NVIDIA AD107 Silicon Powering GeForce RTX 4060 Pictured

[btarunr]

The "AD107" is expected to be the smallest client GPU based on the NVIDIA "Ada Lovelace" graphics architecture. The upcoming GeForce RTX 4060 (non-Ti) is rumored to be maxing this silicon out (enabling all available shaders). MEGAsizeGPU scored one of the first pictures of the "AD107" in the flesh, revealing a small fiberglass substrate, and a visibly smaller die than the "AD106" powering the RTX 4060 Ti.

The "AD107" silicon is expected to feature 3,072 CUDA cores across 24 SM (streaming multiprocessors). The GPU features a 128-bit wide memory interface much like the "AD106," and NVIDIA is expected to use conventional 17 Gbps GDDR6 memory, which works out to 272 GB/s memory bandwidth. 8 GB is the standard memory size for the RTX 4060. With "Ada," NVIDIA has rebalanced the memory sub-system with greater reliance on on-die caches, and the "AD107" features a 24 MB L2 cache. Much like the "AD106," the smaller "AD107" features a PCI-Express 4.0 x8 host interface. The ace up its sleeve has to be power, with even the maxed out RTX 4060 only being rated for 119 W of TGP.



View at TechPowerUp Main Site | Source

 

[Space Lynx]

monolithic design does look prettier than chiplet, I have to admit

 

[BoboOOZ]

This thing is tiny and it has just 4 SM more than the 3050. This better be dirt cheap...

 

[dj-electric]

Making these... and AD106 based products is... should be... relatively cheap. This is good, but not for us. We're a couple of days away from what could be NVIDIA's highest profit margin product based on a small sized silicon.

 

[Dr. Dro]

This thing is tiny and it has just 4 SM more than the 3050. This better be dirt cheap...

Nvidia and dirt cheap don't belong in the same sentence, I'm afraid

 

[Denver]

Is there already an estimate for the die size?

 

[usiname]

Is there already an estimate for the die size?

According Videocardz is 146 mm²
"The GPU is even smaller than its predecessors, measuring only 146 mm², while it is typically 200 mm² for this class."

NVIDIA AD107-400 GPU for GeForce RTX 4060 has been pictured - VideoCardz.com

NVIDIA AD107-400 has been pictured While NVIDIA might have been reluctant to confirm which GPU specifically goes into the RTX 4060 graphics card, hardware leaker MEGAsizeGPU leaves no doubt it is the AD107-400. Not all board partners were ready to showcase their RTX 4060 non-Ti graphics cards...

videocardz.com

 

[BoboOOZ]

Nvidia and dirt cheap don't belong in the same sentence, I'm afraid

Either dirt cheap or collecting dust, I guess. This is really a 4050.

 

[Camm]

Would want to see this in a no external power 75w solution personally.

 

[thunderingroar]

looks as tiny as the gddr chip lol

 

[Octavean]

Either dirt cheap or collecting dust, I guess. This is really a 4050.

Not sure what constitutes “dirt cheap” in dollars and cents but I’ve heard the GeForce RTX 4060 (non-Ti) should be about ~$300 USD MSRP. So cheaper then it‘s predecessor the 3060 (non-Ti) if accurate although you never know when it comes to actual street price. However, any leftover RTX 3060 (non-Ti) stock should receive a price reduction.

 

[Dr. Dro]

Either dirt cheap or collecting dust, I guess. This is really a 4050.

I'd argue it's a 4030, really. GP107 was used in the GT 1030, TU117 in the GTX 1650 and low profile T-series enterprise cards, and GA107 used in mobile MX/RTX 3050, the pattern is always the lowest segments. Here it's being pitched in the 60 class.

 

[Chrispy_]

Nvidia and dirt cheap don't belong in the same sentence, I'm afraid

You are so wrong!
"Nvidia love setting sky-high MSRPs on GPUs that are dirt cheap for them to make"

 

[BoboOOZ]

I'd argue it's a 4030, really. GP107 was used in the GT 1030, TU117 in the GTX 1650 and low profile T-series enterprise cards, and GA107 used in mobile MX/RTX 3050, the pattern is always the lowest segments. Here it's being pitched in the 60 class.

Well I'm giving them the benefit of the doubt because going from Samsung 8nm to TSMC 4nm is actually a 2 generation jump, I could understand they want to give us just half of the benefits of that big die shrink. But it seems like they are trying to turn this into giving us just 15% gen on gen and turning all the rest into green for them. They just got addicted to too much cash flowing in, they don't want to decrease margins in spite of the market changing completely.

 

[Dr. Dro]

Well I'm giving them the benefit of the doubt because going from Samsung 8nm to TSMC 4nm is actually a 2 generation jump, I could understand they want to give us just half of the benefits of that big die shrink. But it seems like they are trying to turn this into giving us just 15% gen on gen and turning all the rest into green for them. They just got addicted to too much cash flowing in, they don't want to decrease margins in spite of the market changing completely.

Nvidia's margins are 25% higher than that of Apple's. Let that sink in.

Despite being childish, "Ngreedia" is quite the accurate insult towards them

 

[AusWolf]

Can we have one in 75 W, low profile form without a power connector for £150 please? Pretty please?

 

[Chrispy_]

Nvidia's margins are 25% higher than that of Apple's. Let that sink in.

View attachment 297162

View attachment 297163


Despite being childish, "Ngreedia" is quite the accurate insult towards them

65% margin is obscene. 25% margin is still much higher than the industry average.

If Nvidia cards were priced even remotely fairly, we'd have the 4060Ti at 125/165 of $399, ie it should cost $299 but Ngreedia!
(am I doing that right?)

Hot damn we need AMD and Intel to take some market share. That's only going to work if people stop buying into the Nvidia BS promises of DLSS and RTX magic. RTX still only works in under 5% of games released every year and DLSS3 frame-gen is just 40ish titles out of, uh, all the the tens of thousands of AAA games ever made.

 

[Cowboystrekk]

This thing is tiny and it has just 4 SM more than the 3050. This better be dirt cheap...

Clock runs almost 1000MHz above 3050 though.

Can we have one in 75 W, low profile form without a power connector for £150 please? Pretty please?

4050 will do that I bet, but price eill probably be 200+ :/

 

[AusWolf]

Clock runs almost 1000MHz above 3050 though.

Clocks don't matter. Performance does. If you want clocks, I suggest buying a 6500 XT. You can max out the power and overclock sliders on basically any model, and run it just a hair below 3 GHz.

4050 will do that I bet, but price eill probably be 200+ :/

Probably. Though I don't think the 4060 (Ti) would lose much performance when limited to 75 W. It surely won't happen, though. A low-power low-profile Nvidia gaming card has been a pipe dream since the 1650.

 

[Cowboystrekk]

Clocks don't matter. Performance does. If you want clocks, I suggest buying a 6500 XT. You can max out the power and overclock sliders on basically any model, and run it just a hair below 3 GHz.

I meant compared to the 3050. With clocks being that much higher, performance will also be quite a bit higher even with the 96-bit memory interface. A 4050 is not that far behind a 3060 so it eill crush the 3050

RTX 4050 vs 3060 Laptop Comparison - 25 Games Tested!

 

[BoboOOZ]

Clock runs almost 1000MHz above 3050 though.

Fair point, but clocks do go up with die shrinks, it happens at each generation, for Nvidia and AMD. You can argue it's a 4050 ti, but it's hard to argue it's really a 4060.

 

[AusWolf]

Fair point, but clocks do go up with die shrinks, it happens at each generation, for Nvidia and AMD. You can argue it's a 4050 ti, but it's hard to argue it's really a 4060.

xx107 chips have always been part of the xx50 series of cards, to be honest. Its small size and memory bus proves the point.
Nvidia does everything they can to milk gamers by throwing them a never needed DLSS3 bone, that's all what it is, imo.

 

[Cowboystrekk]

Fair point, but clocks do go up with die shrinks, it happens at each generation, for Nvidia and AMD. You can argue it's a 4050 ti, but it's hard to argue it's really a 4060.

Yeah, point was that clocks somewhat compensate for lack for SMs, cudacores etc

 

[AusWolf]

Yeah, point was that clocks somewhat compensate for lack for SMs, cudacores etc

Clocks shouldn't compensate for anything. They should go hand-in-hand with core upgrades to make a better product. It's a smaller node with smaller die space needed per SM, so why not? Oh I know! Money.

 

[BoboOOZ]

Well, I don't think any of us argue that Nvidias's pricing and positioning for everything under the 4090 is fair for this generation. They just want to hold on to those huge margins, and it seems that it doesn't bother them they lose gaming market share since there is the AI bubble to thrive on. Their stock just went up by 25% so their strategy must be sound. What is worse, AMD will also join in the AI race. The next generation might be much worse for gamers than this one, if the AI bubble is more than a bubble (and it definitely looks so).