News Alleged Nvidia AD106 and AD107 GPU Pics, Specs, Die Sizes Revealed

[Admin]

Nvidia's AD106 and AD107 graphics processors are small, but will they be cheap?

Alleged Nvidia AD106 and AD107 GPU Pics, Specs, Die Sizes Revealed : Read more

 

[bigdragon]

Good to see the full specifications! I'm concerned there's going to be a performance regression given the reduction in CUDA cores and memory interface. Some early leaks are showing RTX 3070, 3060, and 3050 chips outperforming RTX 4070, 4060, and 4050 chips, respectively. I don't see how Nvidia could justify another price increase when generational performance declines.

 

[InvalidError]

I don't see how Nvidia could justify another price increase when generational performance declines.

Easy: "because we can."

Not like AMD and Intel are a credible threat to Nvidia's stranglehold on the graphics market yet.

 

[helper800]

Easy: "because we can."

Not like AMD and Intel are a credible threat to Nvidia's stranglehold on the graphics market yet.

The dark ages of tech is a higher cost, and lower performance compared to same tier previous gen parts. Are we entering the dark age of laptop gpu performance?

 

[-Fran-]

Here in the UK, the 4090 laptops are north of £4000, with some being over £5000. It's ludicrous.

Given the price structure, all this generation is going to be "top heavy", for sure. That $1000 base price is going to be so barebones it's going to suck for whoever buys such a laptop, I'd say.

I really hope AMD can put up a fight on the mid and lower segment with Navi32 and Navi33. I am not sure if Intel will even try to compete with any ARC models in laptops, but I sure hope they do after the OEMs have enough confidence in their drivers. As a side note (speculative): I'd imagine that is why no major/big OEM has put an ARC GPU in a laptop yet? I only know of Samsung having like 1 model and that's it?

Regards.

 

[renz496]

Good to see the full specifications! I'm concerned there's going to be a performance regression given the reduction in CUDA cores and memory interface. Some early leaks are showing RTX 3070, 3060, and 3050 chips outperforming RTX 4070, 4060, and 4050 chips, respectively. I don't see how Nvidia could justify another price increase when generational performance declines.

never saw this rumor. 4070Ti already at 3090 level except at 4k. somehow 4070 will be slower than 3070?

 

[InvalidError]

never saw this rumor. 4070Ti already at 3090 level except at 4k. somehow 4070 will be slower than 3070?

With Nvidia (and AMD) slotting in smaller dies with narrower memory bus and in many cases less memory for the same marketing tier as previous gen, there is bound to be cases where the new die performs worse than the one it is supposed to replace. I can definitely see why people are getting nervous about where this is going.

 

[bigdragon]

never saw this rumor. 4070Ti already at 3090 level except at 4k. somehow 4070 will be slower than 3070?

I wasn't clear that I meant laptops and not desktops. My bad. The performance regression rumors I've been reading are specific to laptop GPUs. Videocardz and random Twitter Chinese resellers have posted disappointing benchmarks for the 40-series mobile chips showing little generational improvement at 1080p. The potential exists for a regression at 1440p. I'm not following the desktop equivalents.

 

[edzieba]

With Nvidia (and AMD) slotting in smaller dies with narrower memory bus and in many cases less memory for the same marketing tier as previous gen, there is bound to be cases where the new die performs worse than the one it is supposed to replace. I can definitely see why people are getting nervous about where this is going.

There may be memory-limited edge cases for GPU compute, but I cannot recall any being found for gaming yet. e.g. the 4070Ti continues to perform around the 3090 level (usually between 3090and 3090Ti) despite having HALF the memory capacity, buswidth, and bandwidth. The Ada architecture is clearly a lot less sensitive to memory performance than previous architectures.

 

[InvalidError]

There may be memory-limited edge cases for GPU compute, but I cannot recall any being found for gaming yet.

Today's games, maybe, since they do have to target 4+ years old sub-$400 MSRP hardware to get a decent size audience capable of running them decently well.

A few more years down the line where games at near-max settings may commonly use more than 12GB worth of assets and buffers though, the 3090 could age much better than the 4070Ti.

 

[edzieba]

Today's games, maybe, since they do have to target 4+ years old sub-$400 MSRP hardware to get a decent size audience capable of running them decently well.

