There are now only days left until Nvidia unveils the next generation Geforce, where a paradigm shift with ray tracing is promised. As part of this, for the first time in twelve years, Nvidia deviates from the prefix “GTX” of the model names and replaces it with “RTX”.
With this, there is a lot of speculation about what the graphics circuits should be called for the generation. With the current convention, it would be “GT” followed by three digits, where the letter T is short for the architecture Turing. Now SweClockers can reveal that this will not be the case and that “TU” will be the correct prefix.
Specifications: Nvidia Geforce RTX 2080 Ti and RTX 2080
RTX 2080 Ti | GTX 1080 Ti | RTX 2080 | GTX 1080 | |
---|---|---|---|---|
Technical | 12 nm TSMC | 16 nm TSMC | 12 nm TSMC | 16 nm TSMC |
Circuit | TU102-300A | GP102-350 | TU104-400A | GP104-400 |
Circuit surface | 754 mm2 | 471 mm2 | ? | 314 mm2 |
Transistors | 18.6 billion | 12 billion | ? | 7.2 billion |
Architecture | Turing | Pascal | Turing | Pascal |
CUDA cores | 4 352 st. | 3 584 st. | 2 944 st. | 2 560 st. |
Texture units | 272 st. | 224 st. | 184 st. | 160 st. |
Raster units | 88 st. | 88 st. | 64 st. | 64 st. |
Clock frequency | ? | 1 480 MHz | ? | 1 607 MHz |
GPU Boost | ? | 1 582 MHz | ? | 1 733 MHz |
Computational power | ? | 11 334 GFLOPS | ? | 8 873 GFLOPS |
Memory bus | 352-bit | 352-bit | 256-bit | 256-bit |
Memory amount | 11 GB GDDR6 | 11 GB GDDR5X | 8 GB GDDR6 | 8 GB GDDR5X |
Memory frequency | 14 000 MHz(?) | 11 000 MHz | 14 000 MHz(?) | 10 000 MHz |
Memory bandwidth | 616 GB/s | 484 GB/s | 448 GB/s | 320 GB/s |
Power supply | 8+6-pin(?) | 8+6-pin | 8-pin(?) | 8-pin |
Outputs | 1 st. HDMI 2.0b | 1 st. DL-DVI | 1 st. HDMI 2.0b | 1 st. DL-DVI |
TDP | ~250 W | 250 W | ? | 180 W |
The new flagship is called Geforce RTX 2080 Ti and is equipped with TU102, which is a real mastodon with its circuit area of 754 mm2. This houses a total of 72 SM clusters in its full-featured version, but for the variant used for this model, TU102-300A, 4 are deactivated.
That leaves 68 SM clusters with a total of 4,352 CUDA cores and 272 texture units at as yet unknown clock frequencies. However, it is rumored to have a computing power of just over 13 TFLOPS, which would require a clock frequency around 1,500 MHz to reach. However, it is possible that the infamous value applies at base frequency and that the graphics card has a higher maximum turbo frequency.
The memory bus has also been ironed out, from 384 down to 352 bits – exactly the same as for the GP102-350 used in the Geforce GTX 1080 Ti. To this here the brand new memory technology GDDR6 with a capacity of 11 GB that ticks in an efficient clock frequency of at least 14,000 MHz, which would give a bandwidth of 616 GB / s.
This is followed by Geforce RTX 2080, which in turn is based on the graphics circuit TU104, which in its full design has 48 SM clusters. However, two are switched off in the variant used, TU104-400A, which leaves 46 left or 2,944 CUDA cores and 184 texture units.
Here, too, it is not known what clock frequencies it is about, but rumor has it that the graphics card has a computing power of 10.5 TFLOPS. To achieve this, a clock frequency of just under 1,800 MHz is required, something that is not completely impossible as it is only slightly above GTX 1080.
Like the GTX 1080, this one gets a memory bus of 256 bits. Together with the fresh GDDR6 in at least 14,000 MHz, however, this means a bandwidth of 448 GB / s, which is only 40 GB / s after Nvidia’s Geforce GTX 1080 Ti, which is today’s master of what can be considered more reasonable price ranges. The capacity of the memory is similar to its predecessor 8 GB.
On the connectivity front, the Turing-based graphics cards are taking a real step up towards Pascal. First of all, they each get HDMI 2.0b and three Displayport 1.4. The fifth connection is a USB Type-C that is available to support Virtuallink, a standard that is intended to power the VR headset of the future with only one cable.
The sales start for the duo will take place on Monday, August 20, while subsequent models such as the Geforce RTX 2070 will arrive later in October.
► Read all articles about the graphic architecture #Turing