Monday, April 29th 2024
NVIDIA Builds Exotic RTX 4070 From Larger AD103 by Disabling Nearly Half its Shaders
A few batches of GeForce RTX 4070 graphics cards are based on the 5 nm "AD103" silicon, a significantly larger chip than the "AD104" that powers the original RTX 4070. A reader has reached out to us with a curiously named MSI RTX 4070 Ventus 3X E 12 GB OC graphics card, saying that TechPowerUp GPU-Z wasn't able to detect it correctly. When we took a closer look at their GPU-Z submission data, we found that the card was based on the larger "AD103" silicon, looking at its device ID. Interestingly, current NVIDIA drivers, such as the 552.22 WHQL used here, are able to seamlessly present the card to the user as an RTX 4070. We dug through older versions of GeForce drivers, and found that the oldest driver to support this card is 551.86, which NVIDIA released in early-March 2024.
The original GeForce RTX 4070 was created by NVIDIA by enabling 46 out of 60 streaming multiprocessors (SM), or a little over 76% of the available shaders. To create an RTX 4070 out of an "AD103," NVIDIA would have to enable 46 out of 80, or just 57% of the available shaders, and just 36 MB out of the 64 MB available on-die L2 cache. The company would also have to narrow the memory bus down to 192-bit from the available 256-bit, to drive the 12 GB of memory. The PCB footprint, pin-map, and package size of both the "AD103" and "AD104" are similar, so board partners are able to seamlessly integrate the chip with their existing AD104-based RTX 4070 board designs. End-users would probably not even notice the change until they fire up diagnostic utilities and find them surprised.Why NVIDIA would make RTX 4070 using the significantly larger "AD103" silicon, is anyone's guess—the company probably has a stash of chips that are good enough to match the specs of the RTX 4070, so it would make sense to harvest the RTX 4070 out of them, which could sell for at least $500 in the market. This also opens up the possibility of RTX 4070 SUPER cards based on this chip, all NVIDIA has to do is dial up the SM count to 56, and increase the L2 cache available to 48 MB. How the switch to AD103 affects power and thermals, is an interesting thing to look out for.
Our next update of TechPowerUp GPU-Z will be able to correctly detect RTX 4070 cards based on AD103 chips.
The original GeForce RTX 4070 was created by NVIDIA by enabling 46 out of 60 streaming multiprocessors (SM), or a little over 76% of the available shaders. To create an RTX 4070 out of an "AD103," NVIDIA would have to enable 46 out of 80, or just 57% of the available shaders, and just 36 MB out of the 64 MB available on-die L2 cache. The company would also have to narrow the memory bus down to 192-bit from the available 256-bit, to drive the 12 GB of memory. The PCB footprint, pin-map, and package size of both the "AD103" and "AD104" are similar, so board partners are able to seamlessly integrate the chip with their existing AD104-based RTX 4070 board designs. End-users would probably not even notice the change until they fire up diagnostic utilities and find them surprised.Why NVIDIA would make RTX 4070 using the significantly larger "AD103" silicon, is anyone's guess—the company probably has a stash of chips that are good enough to match the specs of the RTX 4070, so it would make sense to harvest the RTX 4070 out of them, which could sell for at least $500 in the market. This also opens up the possibility of RTX 4070 SUPER cards based on this chip, all NVIDIA has to do is dial up the SM count to 56, and increase the L2 cache available to 48 MB. How the switch to AD103 affects power and thermals, is an interesting thing to look out for.
Our next update of TechPowerUp GPU-Z will be able to correctly detect RTX 4070 cards based on AD103 chips.
57 Comments on NVIDIA Builds Exotic RTX 4070 From Larger AD103 by Disabling Nearly Half its Shaders
Release all the hardware you want, there reaches a point where nobody cares anymore when the software is garbage, and software is garbage because instead of making good games based on previous sales figures, all they are doing is forcing devs to focus on the buzzwords. "Our game NEEDS AI and it needs PATH TRACING......and NFT's!!!....huh?...Oh..wait....<side conversation>...ok the NFT's can be patched-in later!".....
Smells like 83/84 again......"so all the new software is shovelware but they still want us to buy new systems every 2 years to re-play the old games but with upgraded graphics, you say?" :)
So it's less than 5%
So I tried it on a 4070 (non-super) and it was pretty lousy even with balanced DLSS, ray-reconstruction and frame-gen. Even if I was getting a reasonable framerate thanks to frame-gen, the number of actual raytracing samples it was working with at a reduced resolution every other frame made the resulting image quality absolute shit-tier garbage in motion.
Sure, path-tracing looks great in still screenshots once you've had a couple of seconds of static image to fill in all the missing data over the course of 100+ frames, but the in-motion, per-frame level of noise and distortion is abysmal on the 4070. With the 4090 having roughly double the raytracing performance in CP2077 according to this chart, I can't imagine the 4090 experience is much better. Ray reconstruction makes the reflections look pretty decent but everything else is a laggy, shadow-crawling mess of delayed lighting and out-of-date temporal information in the scene.
