Slashdot Mirror
AMD Says Power Efficiency Still Key
Larsonist writes to tell us that even though AMD's new architecture wont be released until mid-2007 they are still letting people in on what some of the new features will be. From the article: "While clock speeds have not been revealed, each of the four cores will integrate 64 KB L1 Cache and 512 KB L2 cache. The native quad-core architecture will also include a 2 MB shared L3 cache, which may increase in capacity over time. The processor will have a total of four Hypertransport links - up from three today - that provide a total bandwidth to outside devices of 5.2 GB/s. AMD is also thinking about integrating support for FB-DIMMs 'when appropriate.'"
But will it run Vista? (j/k)
Is Vista going to support 4 cores, or like XP Pro and 2k, limit it to 2 "cpus" so they can charge more for the server version?
Tequila: It's not just for breakfast anymore!
But were the 2048 way systems running at different clock speeds? Or rather were the processors in each node running at different clock speeds? Last time I saw someone on linux-kernel mismatch processors it brought all sorts of interesting issues out.
I'm sceptical that this technique will be very useful. (Of course, AMD is full of smart people and I'm just some net.moron.) I don't think it will be very common for the load on a 4-core processor to be somewhere the middle like 1.5. It's either going to be mostly idle (load close to 0) so you might as well power down the whole chip, or going full blast with the load as high as I think will give me the most throughput. For example, when compiling (and that's when I wish I had more cores) I'm gonna "make -j n" and my load is going to be about n, and that number is going to be chosen to be one more than the number of cores I have (or something like that). If I have a 4-core machine, do you think I'm going to make -j 2? No way.
I can't think of many situations where I would have one core running at 100% and another at 50% and the others idle, for any significant length of time. I can imagine a desktop user clicking on something and maybe for a few milliseconds that load is somewhere around that, but then the work gets done and you're idling again. Or the user asked it to do something "hard" so all cores are near 100% (except maybe while waiting for I/O) for a "long" time.
Am I wrong? What kinds of things does your computer work on, which are a little parallelizable but not very much?
As copyright owner of this comment, I authorize everyone to defeat any technological measure which limits access to it.
How does this tally with a previous story about multi-core architectures being ideal for realtime ray-tracing in games? Is anyone working on a Ray-Tracing evuivalent of OpenGL?
**TODO** Steal someone elses sig.
Intel have publically stated that they will not be shipping 45nm chips until 2008.
Intel shipped a few 65nm processors in 2005, but didn't really get started until 2006, and full conversion might not have happened yet, although all the important plants should have migrated by now.
AMD have been behind on the process node, but that's not the only issue when it comes to making chips, although it is the most major. SS + SOI are other technologies that AMD is far ahead of Intel on, and they help reduce power significantly - hence AMD's low power 90nm processors compared to Intel's 90nm, and even Intel's 65nm P4s, and AMD aren't doing too badly in terms of performance/Watt right now either.
Thank you, Intel marketing department!
I'm afraid that I'm being a thermodynamics fascist, but you really can't efficiently heat your apartment with CPUs. Typically, for every BTU of heat delivered to you by electrical power, 2 more BTUs go up the power station smokestack or are lost in transmission. In contrast, for every BTU delivered from a modern gas furnace, less than 0.1 BTU goes up the chimney.
Okay then. So that means that you can't just shift most of the power to one CPU to make it much more efficient at single threaded tasks, but also that it still should be quite beneficial to run multiple cores at different speeds.
Assuming that each core is like an individual CPU, odds are, one core is better than the other. That would mean that for both cores to run at the same speed, the better one would have to be getting more voltage than it requires, or be running slower than it could be(those two things are basically equivalent). So the optimal solution would be to link the state of the cores, but for each state, set each CPU to the maximum it can go at that particular voltage.
Instead of having idle cores drag the speeds down, maybe active cores should push the speeds up.
Exactly. Intel calls this EPI throttling in one of their recent papers.