Putting to bed months upon months of rumours, Nvidia has officially revealed the Nvidia Turing GPU architecture. Nvidia touts Turing as its single greatest leap in innovation since the advent of CUDA GPU cores in 2006, achieving the holy grail of accelerated real-time ray tracing thanks to its new RT Cores.
“Turing is NVIDIA’s most important innovation in computer graphics in more than a decade,” said Jensen Huang, founder and CEO of NVIDIA, speaking at the start of the annual SIGGRAPH conference. “Hybrid rendering will change the industry, opening up amazing possibilities that enhance our lives with more beautiful designs, richer entertainment and more interactive experiences. The arrival of real-time ray tracing is the Holy Grail of our industry.”
For SIGGRAPH, Nvidia was preoccupied with the benefits to the workstation and high-end industries, espousing the benefits of its RT Cores and Tensor Cores for AI inferencing and top-end visual effects. Thanks to these hybrid technologies, Turing can render physical simulations as 6x the speed of previous-gen Pascal.
“This is a significant moment in the history of computer graphics,” said Jon Peddie, CEO of analyst firm JPR. “NVIDIA is delivering real-time ray tracing five years before we had thought possible.”
But what does it all mean? Well, Real-Time Ray Tracing (RTX) is the big change for gamers. We’ve seen this demonstrated to excellent effect in a number of RTX demos released earlier this year, including the incredible Star Wars footage rendered in Unreal Engine 4. At the time, Nvidia said this was rendered on a $60,000 DGX station with GV100 GPUs. They lied. This was all, in fact, rendered on a single Quadro RTX GPU.
The Turing GPUs are equipped with RT Cores, mini processors which are dedicated to ray-tracing. They compute exactly how light and sound would travel within a 3D environment, producing photorealistic lighting effects. In comparison to Pascal, Turing accelerates these RTX operations by a speed of 25x. This is real-time ray tracing, and it’s happening right now and will be in consumer products.
The secondary benefit, for workstation users at least, is the Tensor Cores. These are focused on AI and deep learning and will be of great benefit for scientific and mathematics use. These are already in use in the Volta architecture used for the Titan V.
And now for the real juice - the hardware itself. For gaming graphics cards we’re going to have to wait another week, but Nvidia was prepared to announce Quadro RTX, its first ever ray-tracing GPU. A grand total of three Quadro RTX models are coming to market. This ranges from the $2300 Quadro RTX 4000, with 16GB GDDR6, all the way up to the frankly obscene $10,000 Quad RTX 8000, complete with 48GB GDDR6 (96GB With NVLink), 4608 GPU cores and 576 Tensor Cores.
The Quadro RTX family is built from two distinct Turing GPUs, one with 4608 CUDA Cores and the other with 3072 CUDA Cores. This is actually a drop in cores compared to Nvidia previous flagship products. However, the redesign GPU and cache architecture allow for a far greater level of performance from fewer cores. This is all achieved without a process shrink, although it means a monstrously-sized 754mm2 GPU, compared to 471mm2 for the GP102 used by the Nvidia Titan Xp.
Graphics Card Quadro GV100 Quadro RTX 5000 Quadro RTX 6000 Quadro RTX 8000 GPU Volta GPU Turing GPU Turing GPU Turing GPU GPU Process 12nm 12nm 12nm 12nm Die Size 815mm² TBD 754mm² TBD GPU Cores 5120 Cores 3072 Cores 4608 Cores 4608 Cores Tensor Cores 640 Cores 384 Cores 576 Cores 576 Cores Base Clock1450 MHz 1132 MHz TBD 1.75 GHz TBD Single Precision 15 TFLOPs TBD 16 TFLOPs TBD Ray Tracing Spec N/A 6 GigaRays/Sec 10 GigaRays/Sec 10 GigaRays/Sec VRAM 32 GB HBM2 16 GB GDDR6 24 GB GDDR6 48 GB GDDR6 NVLINK VRAM N/A 32 GB With NVLINK 48 GB With NVLINK 96 GB With NVLINK Memory Bus 4096-bit 256-bit 384-bit 384-bit Memory Bandwidth 870 GB/s 448 GB/s 672 GB/s 672 GB/s TDP 250W ~160W ~200W ~225W Price $9,000 US $2,300 US $6,300 US $10,000 US Launch Date 08-Dec-2017 Q4 2018 Q4 2018 Q4 2018