Hacker News Reader

Comment by angersock | Hacker News Reader

What, no love for SPARC or MIPS?

Further, no love for Transmeta?

Poor show.

I would certainly agree that MIPS, at least as reference architecture, is an very important part of all modern architectures. The Patterson/Hennesey book [1] is absolutely require reading for many Computer Science/Electrical Engineering/Computer Engineering departments (even my state school required it) in undergrad, and most certainly required in graduate school when Computer Architecture turns to pipelining, cache structures and hierarchies, micro-architecture vs. macro-architecture, instruction set design, etc. Without MIPS, there isn't a baseline architecture to talk about in these courses as all other commercial architectures quickly begin to show their biases and optimization strategies.

[1] http://www.amazon.com/Computer-Architecture-Quantitative-App...

I dont think Transmeta ended up being that influential. Sparc/Mips as the birth of RISC should be there.

I was under the impression that they were munched by Intel and basically became the foundation for the microcode that those chips use now. Am I wrong?

> I was under the impression that they were munched by Intel

Nope, you must be thinking about something else. Transmeta started moving into IP in 2005, sued intel for patent infringement in 2006 (settled for a cool 250mil, Intel got a perpetual, non-exclusive license out of it), licensed tech to Nvidia in 2008 and died in 2009 (was acquired by Novafora in January, Novafora collapsed in July but not before Intellectual Vultures made off with the patent stash in Feb)

Now part of Intel's activity at the time (the PIII and Pentium M side which ended up with Core) may well have been inspired by Transmeta (as many allege), but at the same time Transmeta's activity best matches Intel's pursuit of the NetBurst (P4) shit-cake, which ran opposite to pretty much anything Transmeta looked for (unless Transmeta had patents on lofty promises they couldn't come close to deliver on, then year the P4 fits).

Ah, fantastic. Thanks for the info!

Fun fact: TX/RX Labs (Houston Hackerspace) has a rack of old Transmeta blade servers that a former member donated.

There's speculation that NVidia's Project Denver is an Efficeon reworked to target ARM code rather than x86. So there might be a live Transmeta descendant out in the real world again in a few months.

http://anandtech.com/show/7622/nvidia-tegra-k1/2

http://en.wikipedia.org/wiki/Project_Denver

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.

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.

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.

Yeah, our experience with the Itanic is that now that it's finally in port, it's pretty cool--unfortunately, nobody cares.

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.

I remember the Crusoe - it was ahead of its time, focusing on low heat and power consumption at a time when TDP meant nothing to the customers...

It's a field of heavy-hitters. The 4004, 8088, and Opteron 240 each ushered in a new age in microprocessors.

Do we have anything like that in the works right now? Everything seems to be fairly straightforward evolutions of the x86 or ARM--though maybe the innovation is happening over in FPGAs or business-not-tech venues?

Well, if that sort of thing is indeed in the works, I doubt anyone would tell us- competitive advantage and all that- but it has been ten years since the Opteron 240. Maybe it will be time for the next one soon.

It could even be one of many technologies recently on the market. "Influence" is a legacy; you can't see it on launch day. Maybe one day we will look back on hUMA [0] and say it was one of those Things? The PS4 processor in particular, as the first processor with hUMA.

Although, it does look to me like innovation is more gradual now just as a matter of fact- the business cycle is much shorter now. The 4004 was released in '71, the 8088 in '79, but today every chip vendor releases a new chip every six months, so changes will naturally be spread across releases and it will be harder to point and say "There, right there, that is where the world changed"

[0]: http://en.wikipedia.org/wiki/Uniform_memory_access#hUMA

I agree with hUMA, but don't forget that the game is now mostly about power efficiency. Also, GPUs.

Probably this if anything:

http://ootbcomp.com/

EDIT: They're releasing explanations as they submit patents, so not everything is public yet. I just submitted the most recent talk.

https://news.ycombinator.com/item?id=7027115

I imagine in another few years the Intel Core Duo will make the list, because of its status as the first major multicore CPU on the consumer market.

Realize that's still x86. I'm not sure we'll see that change, though. Another hugely influential Intel chip is behind that story: Itanium marks the point at which it became clear that the market's no longer receptive to completely new ISAs in general-purpose CPUs.

I think the market would have embraced Itanium, if it had been significantly faster.

Also, markets change. Today, 'normal' people don't care about CPU architectures. Who cares what his phone runs on, as long as it runs with low power use?

I would add the P6 architecture to that list, I feel that it was the redemption of many of the ideas in the i960 as well as its designer.

It was the one that finally lifted intel in peoples viewpoint. I think that a lot of people fail to realise just how much of intel's product lineup in the last 14 years can be traced to the P6 microarchitecture.

I guess you can never make a list that satisfies everybody. I would like to see something with DEC Alpha there (not only) for the integrated memory controller. Or Nehalem.

Reply on news.ycombinator.com