By: anon.1 (abc.delete@this.def.com), April 26, 2017 10:15 pm
Room: Moderated Discussions
anon (spam.delete.delete@this.this.spam.com) on April 26, 2017 11:56 am wrote:
>
> See Denver, it's certainly worth trying.
>
Why? I'm genuinely curious. Isn't it transmeta V2 with native ISA support in hardware and a few other tweaks thrown in? Transmeta wasn't exactly a performance beast, so why is this time different? Denver hasn't addressed the fundamental reasons that allow OoO machines to outperform in-order machines esp in branchy integer code. You can create superblocks and hyperblocks all day and do all kinds of fancy scheduling within it, but you can't dynamically adjust to getting the schedule wrong, and you will be wrong some of the time. And hyperblock formation itself is speculative, so if you get it wrong, you have to throw it away entirely unlike OoO machines. Transmeta might've been worth a try. I haven't seen anything about denver that makes me think they addressed the fundamental issues. Or are we conflating activity with accomplishment? We have to do something, so why not do this? The fact that NVidia teams up with IBM to push their GPUs in HPC should tell you what their opinion of Denver is, wouldn't you say?
>
> See Denver, it's certainly worth trying.
>
Why? I'm genuinely curious. Isn't it transmeta V2 with native ISA support in hardware and a few other tweaks thrown in? Transmeta wasn't exactly a performance beast, so why is this time different? Denver hasn't addressed the fundamental reasons that allow OoO machines to outperform in-order machines esp in branchy integer code. You can create superblocks and hyperblocks all day and do all kinds of fancy scheduling within it, but you can't dynamically adjust to getting the schedule wrong, and you will be wrong some of the time. And hyperblock formation itself is speculative, so if you get it wrong, you have to throw it away entirely unlike OoO machines. Transmeta might've been worth a try. I haven't seen anything about denver that makes me think they addressed the fundamental issues. Or are we conflating activity with accomplishment? We have to do something, so why not do this? The fact that NVidia teams up with IBM to push their GPUs in HPC should tell you what their opinion of Denver is, wouldn't you say?