Yeh maybe I'm a bit harsh in my words towards it, but I think in the context of other CPU's available now it stacks up pretty poorly considering you have to worry about whether or not specific games/engines will run ok on it. When it works it works well, but when it doesn't... It really doesn't.

Still a very good CPU from a non-gaming perspective though.
 
I thought about it but Let’s face it we will all lose to those fucking bots, miners and scalpers once again and even if there is no shortage by some miracle, nvidia won’t make the same mistake of pricing the xx80 cards at $700. They will all be $1,200 like the ti and 12gb ones they came out with.

Starfield and avatar are likely going to be out by November or early next year and i just don’t see myself beating out these scalpers and bots for at least the first six months.
Personally I think my only hope is the evga elite early access thing... since I got my 3060 through them.

Then again their email alerts haven't notified me for newer cards like 3080 12gb and 3090 ti but maybe that's because they're the same series...?

Anyway really hoping to get an 4060/TI/70... preferably the 70 if it at least has 12gb vram. But i'll buy whichever one I can get as long as it doesn't have less than 3060 vram
 
Last edited:

Hoddi

Member
But it still could be IO related, just not total transfer size, but number of small transfers or the latency of a few key transfers that require synchronisation - such as a shader in a region you are moving towards that is outside of the current region needs compiled (and could be long and complex) or a small; data transfer that needs complete and will stall the CPU until complete.

Someone said in another thread that the PS5 transfers 10GBs more data than the PC from doing SMART analysis on PC of an nvme in the PS5 - but that could just be the section on PS5 outside of the PC sample causing that difference.
I think you're referring to my own post here.

The reason I tried allocating memory elsewhere was to prevent Windows' memory manager from caching data in system memory because it will otherwise just read the data from RAM. I was intentionally crippling the memory system to make it more IO dependent but it still didn't push read rates above a few hundred MB/s. It just didn't touch the disk much at all so I can't see how it should be IO bound.
 
Last edited:
Top Bottom