"Mentioned when appropriate..."
No one brought it up though. We were discussing ETAs and then the OP threw out a reasonable speculation about power efficiency, and then you come in with your pitch about throughput, which you must admit came from another realm of discussion that was entirely unsolicited.
I understand you feel strongly about these features but regime changes do not happen overnight, and in that meantime you are going to have to be the community charlatan citing fringe cases and synth benches where these technologies perform exceptionally and somehow ignoring the cases where they do not. It's getting on everyone's nerves because there is actually a lot more to haswell than the two things you won't shut up about. People would go easier on you if you would talk about other things, but you don't. This is all you have to talk about, so irrespective of the validity of your posts, it *feels* like we have to sift through your spam to have a balanced topic-driven discussion with everyone else. I don't dislike you or anything like that. It's just the impression I get. And when you say "Deal with it," in the tone that you did, it really drags down the level of discourse and your repute.
If you want to admire intel for something, take note of the intermediate evolution of the core family that did so much for throughput while still confined to preexisting and shortsighted instruction set extensions. The fast radix divider, super shuffle engine, physical register file, and the uop cache all achieve better than linear improvements in throughput per joule and are universally applicable, particularly the uop cache which sees ~85% hit rate no matter your application. That is engineering to me; being a badass under incredible constraints.
To me, AVX and TSX are great technologies, but the instruction set and underlying hardware are brand new and being built and implemented together from the ground up, which is a bit less challenging and far less impressive to me than getting a better-than-linear improvement from a retrofit such as the radix-16 divider.
But yeah, just do a search. You mention TSX in almost every thread, sometimes even in a lazy unclear manner:
Mainstream Haswell chips will have 64 compute cores, made available through the AVX2 instruction set extension. It offers the best performance/Watt improvement imaginable.
And more cores is worthless without more multi-threaded software. Haswell also adds TSX technology to make that happen.
I mean come on. It reads like you are doing a crappy SEO job for someone. I don't mean to beat up on you, but the socket 1150 haswell will have a max of 4 physical cores with 2-way hyperthreading. Wider, smarter instructions can theoretically increase throughput to something equivalent to 64 simple cores, but that post right there remains completely unsubstantiated speculation and what's worse is you make the reader feel like they are being harassed by a door to door evangelist and there is no thread we can go in to get away from you. Stop promising something you have never seen. Stop it.
When you are old like me, it's more fun to just watch things transpire. Stop trying to make everyone think and feel the way you do. Personally I hope Haswell is a disgrace because I know intel will survive it and you may not and you'll learn more from that than from a bunch of people barking at each other who are more interested in being correct than being corrected.
Facebook
Twitter