After many rounds about the name for Nvidia’s next architecture, Volta looks to be a parenthesis in the company’s history and that it did not become Ampere, or that the use of this name will have to wait until later. Instead, the news agency Reuters sued in February’s report on naming after “the father of computer science” – Alan Turing.
Specifications: Nvidia Quadro RTX 8000 “Turing”
RTX 8000 panel | RTX 5000 panel | Tesla V100 | |
---|---|---|---|
Technical | 12nm TSMC(?) | 12nm TSMC(?) | 12nm TSMC |
Circuit | “Turing” | “Turing” | GV100 |
Circuit surface | 754 mm² | ? | 815 mm² |
Transistors | 18.6 billion | ? | 21 billion |
Architecture | Turing | ? | Time |
CUDA cores | 4 608 st. | 3 072 st. | 5 120 st. |
Tensor cores | 576 st. | ? | 640 st. |
Texture units | 288 st.(?) | ? | 320 st. |
Raster units | 96 st.(?) | ? | 128 st. |
Clock frequency | ? | ? | 1 370 MHz |
GPU Boost | ~1 736 MHz | ? | 1465 MHz |
Computing power (HP, FP16) | 32 000 GFLOPS(?) | ? | |
Calculating power (SP, FP32) | ~16 000 GFLOPS | ? | 14 899 GFLOPS |
Calculating power (DP, FP64) | ? | ? | 7 450 GFLOPS |
Memory bus | 384-bit | 256-bit | 4 096-bit |
Minnestyp | GDDR6 | GDDR6 | HBM2 |
Memory amount | 24 / 48 GB | 16 GB | 16 GB |
Memory frequency | 14 000 MHz | 14 000 MHz | 1 750 MHz |
Memory bandwidth | 672 GB/s | 448 GB/s | 900 GB/s |
TDP | ? | ? | 300 W |
It is in connection with Siggraph 2018 that architecture sees the light of day. The first to be presented is a real mastodon for a graphics circuit of 754 mm², which makes it one of the largest circuits ever manufactured. However, it is slightly less than 815 mm² for the Volta-based circuit GV100.
This includes no less than 18.6 billion transistors, even an impressive number that is surpassed by the GV100. Nvidia does not state which manufacturing technology was used, but the size of the circuit combined with the number of transistors points to TSMC’s 12-nanometer technology.
As expected, the architecture also introduces the GDDR6 graphics type, which succeeds both GDDR5X and the more than ten-year-old faithful servant GDDR5. The first graphics cards to be unveiled are part of the Quadro series and have the memory clocked at an effective frequency of 14,000 MHz, which for the top model with a 384-bit memory bus gives a bandwidth of 672 GB / s.
Nvidia also confirms that they use GDDR6 memory from Samsung, with 16 Gb (2 GB) per capsule and significantly higher speeds compared to GDDR5. The Quadro RTX 8000 is equipped with a full 48 GB of graphics memory, and the new memory modules enable Nvidia to increase the amount of memory even for the consumer-oriented cards.
In addition to the top model Quadro RTX 8000, RTX 6000 and RTX 5000 are also mentioned, which according to Anandtech get a 256-bit memory bus with GDDR6 at the same clock frequency for a bandwidth of 448 GB / s. This is especially interesting as the graphics circuit that makes up the RTX 5000 will probably also replace the GP104 (GTX 1080 / GTX 1070) for ordinary consumers.
In accordance with previous information, the company attaches great importance to ray tracing, where the prefix RTX indicates just this. Just like the GV100, the new Quadro models are equipped with Tensor cores and the new RT cores, which will be used for just ray tracing. According to Nvidia itself, the cards are six times faster when it comes to real-time rendering compared to Pascal-based models.
Another new feature that is introduced is support for Virtuallink, which is a cable standard for virtual reality developed by, among others, Nvidia together with several other players. In short, it is a Displayport 1.4 cable that has been modified for higher power supply and bandwidth and Nvidia’s Turing-based cards will be the first to come out with support for this.
Nvidia’s new Quadro card will be launched sometime during the fourth quarter of the year, with a recommended price tag of USD 10,000 for the top model Quadro RTX 8000, corresponding to approximately SEK 100,000. Whether the same launch period applies to the company’s consumer-oriented cards is still unknown.
Source: Anandtech
Read more about Turing!
► PNY reveals specifications and price for Geforce RTX 2080 and RTX 2080 Ti
► Nvidia Geforce GTX 2060 prestandatestas i 3DMark
► Nvidia’s Turing graphics card from MSI appears in the picture