A few more years down the line where games at near-max settings may commonly use more than 12GB worth of assets and buffers though, the 3090 could age much better than the 4070Ti.

It's going to be many, many years before developers can start assuming more than 4/8GB as a GPU capacity baseline (i.e. when 12GB+ becomes the 'cheap card' baseline, which is not going to be any time soon). And at the same time, NVME drive ubiquity can also be assumed, so asset streaming comes along to alleviate that potential bottleneck.

 

[-Fran-]

It's going to be many, many years before developers can start assuming more than 4/8GB as a GPU capacity baseline (i.e. when 12GB+ becomes the 'cheap card' baseline, which is not going to be any time soon). And at the same time, NVME drive ubiquity can also be assumed, so asset streaming comes along to alleviate that potential bottleneck.

Have you ever heard of "VR"?

Not trying to be a smarty pants here, but I truly wonder why no one even considers that, most of the use for these high end cards nowadays is 4K gaming alone. VR games can easily push 12GB if they want to. Alyx can push 10GB and it's not even an open game. Plus, you can't have massive FPS drops or you are prone to puking all over.

The only reason I had to get a 6900XT is because I was in need of VRAM. And I'm seriously considering getting a 7900XTX because of the 24GB and excellent raster performance for the price.

Regards.

 

[InvalidError]

It's going to be many, many years before developers can start assuming more than 4/8GB as a GPU capacity baseline (i.e. when 12GB+ becomes the 'cheap card' baseline, which is not going to be any time soon). And at the same time, NVME drive ubiquity can also be assumed, so asset streaming comes along to alleviate that potential bottleneck.

Looks like the part about "near-max settings" flew 100' over your head. We are talking RTX3090 vs RTX4070Ti staying power here, not budget gaming.

Current games with graphics maxed out already push 10+GB of VRAM and that earned Nvidia a fair amount of criticism from reviewer as insufficiently future-proof with modern games already pushing 10+GB on Ultra.

Streaming assets from NVMe at 15GB/s including decompression gain is no substitute to sufficient VRAM at 300+GB/s.

 

[renz496]

Looks like the part about "near-max settings" flew 100' over your head. We are talking RTX3090 vs RTX4070Ti staying power here, not budget gaming.

Current games with graphics maxed out already push 10+GB of VRAM and that earned Nvidia a fair amount of criticism from reviewer as insufficiently future-proof with modern games already pushing 10+GB on Ultra.

Streaming assets from NVMe at 15GB/s including decompression gain is no substitute to sufficient VRAM at 300+GB/s.

future proof is one of the main enemy for hardware maker.

 

[InvalidError]

future proof is one of the main enemy for hardware maker.

At least, as far as VRAM burden is concerned, budget buyers willing to compromise on details have the option of lowering texture resolutions to make it fit as games become more VRAM-hungry and those GPUs can still get a decent useful life, albeit possibly a rung or two behind on the performance ladder from where they could have been.

 

[edzieba]

Have you ever heard of "VR"?

Yes, very well indeed, since prior to the DK1 being announced. There is nothing inherent to a low-latency rendering pipeline that requires large quantities of VRAM, and the popularity of standalone HMDs with both some of the highest panel resolutions and less than 10GB of total system RAM indicates this is no more an inherent requirement as render resolution (which also only scales very loosely with VRAM requirements). VRAM burden of current games is vastly overstated.

Looks like the part about "near-max settings" flew 100' over your head. We are talking RTX3090 vs RTX4070Ti staying power here, not budget gaming.

Current games with graphics maxed out already push 10+GB of VRAM and that earned Nvidia a fair amount of criticism from reviewer as insufficiently future-proof with modern games already pushing 10+GB on Ultra.

Streaming assets from NVMe at 15GB/s including decompression gain is no substitute to sufficient VRAM at 300+GB/s.

