'Radical Redesign' Urged for Future Computers

CommanderFrank

Cat Can't Scratch It
Joined
May 9, 2000
Messages
75,399
A professor from the University of Maryland is proposing radical changes to the present day computer architecture. Dr. Vishkin has designed a new architecture abstraction to accommodate today’s multi core processors in a more efficient manner.

Such an architecture, Vishkin states, would require changes in hardware design. For the approach to operate, the chips would require a high-bandwidth, low-latency network between the processors and memory
 
it is exactly the way systems work with hyper threading and modern gpus work. Nothing new here except some smutz got a shit load of money for claiming the re-invent the wheel which had already been done.
 
Yes, good ideas, none of which are new though... followed by the fact he has six patents $$$ on related ideas. Of course he's going to push it.
 
Ummm... why is THIS guy getting credit for something they have been working towards for years? (and already done in some cases)
 
I'm more curious about what Intel has in the works for after Haswell and Rockwell.
 
I'm with everyone else in thinking isn't this how current multicore cpus and stuff work. Each core can process a different instruction. I no longer get that ACM magizine otherwise i would read his article to see if we are all missing something.
 
Newsflash: getting together high bandwidth, low latency, multiple processor cores, and memory is hard.

It doesn't sound like programming for parallel instructions is going to be any easier than what you can do now with programming for multiple threads or processes (which execute in parallel given OS support and multiple CPU cores). Maybe it would be a little more lightweight than a thread, but it sounds like this guy doesn't know the crazy stuff a modern processor will do to your serial instruction stream anyway.
 
The PCworld article leaves a lot to be desired; but what I think he's advocating is improving the communication rate between cores so that a single thread could run on several at the same time.

I don't see how this would differ than replacing the current setup of an intel quadcore CPU with 3 integer, 1 FPU, 1 MMX, and 1 SSE pipeline in each of 4 cores with a single core with 12, 3, 3, and 3 pipelines and enough decode units/etc to support 8 way hyperthreading. The problem with this approach is that very few applications can keep all the pipelines they're using full at the same time which is why Intel and Amd went to multi-core architectures instead of just adding more pipelines to a single core. Baring a major breakthrough in compiler/language design this isn't going to change; and while a skilled asm programmer can do so the cost of doing so is prohibitive outside of mature high performance computing applications where the vast majority of the computing cycles are spent in a small segment of code.
 
Instruction level parallelism has already been exploited by pipelining. It'd be nice to have VLIW (packing multiple instructions into an instruction word, and still pipelining) like Itanium which a compiler can optimize for, but hey, the pipeline alone does away with the 1-instruction-only idea, even without Hyper-threading (or any other SMT).

Thread and data level parallelism already can be leveraged by us programmers, assuming OS support (been there for quite a while, so moot point).
Such an architecture, Vishkin states, would require changes in hardware design. For the approach to operate, the chips would require a high-bandwidth, low-latency network between the processors and memory. The hardware would have a single processor core to control all the other cores. If the code is serial, it can be executed on that core. If there are additional instructions, the central processor can dole out additional instructions to the other cores.
Gasp. :p He turned an ASMP operating system into ASMP hardware. So instead of proposing less-scalable-than-SMP operating systems (relatively inexpensive to rewrite) we are told that less-scalable-than-SMP hardware (expensive to re-design and re-build). Using ASMP may be great for very specialized systems, but it's gonna suck some major ass addressing lots of software over time by effectively removing the scheduler from the OS to the single "controller" core. :rolleyes:
 
Gasp. He turned an ASMP operating system into ASMP hardware. So instead of proposing less-scalable-than-SMP operating systems (relatively inexpensive to rewrite) we are told that less-scalable-than-SMP hardware (expensive to re-design and re-build) is the way to go. Using ASMP may be great for very specialized systems, but it's gonna suck some major ass addressing lots of software over time by effectively removing the scheduler from the OS to the single "controller" core.
 
Redesigning the hardware is not so much needed as is getting all software makers to start programming and writing their software to properly use multi-core processors. They are all woefully behind in this aspect.
 
You could dream up any number of instructions as long as the input for one is not the output for the another

The problem with most programs is that most functions are interdependent, and usually sequential in nature because the user's activities occur in a sequence. The only place this works well is when the programs are running independently and not attempting to utilize the same hardware resources simultaneously, or in output situations where no returned values are required, as is often the case with graphics and audio acceleration. It's more efficient for a sequence of instructions to be simultaneously distributed to multiple CPU's for processing and the results reordered back into memory.
 
Newsflash: getting together high bandwidth, low latency, multiple processor cores, and memory is hard.

It doesn't sound like programming for parallel instructions is going to be any easier than what you can do now with programming for multiple threads or processes (which execute in parallel given OS support and multiple CPU cores). Maybe it would be a little more lightweight than a thread, but it sounds like this guy doesn't know the crazy stuff a modern processor will do to your serial instruction stream anyway.

Huh.. what?

You would still have to use threads even if no thread interfered with data from other threads.

Eliminating locks in threads is possible in some instances and can speed things up exponentially compared to using locks to make sure data doesn't get messed up.

The person who wrote the article needs to get educated about systems that have been around for 10+ years.
 
Wasn't this all attempted a few years ago with the failed BTX form factor?
 
Am I the only one that _doesn't_ think this is a load of crock? Seems pretty reasonable.

Current architecture is set up for serial instruction execution (thread), and all parallel work must be done by either creating separate discrete threads, or fancy runtime based profiling to determine if instruction level parallelism can occur.

He is suggesting, why not do away with the hacks to the current system, and design parallel from the ground up. Do the parallelization at the software level (both code and compiler) and then have a much simpler "processor pool" that can execute an arbitrary amount of non serial-dependent instructions.

No guarantee that it would be better, or worth the effort, but it doesn't seem to be that far out. ie. to suggest the idea that by adapting a primarily serial design to add in support for parellization (where now virtually all computers are multi-core) might be more efficient than just designing for parallel from the ground up, at every step of the food chain.

*Shrugs*
 
A big hint: anything you read in general purpose media is wrong on specifics. It is likely completely wrong as well.

An out of date paper on the subject.

The article points out that this is returning to a topic that was quite in 1990 but was ignored after the mid 1990s. It notes how the cell processor (of PS3 fame) appears closest in hardware design of commercial processors to what they are doing. Performance was compared against an Opteron, a chip not exactly designed without parallelism in mind.

Project website. And yes, it does appear to be a "radical redesign" of computers.
 
Hard to really say anything about it without looking more into it, but just reading the article and hearing what the guy reportedly said about the architecture, I have 2 major issues:

1. 1 core to dole out work to the others? OK... sure, that's real scalable.
2. "For the approach to operate, the chips would require a high-bandwidth, low-latency network between the processors and memory."

L O L. I believe ALL chips would love to have this, but the reason why you often don't have both (or either in extreme cases) is because it is expensive...
 
Won't happen. CPUs 30 years from now will still have to be able to work exactly like the 80386, just in case some company still depends on some specialized software that only runs on Windows 3.1.
 
Do the parallelization at the software level (both code and compiler) and then have a much simpler "processor pool" that can execute an arbitrary amount of non serial-dependent instructions.
That would be very, very tedious. Programs tend to be serial by nature owing to the process of handling data. You have at the very basic, input -> calculation -> result. Everyone knows the "Hello world" program. Every complex program is simply an outgrowth of the same process, just with a lot more math in the middle. You cannot eliminate the serial nature of programs entirely. You can farm out some of the calculations, but that's about it.

adapting a primarily serial design to add in support for parellization (where now virtually all computers are multi-core) might be more efficient than just designing for parallel from the ground up, at every step of the food chain.
I'd agree on that. Actually it's already done to a large degree already in the PC. Games are a great example. You have work farmed out to the GPU, the Audio Processor, etc, and taken off the CPU whereas in the older x86 systems and prior you had only the CPU to do all the work.

*Shrugs*[/QUOTE]

I believe ALL chips would love to have this, but the reason why you often don't have both (or either in extreme cases) is because it is expensive...

This hits the proverbial nail on the head. Redesigning the entire board architecture would increase cost, and really, how much redesign can you practically do? Unless you embed the RAM directly in the CPU die and do away with the memory bus you're always going to have a bottleneck, and on-chip RAM is putting too many eggs in one basket, causing a higher chance for defects and failures, and limiting the flexibility of the system. Larger on-chip caches DO have a dramatic effect on speed, but even then there are limits.

The problem with redesigning the mainboard architecture is that, right now, it just works to darn well. There are other, more practical and cost effective ways to speed up the PC without reinventing the wheel.
 
Back
Top