NVIDIA has announced new details about its future GPU architecture roadmap at the GPU Technology Conference in San Jose, California. The company’s next Pascal GPU architecture will debut next year while the Volta GPU architecture will arrive in the year 2018.
Let’s talk about the Pascal architecture first. The main highlights of Pascal will be Mixed Precision, 3D Memory, and NVLink. Mixed-Precision computing will allow Pascal architecture based GPUs to be able to compute at 16-bit floating point accuracy at twice the rate of 32-bit floating point accuracy. 3D memory will allow Pascal architecture GPUs up to three times the bandwidth and three times the frame buffer capacity of current GPUs based on the Maxwell architecture. NVLink will help Pascal GPUs to move data between the CPU and GPU much faster than what is possible currently with the PCI-Express connector. NVIDIA is claiming 12 times faster data movement.
Another area where Pascal GPUs will be a huge leap forward will be memory. Think the 12GB memory on the GTX Titan X is impressive? Pascal will allow maximum memory of up to 32GB, that’s 2.7 times higher than the Titan X. Overall, NVIDIA claims we could be looking at around a 10x improvement in performance over Maxwell when using multiple GPUs. With single GPUs, you can expect around 5x jump from Maxwell.
The successor to Pascal, Volta, will debut only by the year 2018. Volta will double the maximum memory capacity to 64GB, twice higher than Pascal.