![]() That pretty much means it has to do speculative execution. The Thunder very likely does speculative execution it is a high-end CPU with high core count and high parallelism, intended for big network servers (they are trying to compete with Xeon server-class chips) and also being used in supercomputers. Similar things happen in computers with disks (whose performance depends crucially on the workload, and on workload parallelism): By measuring the performance of its own IOs, a process can make inferences about what IO other processes are doing. The traditional joke about it is that the best way to measure the level of international tension is to watch how many pizzas are delivered late at night to the Pentagon. The same problem of "information leakage due to observing the performance effects of operations that you are not allowed to see" happens in many other realms. But speculative execution is here to stay given the physical limits on clock frequencies, parallelism (of which speculative execution is one variant) is necessary. In effect, chip designers will have to learn new skills. The problem is that it has effects that can be measured by performance, and that's where all these problems are coming from. I understand why the chip designers did it that way (because otherwise the permission checking engine becomes the bottleneck), and in theory it looked like this wouldn't hurt, because speculative execution by design can't have visible effect. The problem is speculative execution without doing permission checking. The problem isn't speculative execution per se. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |