Nvidia GeForce RTX 4070 Ti: AD104 graphics chip is half the size of the RTX 4090 GPU


On January 5, 2023, the third graphics card of the Ada Lovelace architecture, previously presented under the name GeForce RTX 4080 12 GB, is to be released two months late as the GeForce RTX 4070 Ti. Now a photo of the presumed AD104 graphics chip is circulating on the internet for the first time – and it's quite small.

The AD104 is half the size of the AD102< /h2>

The GeForce RTX 4090 (test) as Nvidia's new top model will soon be on the market for two months. The GeForce RTX 4080 (test) followed exactly two weeks ago – but only with the AD103 as the graphics chip and 16 GB of graphics memory. The third RTX 4000 graphics card with a smaller AD104 GPU and 12 GB of graphics memory, which was once presented, is still a long time coming, although the forthcoming launch of an RTX 4070 Ti was only confirmed a few days ago with eight custom models from Gigabyte that appeared. It is most likely the exact graphics card that was originally supposed to appear as the GeForce RTX 4080 with 12 GB of graphics memory – but that is still not certain.

The fact that the next Nvidia graphics card will almost certainly be based on the AD104 chip is confirmed by a photo of the said GPU that Twitter user @Zed__Wang aka MEGAsizeGPU circulated. Wccftech, in turn, compared the package to a GA104, i.e. the direct predecessor from the Ampere generation, which is used in the GeForce RTX 3070 (test) and GeForce RTX 3070 Ti (test), for example. And the comparison is surprising at first glance: the supposed graphics chip of the RTX 4070 is a good deal smaller. According to Twitter user @harukaze5719, the die's area is around 295mm², while the GA104 is 392mm². For comparison: The mighty AD102 of the GeForce RTX 4090 measures 608 mm², so it is about twice as big.

< img src="/wp-content/uploads/df47bf6390e8b883e213acfa072e9c43.jpg" /> Nvidia GA104 and AD104 in comparison (Image: Wccftech)

There is no reason to worry that the GeForce RTX 4070 Ti cannot live up to its name. The recording also reveals that it is actually an AD104-400-A1, i.e. the maximum configuration without deactivated execution units. That model is said to have 60 streaming multiprocessors with a total of 7,680 FP32 ALUs, whereas the GA104 of the GeForce RTX 3070 only has 46 SMs with 5,888 execution units. The FP32 computing power is said to have even doubled. There is only a cut from 256 bits to 192 bits in the memory interface, but the actual memory bandwidth should not collapse due to an increased throughput of 21 Gbps instead of 14 Gbps – but remains below the level of the RTX 3070 Ti.

RTX 4000 Series Specifications RTX 3090 Ti RTX 4090 RTX 3090 RTX 4080 16GB RTX 3080 RTX 4070 Ti* RTX 3070 Architecture Amp Ada Amp Ada Amp Ada Amp Manufacturing Samsung 8nm TSMC N4 Samsung 8nm TSMC N4 Samsung 8nm TSMC N4 Samsung 8 Nm GPU GA102 AD102 AD102 AD104 GA104 SM 84 128 82 76 68 60 46 FP32-Alus 10.752 16.384 10.728 8.704 7.6888 RT cores 84, 2nd Gen 82, 2nd gene 76, 3rd gen 68, 68, 2nd Gen 60, 3rd Gen 46, 2nd Gen Tensor Cores 336, 3rd Gen 512, 4th Gen 328, 3rd Gen 304, 4th Gen 272, 3rd Gen 240, 4th Gen 184, 3rd Gen Base Clock 1560MHz 2230MHz 1400MHz 2210MHz 1440MHz 2310MHz 1500MHz Boost Clock 1860MHz 2520MHz 1700MHz 2510MHz 1710MHz 2610MHz 1730MHz FP32 Compute Power 40.0 TFLOPS 82.6 TFLOPS 35.7 TFLOPS 48.8 TFLOPS 408 TFLOPS 408 1 TFLOPS 20.4 TFLOPS Memory 24 GB GDDR6X 24 GB GDDR6X 24 GB GDDR6X 16 GB GDDR6X 10 GB GDDR6X 12 GB GDDR6X 8 GB GDDR6 memory interface 384 bit 384 bit 384 bit 256 bit 320 bit 192 bit t 256-bit Memory throughput 21 Gbps 21 Gbps 19.5 Gbps 22.4 Gbps 19 Gbps 21 Gbps 14 Gbps Memory bandwidth 1,008 GB/s 1,008 GB/s 936 GB/s 717 GB/s 760 GB/s 504 GB/s 448 GB/s TDP 450 watts 450 watts 350 watts 320 watts 320 watts 285 watts 220 watts Price €1,329 €1,949 €1,199 €1,469 €759 €1,099 €549 *Speculation

The smaller size, on the other hand, can be easily explained: Nvidia does not rely on an adapted 8 nm production from Samsung to manufacture the Ada Lovelace graphics chips, as it did with Gaming Ampere, but has them manufactured by TSMC in N4. With the AD102, which works in the GeForce RTX 4090, the more modern lithography resulted in around 125.5 million transistors per mm², whereas the packing density of the predecessor GA102 was only 44.6 million transistors per mm². When stepping from GA104 to AD104, the relationship will be similar. In all likelihood, this will again bring benefits in terms of power consumption, as the two larger Ada Lovelace graphics cards have already demonstrated. >In terms of performance and price, one can only speculate

In terms of performance, the expected GeForce RTX 4070 Ti will presumably be at the level of a GeForce RTX 3090 (test) or even RTX 3090 Ti (test), which means that the distance to the GeForce RTX 4080 should be in the range of 20 to 40 percent – depending on the game and resolution. The price is still unknown. The smaller GeForce RTX 4080 with 12 GB VRAM was originally supposed to cost around 1,100 euros according to the recommended retail price. It is quite possible that the GeForce RTX 4070 Ti will adopt this RRP unchanged.