Honestly, I liked the raytraced reflections in CP2077 but the shadow and GI stuff isn't great even without path-tracing, it simply doesn't work in motion, only in static screenshots (by which I mean screenshots taken when the camera is static). With path tracing and ray-reconstruction, the GI and reflections are fantastic but the low framerate means that the already dubious RT shadows and ambient occlusion delay are an even worse, unmissable eyesore that stick out so badly that it completely ruins the entire image in motion.
There's simply not enough raw information per frame to handle shadows in realtime. "Slow light" simply just work in an FPS game where even when you're stood still the shadows crawly like a Lovecraftian ooze and every moving object has a bright trail absent of shadow where shadow should be at its darkest - as well as any new part of the image coming into view as you turn your view to glance sideways always being fullbright for several frames as the temporal-filter slowly fills in shadows and occlusion. It's hard to ignore when it's often half of your entire screen lit incorrectly in something as simple as turning your head to look left at any street junction or turn of any corridor.
I'm at work, but I will take a screenshot at home if I remember of CP2077 at 4K with DLSS3.5 FG RR path-traced as I turn my camera to check a street junction. From memory, these screenshots look like total ass for around 1/3rd of the image that's just come into view. Anyone accepting that as a superior image quality to baked-and-faked is crazy, in my opinion.
In-motion screenshot as I turn to to face the street on the stairs to V's Megacomplex apartment:
This screenhot is 100% authentic gameplay and highlights the shortcomings of both DLSS and raytracing with it's jaggies and godawful splotchy lighting. There's obvious visible crawl of any and all lighting in the scene and you can repeat this experiment yourself by simply enabling overdrive and hitting PrintcSreen in typical gameplay motion. The uneven shadow/lighting mess in the screenshot above isn't static, it crawls like a tentacle-endowed Lovecraftian horror so an animated .gif of this in slow-mo would look orders of magnitude worse!
What an RT Overdrive screenshot looks like if you give it a second or so (40-50 frames of temporal data):
This is what RT frames should look like, but this is the temporal average of several hundred frames and doesn't represent real gameplay! In other words, it's a hoax. The lighting is a total mess in motion and it looks like ass - nothing like this second static screenshot. Yes, I said 'ass' because I lack the vocabulary to better describe this splotchy, irregular mess, but if the 4070 isn't remotely close to a reasonable RT lighting output, I'm going to guess that the 4090 is only half as bad, and "semi-ass" is no compliment for $2000 of cutting-edge hardware that's beyond the reach of 99% of all gamers.
Meanwhile, here is what "raster" looks like (in motion) to compare against the first screenshot in this post:
This "fake" lighting will run at 160+ FPS without framegen-lag on my 4070. Youtube proves that you can get decent, true 60+fps results from entry-level GPUs costing $200, which is a miles better experience than the approximately 30fps input lag of framegen with path-tracing on far more expensive hardware on my $600 4070. Sure, the lighting is different, but it isn't anywhere near as bad as the RTX splotchy, blurry, mess. Anyone with half-decent hardware can make their own decision but I'm not personally keen on the inaccurate, Lovecraftian nightmare of RTX realtime* lighting. It's low quality, inconsistent garbage that requires you to spend quadruple on hardware for the low-quality result.
Realistically, human vision isn't a 1:1 translation of screenshots, but you need to be practically blind to miss the artifacts and problems with realtime RT. We're so far off the 'realtime' RT reality that Nvidia aspire to that I don't ever think we'll get there. Game developers increase scene complexity to match GPU capabilities. Unless a GPU magically appears with 20x more horsepower than the current expectations, the in-game scenes will always be too complex to render on any given GPU of the same generation.
* - intentional obvious sarcasm.
Let us not forget that this example I'm using is the definitive Nvidia-sponsored, Nvidia-funded, total RTX experience in its best possible light. No other game comes close to this level of DLSS and raytracing development effort.
I would guess the higer end cards of the RTX 8000 series might be the first to do real-time RT at 720p 30fps.
TBH we need a new way of rendering games, I'm not sure current style RT rendering can ever be truly realtime.
The problem is that game graphics keep getting pushed forward to match the capabilities of current GPUs, so once the RTX 8000 series arrives, we will probably have games that push an order of magnitude more polygons...
When the RTX 11070TiSuper comes out in 2034, you can be sure there will be game developers making games that will bring it to its knees and max out its VRAM.
More realistic examples of fully-raytraced games are Portal RTX and Quake II RTX. For Quake II, it took a 2080Ti to get 4K60 path-traced, in a 22-year-old game at that point in time.
I'm not sure the 2080Ti can path trace or not. I'm seeing no benchmarks anywhere so maybe it's a "30-series-and-above" feature in CP2077, the single heaviest RT title I can think of.
Not only MSI, I guess.
I bought this GIGABYTE WINDFORCE RTX 4070 and it has an AD103 chip too.