The network has been browsing for a long time about the possible specifications and performance of NVIDIA GeForce RTX 4000 graphics cards based on the Ada architecture. There are still a few months left until the premiere of the three most powerful cards: GeForce RTX 4090, GeForce RTX 4080 and GeForce RTX 4070, while the latest reports show that the manufacturer is still adjusting individual elements of the specifications. Currently, the last of these cards seems surprising.
Unofficial specifications for NVIDIA GeForce RTX 4000 graphics cards appeared on the web.The manufacturer is still working on the final parameters, so the latest information may change before the premiere.
NVIDIA GeForce RTX 4000 with very high power limit for desktop cards and significantly reduced for mobile versions
According to information posted by kopite7kimi on Twitter, the NVIDIA GeForce RTX 4070 card may have worse specifications than previously thought. The AD104-275 chip will offer 7168 active CUDA FP32 cores, and the card will use 10 GB of 18 GB GDDR6 memory on a 160-bit memory bus. Until now, it was believed that the card in question would receive 12 GB of memory on the 192-bit bus. The TGP power is expected to be around 300W.
NVIDIA GeForce RTX 4090 | NVIDIA GeForce RTX 4080 | NVIDIA GeForce RTX 4070 | |
general engineering | Ada | Ada | Ada |
drawing layout | M 102 – 300 | M 103 – 300 | M 104 – 275 |
lithography | TSMC N4 | TSMC N4 | TSMC N4 |
SM . blocks | 128 | 80 | 56 |
SP units | 16384 | 10240 | 7168 |
Effective VRAM clock | 21 Gbps | 18 Gbps | 18 Gbps |
VRAM | 24 GB GDDR6X | 16 GB GDDR6 | 10 GB GDDR6 |
Memory bus | 384 bit | 256 bit | 160 bit |
TBP . coefficient | ~ 450 watts | ~ 420 W | ~ 300 watts |
NVIDIA prepares triple-fan cooling for GeForce RTX 4000 graphics cards with AD102 core
The NVIDIA GeForce RTX 4080 in its current form uses the AD103-300 chip, which is based, among other things, on 10240 CUDA cores and 16 GB of GDDR6 memory. Interestingly enough, kopite7kimi notes that the card will abandon GDDR6X memory cubes in favor of slower GDDR6 (18Gbps). The change may be dictated by a significant increase in the L2 cache, thanks to which there will be no need to give GDDR6X bones as fast as possible. The best GeForce RTX 4090, in turn, is offering a 16384 CUDA processor (previously about 16,128 cores) and 24 GB of GDDR6X memory on a 384-bit bus.
And TBP, 450/420? / 300 watts.
– kopite7kimi (@kopite7kimi) June 23, 2022
Source: Twitter @kopite7kimi
“Prone to fits of apathy. Introvert. Award-winning internet evangelist. Extreme beer expert.”