Perspective: Eyeing the next wave in RISC computing

Despite what has been written (and rewritten) about the potential obsolescence of platforms based on RISC processors, they keep on going and going.

Indeed, RISC processors have never been more critical to driving innovation in computer design than they are today.

Design innovation isn't driven by the economics of large volume, the agendas of microprocessor companies or the plans of companies building commodity computers. Innovative designs come from companies with expertise in building entire systems from scratch--from chip to crossplane to code. These companies work closely with end users to anticipate new demands, and they utilize RISC-based technology to create the leading-edge systems needed to meet those demands.

In the 1980s, RISC (reduced instruction set computing), changed the rules of computing. The premise of RISC was that earlier CISC (complex instruction set computing) processors used only about 20 percent of the instructions they implemented. By comparison, RISC processors required far fewer transistors, making them cheaper to manufacture.

It is time for RISC designs to once again show the way to a new era in computing.

Further, because of their streamlined design, RISC processors could execute far more instructions every second. The result was an unmatchable price-performance advantage for RISC over older CISC-style designs, with RISC designs coming to dominate workstation, server and high-end embedded markets.

In the early '90s, RISC processors pioneered another advance, becoming the first processors to leap to 64 bits. That provided a huge increase in address space required for applications and data sets that were fast approaching the limits of 32-bit processing power.

About a decade after the first commercial RISC processors appeared, the manufacturers of CISC processors began to adapt the principles of RISC design to their own products. Today, CISC processors such as the Pentium 4 and Opteron translate their complex instructions into sequences of simple RISC-style instructions, then rely on a powerful "under the hood" RISC-style engine to execute the resulting "micro-ops." Similarly, about a decade after RISC processors had shown the way to 64-bit processing power, the first 64-bit x86 architecture processors appeared.

Now the computer industry faces a new set of challenges. Clock frequency gains have declined precipitously. Memory access times have risen from a few clock cycles to hundreds of cycles. Power budgets and associated thermal problems are rapidly escalating out of control. In short, it's time for RISC designs to once again show the way to a new era in computing.

At Sun, we've termed this brave new future of RISC processor design radical chip multithreading (CMT). Radical CMT processors differ from today's dual-core and symmetric multithreaded designs in much the same way a horseless carriage differs from a Formula One racer. That is to say, radical CMT designs aren't awkward adaptations of pre-existing technology to a new and unforeseen use. They are designs crafted from the ground up to employ available resources as efficiently as possible in pursuit of their throughput goal. We're not talking about executing two to four threads in parallel, but 30 or more threads simultaneously.

We're not talking about executing two to four threads in parallel, but 30 or more threads simultaneously.

More specifically, radical CMT designs will be based on new core designs, tailored to provide exactly the right level of per-thread performance needed for a targeted class of applications. This design then will be replicated eight or more times on a single die, along with appropriate mechanisms to ensure the cores operate together effectively. Each core also will be designed to gracefully switch between threads, so that when one thread stalls on a cache miss or branch misprediction, another can keep the core's execution resources productively occupied.

Although radical CMT, like previous waves of innovation pioneered by RISC processors, eventually will spread to commodity processors, once again this process is likely to take a decade or so. By the time commodity processors catch up, RISC processors will be ready to move on again, pioneering the next critical wave of innovation in processor design.

Technology is a process, not a result. For RISC, this means an endless migration toward the next big industry challenges that wait just over the horizon.

Biography
David Yen is an executive vice president at Sun Microsystems, where he specializes in processors and network products.

More Perspectives

not really.

Simultaneous multithreading made it to an x86 processor before it made it to a RISC (Tera doesn't count). Multiple cores are also coming to x86 within 1-2 years of RISC. The 10 year lead is nonsense.

Sun is clutching at straws. They would be better off pooling resources with AMD to make a better Opteron. The vast majority of people reading this article would be better served by low-cost, high performance commodity x86 processors from Intel or AMD than a Sun "RISC".

Compare: http://www.spec.org/cpu2000/results/res2004q4

A 4 opteron HP proliant has better CINT and CFP performance than a 4 ultrasparc (dual core) Sun Fire V490 at a price that is probably much better.

