3 ms·

Not at all. The Transmeta is a really fascinating CPU architecture based around "code morphing", modern CPUs use something at a bit of a lower level: micro-op

Not at all. The Transmeta is a really fascinating CPU architecture based around "code morphing", modern CPUs use something at a bit of a lower level: micro-op translation.

seanmcdirmid 13y ago
Intel was doing micro-op translation way before 2005...it was something they introduced with the Pentium Pro. Their patent problems with transmeta was probably coincidental infringement (in that one can patent the obvious). Transmeta's code morphing was quite different: they were going to VLIW instead of RISCy Pentium Pro operands. It never worked very well, but for the same reason Itanium/Merced failed: it is just really hard to transparently schedule VLIW instructions efficiently without programmer intervention.
- InclinedPlane 13y ago
  I thought that was what I said, maybe I wasn't very clear. Anyway, it's a really interesting range of problems, as is what's worked and what hasn't. Personally I think code morphing still has a lot of promise, but mostly in software level "emulation"/trans-coding. Something interesting about Itanium is that eventually it's actually become pretty good. But it's now in a market niche that it's not going to be able to break out of. The big problem with IA64, in my opinion, was that they neglected IA32 performance with the first gen. They made the classic error of prematurely committing to a new design before it was proven, which contrasts greatly to the AMD64 example where they did everything possible to ease the transition and maintain, or even improve, performance with x86 only code.
  - angersock 13y ago
    Yeah, our experience with the Itanic is that now that it's finally in port, it's pretty cool--unfortunately, nobody cares.
  - seanmcdirmid 13y ago
    I believe Intel eventually fixed Itanium with more support for dynamic instruction scheduling which sort of goes against the very concept of VLIW. SIMD (via GPUs) rather than MIMD then took the performance crown for HPC tasks and hasn't looked back. VLIW still works very well in DSP-style applications, where static scheduling is quite feasible.

Not at all. The Transmeta is a really fascinating CPU architecture based around "code morphing", modern CPUs use something at a bit of a lower level: micro-op translation.

seanmcdirmid 13y ago

Intel was doing micro-op translation way before 2005...it was something they introduced with the Pentium Pro. Their patent problems with transmeta was probably coincidental infringement (in that one can patent the obvious). Transmeta's code morphing was quite different: they were going to VLIW instead of RISCy Pentium Pro operands. It never worked very well, but for the same reason Itanium/Merced failed: it is just really hard to transparently schedule VLIW instructions efficiently without programmer intervention.

InclinedPlane 13y ago
I thought that was what I said, maybe I wasn't very clear. Anyway, it's a really interesting range of problems, as is what's worked and what hasn't. Personally I think code morphing still has a lot of promise, but mostly in software level "emulation"/trans-coding. Something interesting about Itanium is that eventually it's actually become pretty good. But it's now in a market niche that it's not going to be able to break out of. The big problem with IA64, in my opinion, was that they neglected IA32 performance with the first gen. They made the classic error of prematurely committing to a new design before it was proven, which contrasts greatly to the AMD64 example where they did everything possible to ease the transition and maintain, or even improve, performance with x86 only code.
- angersock 13y ago
  Yeah, our experience with the Itanic is that now that it's finally in port, it's pretty cool--unfortunately, nobody cares.
- seanmcdirmid 13y ago
  I believe Intel eventually fixed Itanium with more support for dynamic instruction scheduling which sort of goes against the very concept of VLIW. SIMD (via GPUs) rather than MIMD then took the performance crown for HPC tasks and hasn't looked back. VLIW still works very well in DSP-style applications, where static scheduling is quite feasible.