3 ms·
Key quote from the linked article: > the company had deployed thousands of servers that were sitting unused, and only ever had 1,600 concurrent users of the se
by ehsankia 8y ago
Key quote from the linked article:
> the company had deployed thousands of servers that were sitting unused, and only ever had 1,600 concurrent users of the service worldwide
They were burning through all of their money because they highly overestimated the audience. That's one of the main problems the cloud was made to solve.
- Const-me 8y agoThis particular cloud will only solve the problem if google will find a way to sell the time of these AMD GPUs to someone else. Traditional GPGPU customers, and machine learners, usually use CUDA instead of OpenCL / VulkanCompute / DirectCompute. At least from my position is looks that way. I’m a freelance developer and my clients are picking CUDA in ~75% cases. The rest is DirectCompute, if computing on client PCs and hardware compatibility is required.