Exquisite and expensive memory architecture? I mean, if you have any information on the price Nvidia paid, please share. I'm pretty sure nobody knows. It's pretty much a logical conclusion that Samsung signed a complete package deal to Nvidia for the foundry and the memory. Nvidia isn't some nobody picking parts at the store at full price per unit.
Less expensive GDDR6 vs GDDR6x, yes definitely, but is that the whole picture? No. SRAM has a cost, enormous cost in fact, if it's not reflecting as much as a "component list of material" like GDDR6 is, it's definitely an expensive cost with how much area is dedicated to it and to be manufactured in one of the planet's best foundry.
Why are you questioning a statement just to prove it in the following sentences? The GA102 uses 10 or 12 chips of GDDR6X, which is made on a single foundry using the first PAM4 mass market implementation to date. Navi 21 uses 8 GDDR6 chips rated at 16Gbps chips that can be made by Samsung, SK Hynix and Micron. GDDR6X has an even more exclusive list of producers than HBM.
Of course the external memory subsystem and the PCB of GA102 is more expensive and therefore harder to scale down in price than Navi 21's.
You can argue whether or not the Navi 21 chip is more expensive to fab than a GA102, though considering how long 7nm has been around plus the die size difference I doubt you'd find any relevant conclusion.
And all that for.. ~20 to 30 Watts difference.. What does it say?
Can we really say it's the memory choice? The foundry node difference alone should make this an even bigger difference. Nvidia can take their 5% lead at 4K and undervolt their cards with a patch to meet AMD's performances and be like what, 60W less?
What's your suggestion here? That only Nvidia cards are able to lower the clocks and voltage to increase performance per watt?
There's a SoC out there with a 12TFLOPs RDNA2 iGPU + 8 core Zen2 + 320bit GDDR6 consuming 200W at the wall.
None of the current desktop GPUs are operating within their ideal clock/power curves.
We'll see how the notebook Ampere cards compare to the RDNA2 ones. I happen to know of a pretty good source who claims RDNA2 is most of all an architecture engineered for power efficiency because of laptops and consoles, and Navi 22/23 will excel in the mid to high-end gaming notebook market.
I'll reverse your last question : How is Nvidia competing in rasterization against a card almost exclusively made for it, when they dedicated so much to RT & ML silicon area? The fact that they're competing in rasterization performance despite that, is mind blowing.
They did not. The area dedicated to RT and ML is tiny compared to the area dedicated to general compute units, especially in Ampere where they doubled the FP32 throughput.
I can clearly see how your mind is blown, though.
I'm talking about the BVH acceleration though. No one in their right mind would prefer using the TMUs for testing intersection when they could be used for evaluating shaders. I don't understand your point here. Nvidia is ahead on this. Period.
Nvidia being ahead in RT performance doesn't mean Nvidia is ahead in producing the best real-time rendering visuals at a given cost and power consumption.
They certainly weren't chosen for Sony or Microsoft's consoles. And this time around they had their own high-performance CPU cores to make a high-performance console SoC so that excuse it out the window. And for BC, Microsoft's platform works over a virtual machine regardless.
DLSS2 is nice but it's neither widespread nor will it be without a competitor for long, according to their competitor.
UE5 demo is stellar because of it's geometry tessellation. That is all. Lumen isn't all that as there are several other custom GI solutions that are the same. You also can't possibly make something subjective be objective as fact. No matter how pretty those static rocks look (which can't deform), the overall look of the demo itself isn't very impressive. We can go into more detail if you want or maybe you can make a graphics analysis thread on the pros/cons why you think geometric tessellation is everything to a rendering pipeline. I'd gladly objectively prove to you that it's lighting/shading instead.
I think you keep missing my point. The underlying technology and how many compute cycles are being spent in what are irrelevant for the general audience. It doesn't matter if the rocks are
supposedly static (
like 99% of the geometry in videogames except maybe for characters, so point being?) or how much tessellation there is.
You can't change the fact that Demon's Souls is at the moment the best-looking 9th-gen console title according to critical acclaim, while missing out on raytracing entirely.
There's also a good number of sources claiming the dev teams of the big engine makers (idTech, Frostbite, etc.) got into disarray the day that UE5 demo released to the public and they all started working on similar approaches right away. It's also a fact that it got into mainstream media news worldwide, despite how little you apparently think of it.
And these were all done without calculating a single BVH intersection.
Compared to these mass reactions, it matters very little that you're able to pinpoint inaccuracies in any of the implementations.
To both claim to be a reputable responsible professional , and in complete opposition behave the way you do here.
AFAIK
VFXVeteran
started in the movie industry and is now in the movie industry still, not game industry. I don't know if he's worked in games in the past.
I do enjoy his analyses but the conversation at hand isn't really technical anyways.