NVIDIA on Multi-Chip Module (MCM) GPU: Multiple GPUs Under One Package
NVIDIA is spearheading a new project that will pack multiple GPUs under one package. This is the GPU leader's answer to the rapid increase in the development of GPU technologies nowadays.
Based a recent story posted by Tech Report, the Multi-Chip Module GPU (MCM-GPU) proposal that would comprise several GPU modules integrated as a single package was formulated by the researchers and engineers from Arizona State University, the University of Texas at Austin, the Barcelona Supercomputing Center, and NVIDIA.
According to the report, the proposal originated with the idea recognizing NVIDIA's soon-to-be dilemma of bringing in more power on one chipset based on today's fabrication technology. Previously, NVIDIA was able to improve GPU performance by increasing multiprocessor (SM) count. However, the problem also lie somewhere in that department as it has become a problem to bring in more transistors into a single die.
Like what NVIDIA's CEO Jen-Hsun Huang said during GTC 2017, "The part that is really shocking is this is rectile limits. Rectile limits basically means that is at the limit of photo lithography meaning you can't make a chip any bigger," Huang added as posted by WCCFTech.
Take for instance NVIDIA's V100 GPU which pushed the limit on its 12-nm process. Project cost and larger dies even added to the initial problem. Initial take on the solution was to put multiple GPUs on the same PCB like with the Tesla K10 and K80. However, it rises to other problems for which NVIDIA has yet to solve.
So instead of taking this earlier approach, a new and better approach is seen and that is to take advantage of allowing multiple NVIDIA GPU modules (GPMs) onto one package. Smaller and cheaper, this makes it a perfect candidate to enter the manufacturing stage.
Such would be the case of the next GPU after NVIDIA's next generation Volta architecture. Volta and Pascal both have FP32 and FP64 cores while the former houses a dedicated Tensor cores for INT8 operations to accelerate neural networking and deep learning performance. Considering that Volta is bigger and better than the already powerful Pascal, what would NVIDIA do next after Volta? The answer of these engineers and researchers are simple - Multi-Chip Module GPU (MCM-GPU).
© 2017 iSports Times All rights reserved. Do not reproduce without permission.