RTX 4070 will be getting more units and not just 160-bit memory
We have been getting leaked spec bits of the upcoming Nvidia GeForce RTX 4000 cards based on the “Lovelace” architecture for some time. But things still aren’t set in stone yet (or at least not for all models), so things are still subject to change. And that is what has now happened with the GeForce RTX 4070 SKU. Nvidia is said to have changed the specs and give this graphics card higher performance and less cut-down memory.
This information was brought to us yet again by leaker Kopite7kimi, who at least in the past had the best information about Nvidia’s plans. According to previous reports, the GeForce RTX 4070 was supposed to use the AD104 chip (in version AaD104-275), which was supposed to have enabled 7168 shaders (56 SM blocks) out of its 7680 shaders (60 SM).
But also the memory subsystem was supposed to be stripped down. The AD104 GPU itself is said to have a 192-bit bus and support for GDDR6X memory, but Nvidia was said to have planned to reduce the width to 160 bits for the RTX 4070, which would have had the advantage of only requiring 10 GB of memory capacity, the (production) cost of the card would have been lower as a result. In addition, GDDR6 memory was to be used at 18.0 GHz effectively.
Now Kopite7kimi reports that the plan has changed and the card (which reportedly has an internal designation of PG141-SKU331) will have more powerful specifications. Firstly, a fully active AD104 chip with 7680 shaders/FP32 units or CUDA cores is to be used. Nothing is known about the clock speeds yet, but the increase in the number of units should logically mean that the card will eventually have a bit more performance.
There is an update of RTX 4070.
12G 21Gbps GDDR6X
— kopite7kimi (@kopite7kimi) August 4, 2022
Nvidia has also changed its plans for memory. The GeForce RTX 4070, as per the PG141-SKU331 specification, is to use a full 192-bit bus and also have GDDR6X memory fitted. These will run at an effective clock speed of 21.0 GHz, giving the graphics card half the bandwidth (504 GB/s) of the GeForce RTX 3090 Ti. However, according to Kopite7kimi, the performance (at least in lower to medium resolutions?) could be almost the same or better. Its score in Time Spy Extreme is said to be exceeding 11,000 points. The significantly lower bandwidth could in fact be compensated by a significantly increased L2 cache (with 48 MB capacity), similar to AMD’s GPUs. The 192-bit bus also means that the graphics card will provide 12 GB of memory capacity.
However, power draw should be higher than it is now with the GeForce RTX 3070. Kopite7kimi mentioned a few days ago that the fully enabled configuration of the AD104 chip could have a “limit” of 400 W, and that was perhaps referring to the RTX 4070 (PG141-SKU331). But that number was probably just a Power Limit (or alternatively a power draw number for overclocked models), whereas the reference TDP of the whole card should be only 300 W, as this more moderate value was later mentioned by the leaker specifically among the parameters of the RTX 4070 model.
As I have mentioned before, there is an AD104 SKU with a 400W limit.
a full-fat AD104 with 7680FP32
21Gbps 12G GDDR6X
It can easily match RTX 3090 Ti.
— kopite7kimi (@kopite7kimi) August 1, 2022
However, even this is a pretty significant increase in power draw, as a mainstream card will now “chug” as much as what was the maximum for high-end graphics two years ago. As a comparison, before the GeForce RTX 3070 (220W) that is now offered, the GeForce RTX 2070 had only 175W TDP in 2018, and in the previous Pascal generation, the GeForce GTX 1070 consumed even less with just 150W TDP.
The GeForce RTX 4070 will probably come out a little later after the initial release of Lovelace-based graphics cards. Current rumors say that of those, Nvidia will only release the flagship GeForce RTX 4090 this year, while the remaining cards would probably come early next year, maybe sometime in January or March.
English translation and edit by Jozef Dudáš