Indeed

I think you're right. You can easily compare prices on the DL585 vs. the v490 via their respective online ordering sites.

I just tried it, and a 4 CPU DL585 with 8GB RAM and similar storage is $11K cheaper than a 2 CPU (but 4 core) v490. The big kicker is the Sun box needs a $6K Sun Silver warrenty option (included in my $11K differential) to bring it up to the same 3 yr parts/labor you get with the HP box.

[publicstatic]

Too late for Sun

Sun never know its problem is EXECUTION. Now this company can only "sell" vision and spec. I don't care who is the inventor of CMT or similar concept. But I'm sure Sun WON'T MAKE MONEY from this vision because it usually fail to productize their concept. Look at Java! Who makes money from Java ? IBM or BEA or Sun ? Sun hardly to build a competitive app server years ago, now it's too late. Look at Java Enterprise System, how badly the actual release date match with the roadmap Sun announced 1 year ago, and how badly Sun brand the component products, so many confusing names and version.

Sun Processor Initiatives

Surprisingly many technical people are confused about such issues as RISC. This is due to a lack of appreciation of product management and marketplace issues. It is also due to their specialized knowledge and interests. But, so what?

It is important to understand that Sun has to remain active in processor design issues and manufacture, albeit, their chance to displace Intel or AMD in any important facet of desktop processors is unlikely.

With the advent of 'supercomputing' on the desktop we may see many developments that will surprise the public.

As a researcher in Artificial Intelligence, I am anticipating a considerable change in the marketplace with respect to desktop AI capability. At first, this development will take the form of specialized supercomputing in a difficult to use programming format.

WAN or wide area network computing has been a large disappointment (lol, with the exception of DOS attacks by hackers) so, I'd say forget the internet as a large scale platform for significant processing for the time being.

Muddled thinking aside, yes, there is a place for the processor innovations that Sun is spearheading.

What the public does not appreciate is that Sun is responding to a more specialized marketplace.

[wjc-qut]

RISC means RISK!

Well - we seem to have forgotten the original work and papers from IBM on the RISC concept from the 1970s and 1980s. The important point, made here also, is that all hardware based protection vitally needed to support system and middleware software for trusted and trustworthy computer systems, as carefully researched in the famous MULTICS project at MIT and elsewhere, is abandoned for speed of execution and minimistaion of structural overheads. Remember, the Intel x86 memory management architecture (from the iAPX-286 onwards) was based around these vital research results. What we need today is hardware enforcement of basic security structures, e.g. the memory address "bounds" and capability registers inherent in the segmented memory architecture model - meaning we can forget "buffer overflow", the protection rings concepts, meaning we can safely allow for diverse enterprises providing separate device drivers, etc.

All this is sacrificed in the RISC concept in the name of performance and manufacturing efficiency. While RISC has its place, e.g. in isolated, unconnected, trusted, environments and elsewhere, what is sorely needed today is a re-examination of the work of 25-30 years ago ( e.g. by Roger Schell, by the late and great Roger Needham of Microsoft who really put the segmented/capability memory architecture needs so well in the 1970s but was later totally ignored, even by Microsoft itself, and others).

By all means RISC - but NOT at the acceptance of unacceptable RISK in the modern, globally connected world!

Let's re-examine computer hardware architecture from the point of view of the new imperatives for trust in global interconnection. RISC was never, never, never originally intended to "play" in that domain - I wonder if it could?? Hmmm... Great challenge there...Over to SUN.... and IBM... and ?

Regards from Australia,
Bill Caelli

RISC/Sparc is useful in some areas, but trend is towards commodity computin

Sure, I use SUN/Sparc systems if I need ultra reliable systems for things like NFS servers or UNIX services (NIS, NTP, etc.). Servers that shouldn't ever go down.

But for large scale computing, you just can't beat 100s of low cost commodity processors connected with low cost commodity switching.

At least in my area (high performance technical computing), the RISC platforms have remained as the infrastructure systems, but the really massive computing has migrated to AMD/Linux commodity hardware. I don't see that changing anytime soon.

I can throw together inexpensive commodity systems running a commodity OS to do just about any job, with increasing reliability and ease of management. This makes buying a 'proprietary' system from SUN a harder sell these days, regardless of the wonderful advances this author advances.