3 ms·
You can build an ASIC with fast external memory, it adds to the cost but then you can handle larger models similar to a GPU. Software support is an issue but fo
by hedgehog 10y ago
You can build an ASIC with fast external memory, it adds to the cost but then you can handle larger models similar to a GPU. Software support is an issue but for deep learning applications there's no reason in principle you couldn't add support to TensorFlow etc for new hardware to make it simple for application developers to adopt. Movidius has announced that they're doing this and it's likely that other ML chip vendors will do the same.
- emcq 10y agoExternal memory kills your power budget. Efficient processors, whether they be GPUs, cpus, or asics all gain efficiencies by keeping data local to computation. Having your chip focused on off chip data is difficult to optimize power for if you don't have compute bound problems.