Remember that VRAM occupied != VRAM required due entirely to opportunistic asset caching, and its asset caching that DirectStorage and similar techniques target. VRAM required for rendering is far lower, but identifying it is far more invovled than seeing what number a process monitor spits out.
'Max settings' is no indication of actual requirements either, as those are not optimised for any actual available GPU but simply based on how far the various parameters an engine has to tweak can go without much effort (e.g. nobody is going to re-render textures just to allow for a higher texture resolution setting, but if there is an option to push out LoD distance where those textures are used without breaking rendering that's an low-effort add to the ULTRA MAX NIGHTMARE EXTREME BBQ preset). Anyone actually playing rather than benchmarking uses the visually identical but actually optimised present one stop below the maximum-e-peen setting, possibly with some slider tweaking for visual preference (though usually that ends up being counterproductive).

 

[InvalidError]

Remember that VRAM occupied != VRAM required due entirely to opportunistic asset caching, and its asset caching that DirectStorage and similar techniques target. VRAM required for rendering is far lower, but identifying it is far more invovled than seeing what number a process monitor spits out.

It is simple: run a background task that hogs VRAM and see how much extra VRAM the background task can consume before frame rate starts dropping.

Also keep in mind that Chrome, Firefox and a whole bunch of other things one may have running on the background or secondary monitors can consume well over 1GB of VRAM each when using hardware acceleration. I have to disable hardware acceleration in everything I can with my GTX1050, otherwise many games are glitchy as heck from having almost no spare VRAM to work with. Windows itself can use 400+MB of it just to render the desktop. You need to budget enough VRAM for all of that stuff too.

 

[-Fran-]

This may be relevant to the VRAM topic:

https://www.techpowerup.com/review/hogwarts-legacy-benchmark-test-performance-analysis/6.html

EDIT:

View: https://www.youtube.com/watch?v=qxpqJIO_9gQ


Regards.

 

[InvalidError]

This may be relevant to the VRAM topic:

https://www.techpowerup.com/review/hogwarts-legacy-benchmark-test-performance-analysis/6.html

Long story short: if you want to play it mostly maxed out, you need 11GB of VRAM at the absolute minimum. The RTX2080Ti with 11GB of VRAM and 12GB RTX3060 completely destroy the 8GB 3070Ti. While the 10GB RTX3080's performs similarly on average, its 1% lows get tanked too.

8GB of VRAM isn't enough for anything more powerful than an RTX3050.

 

[edzieba]

Long story short: if you want to play it mostly maxed out, you need 11GB of VRAM at the absolute minimum. The RTX2080Ti with 11GB of VRAM and 12GB RTX3060 completely destroy the 8GB 3070Ti. While the 10GB RTX3080's performs similarly on average, its 1% lows get tanked too.

That's not what the graphs show.
In average framerates, the 3070 (3070 Ti is not listed) performs essentially the same as the 2080Ti, just as it does in other benchmarking, and the 3060 performs consistantly behind both cards, with the gap widening as render resolution increases.
In minimum framerates, the 3070 performs essentially the same as the 2080Ti (as before) and the 3060 performs consistantly behind both cards, with the gap widening as render resolution increases (as before). Except at UHD with raytracing enabled, where the 2080Ti pulls ahead of both the 3070 and 3060 by a significant margin (with the 3070 and 3060 performing similarly).

Or in other words: if VRAM capacity is a factor, it's only one that comes into play with Raytracing enabled, at UHD, and only for minimum frametimes. Outside of that situation, the 3070, 2080Ti, and 3060 have the same relative performance as tested in other games.

 

[KyaraM]

Long story short: if you want to play it mostly maxed out, you need 11GB of VRAM at the absolute minimum. The RTX2080Ti with 11GB of VRAM and 12GB RTX3060 completely destroy the 8GB 3070Ti. While the 10GB RTX3080's performs similarly on average, its 1% lows get tanked too.

8GB of VRAM isn't enough for anything more powerful than an RTX3050.

That game is so horrendously programmed with performance issues on such a scale over the entire GPU stack it's not even worth looking at right now. Also, iirc the test was from pre-launch; it got a day 1 patch that helped a lot, and community fixes as well. Besides, as stated directly above, your assessment isn't even correct...

Btw, AMD CPUs are worse in this game than Intel and seem to show some weird behavior, so that is a factor, too.

Edit: my 3070Ti at 1440p is getting well over 60FPS in Hogsmeade with DLSS. It's very, very far away from the 17FPS HUB got.