The immediate future for Nvidia at the launch of its first duo of consumer graphics cards in the Geforce RTX 3000 series looked like a simple match. The company had no direct competition from AMD, expectations for the red team’s upcoming graphics circuit “Big Navi” were low, while performance results from the “Ampere” -based graphics card RTX 3080 made enthusiasts flock to the pre-booking queues.
A few weeks later, nowadays, the graphics card landscape looks different. While Nvidia certainly sold a lot of graphics cards based on the Geforce RTX 3080 and its later launched big brother RTX 3090, the majority of the graphics cards are back-listed due to possible manufacturing problems at the manufacturing partner Samsung. As a solution to this, Nvidia is said to have booked manufacturing with TSMC starting next year – but exactly which circuits will be manufactured there is still unclear.
Now information emerges from the leak Kopite7kimi, whose previous “Ampere” leaks turned out to be true, indicating that a new graphics circuit with the code name GA102-250 is on its way. The circuit will offer performance between RTX 3080 and RTX 3090 and is thus a candidate to become an RTX 3080 with the suffix Ti or Super. Like the RTX 3090, the new circuit is said to have a memory bus 384 bits wide, but the number of CUDA cores remains at 9,984.
RTX 3080 Ti | RTX 3090 | RTX 3080 | |
---|---|---|---|
Technical | 8 nm Samsung | 8 nm Samsung | 8 nm Samsung |
Circuit | GA102-250* | GA102-300 | GA102-200 |
Circuit surface | 628 mm²* | 628 mm² | 628 mm² |
Transistors | 28.3 billion* | 28.3 billion | 28.3 billion |
Architecture | Ampere | Ampere | Ampere |
CUDA cores | 9 984 st.* | 10 496 st. | 8 704 st. |
RT cores | 78 st.* | 82 st. | 68 st. |
Tensor cores | 312 st.* | 328 st. | 272 st. |
Texture units | 312 st.* | 328 st. | 272 st. |
Raster units | Unknown | 112 st. | 96 st. |
Clock frequency | 1 420 MHz* | 1 395 MHz | 1 440 MHz |
GPU Boost | 1 710 MHz* | 1 695 MHz | 1 710 MHz |
Calculating power (FP32) | 34 145 GFLOPS* | 35 581 GFLOPS | 29 768 GFLOPS |
Memory bus | 384-bit* | 384-bit | 320-bit |
Memory amount | 12/24 GB GDDR6X* | 24 GB GDDR6X | 10 GB GDDR6X |
Memory frequency | 19 000 MHz* | 19 500 MHz | 19 000 MHz |
Memory bandwidth | 912 GB/s* | 936 GB/s | 760 GB/s |
* The specifications are based on leaked information and may be considered speculative until they are officially unveiled.
At the far end of the memory bus are placed like RTX 3080 and RTX 3090 graphics memory of the type GDDR6X, and the width of 384 bits means a memory amount of either 12 or 24 GB. According to Wccftech, the bandwidth may be 912 GB / s – a stone’s throw from the RTX 3090 and its 936 GB / s. Other parameters, such as computing power in FP32 loads, are also expected to be closer to RTX 3090 than RTX 3080, where the new graphics processor is said to be powerful with 34 TFLOPs – to compare with 36 and 30 for the siblings respectively.
Where Kopite7kimi has certainly previously been right in its predictions about graphics cards in the “Ampere” series, Nvidia seems to be sitting in a much more precarious seat than originally planned. With pressure from AMD and “Big Navi”, the GA102-250 could very well be a graphics processor that is being pushed forward as yet another product with scant access to meet AMD in the top segment – if the circuit even sees the light of day in the near future.