GeForce GTX Titan Z is the culmination of the work of Kepler architecture, Maxwell architecture is about to enter the Tegra line of mobile products, NVIDIA’s next generation GPU architecture has also been ready to come out, this is the epoch-making “Pascal”.
According to the road map, which will be launched in 2016, is still the rhythm of two generations.
A revolutionary new NVIDIA GPU architecture “Pascal”: two new technologies with no way out!
Tegra has to do with American superhero comics code, GeForce series is all history of the famous physicist, the next one is Blaise Pascal (Blaise Pascal) Born in June 19, 1623, died in 1662 On August 19, French theologian and religious philosopher, mathematician, physicist, chemist, musician, educator, meteorologist. Early studies of natural and applied sciences, the study of fluid mechanical calculators manufacturing and make an important contribution to the expansion of Torricelli’s work to clarify the concepts of pressure and vacuum, there is no effort to defend the scientific method, then mathematically contributed to insinuate geometry, probability theory two important new area of research.
“Pascal” There will be two groundbreaking cattle X place, one called ” NVLink “. This is NVIDIA, IBM CPU-GPU co-developed fast-track, the next generation of IBM Power processors will support.
Traditionally, between the x86 CPU, GPU connected via PCI-E bus, it is easy to share data when bottlenecks, high bandwidth IBM Power is particularly evident, NVLink bandwidth PCI-E 3.0 can reach 5-12 times, enough comparable CPU memory bandwidth, allowing CPU-GPU, GPU-GPU quickly communicate with each other.
However, this technique requires CPU, GPU support, so the predictable main application will be IBM Power processors, NVIDIA Tesla computing card, which is the main field of high performance computing , the consumer-grade GeForce graphics cards may see not short-lived fad to.
NVIDIA also made it clear that it would continue to support the PCI-E.
Another trick is ” 3D Memory “, which is the three-dimensional stacked memory to replace the current popular GDDR5.
3D Memory uses silicon vias (TSV) technology will be more stars DRAM stacked together, to achieve greater capacity, higher bandwidth, lower power consumption, especially in the current GDDR bandwidth and capacity of up to 2-4 times.
Specific capacity has not been determined, but it is possible to learn Intel, adding additional embedded cache.
NVIDIA also announced a new card design sample map, you can see the central core and a GPU integrated multiple satellites suspected cache, densely arranged around it seems that 3D Memory, and the volume and the entire card is very small, only the current about 1/3 .
Finally, say, in the roadmap released last year, Maxwell’s successor, called “Volta” (Volta), also has a stackable memory, now it’s gone. NVIDIA seems to change a name, Volta brother rest in peace.