The Leading Source for Global News and Information Covering the Ecosystem of High Productivity Computing
From the Editor | Main Blog Index
January 08, 2009
The hardware and software challenges of multicore/manycore CPUs have been flogged in this publication for a number of years. The assumption was that geek ingenuity would eventually power through the roadblocks. The memory wall problem would yield to innovative hardware architectures, and new software development approaches would make multithreaded computing practical enough for widespread use. But what if that doesn't happen?
There's a good article in the January/February 2009 issue of Technology Review that outlines multicore computing challenges and talks about some of the software strategies being pursued by Intel, Microsoft and others in the industry. But the most interesting part of the article is toward the end, where the author allows for the possibility that the whole multicore paradigm may just fall apart:
So what's the downside if multicore computing fails? What is the likely impact on our culture if we take a technical zig that should have been a zag and suddenly aren't capable of using all 64 processor cores in our future notebook computers?
For a positive spin on this outcome, the author quotes Apple Computer co-founder Steve Wozniak, who apparently believes the end of Moore's Law-driven microprocessor evolution would be a good thing:
"I can't wait!" says Steve Wozniak, the inventor of the Apple II. "The repeal of Moore's Law would create a renaissance for software development," he claims. "Only then will we finally be able to create software that will run on a stable and enduring platform."
Of course, the other way to create a stable platform is to build scalability into the software model so that the number of cores is transparent to the application. The idea is that jumping from 8 to 64 cores automatically gives an application better performance, without recoding or even recompilation. That's the thrust behind the work Intel, Microsoft and university researchers are doing today.
Some industry luminaries, like Professor David May at Bristol University, thinks replicating cores using legacy architectures is the real problem, given that conventional CPUs like the x86 were never designed for parallel processing. He elaborated his position in October in an Electronics Weekly article on the pitfalls of multicore programming:
Current attempts to use multi-cores in the mainstream computing world, like the efforts made by Intel and Microsoft with a bunch of US universities, may be doomed. "I think they (Intel and Microsoft) are trying to solve a different problem," said May, "they're taking all the PC applications and putting them on multi-cores. That's a very different problem and, in my view, they won't be very successful. Taking sequential programmes and trying to make them run in parallel is virtually impossible."
May is also the CTO of XMOS Semiconductor, a company that has developed a multicore architecture that uses "software defined silicon" to combine some of the best attributes of ASICs and FPGAs. The resulting processor is aimed at the consumer electronics market.
Perhaps along the same lines is Creative Technology's just-announced Zii processor, which also claims to use software defined silicon in its newly minted 10 gigaflops chip. Like the XMOS silicon, Zii is targeted for the consumer space, although the Web site video hyperventilates about building a petaflop supercomputer with a mere six racks of Zii processors. Maybe if they were IBM, they'd actually attempt it.
In any case, for most kinds of client-side computing, the x86 architecture may truly be a dead end. Since the Internet became the center of the computing universe, PCs have been morphing from general-purpose computing appliances to thin clients. This will continue as more and more computing is moved into the cloud. As clients get ever thinner, the main computing load is data transcoding, which generally can be accomplished with greater efficiency using more specialized silicon like GPUs, FPGAs, DSPs and maybe these new-fangled software defined silicon gadgets. In that sense, PCs are becoming more like handheld devices.
Where would that leave server-side computing, especially HPC? For throughput and capacity computing, CPU-based architectures still offer a reasonably-natural computing architecture. But for many HPC applications, and for capability supercomputing in particular, the inherently parallel architectures of GPUs, Cell processors and FPGAs offer a better fit (although a CPU companion is still needed at this point). The high level of interest with GPGPUs, Cell processors and FPGAs is one indication that supercomputing might be turning away from conventional CPUs.
Economics will dictate that mainstream HPC will continue to rely on the same processor architectures used in consumer electronics. But one day, those chips may be something other than x86.
Posted by Michael Feldman - January 8 @ 4:49PM
(Digg, Technorati, more)
There are 1 discussion items posted.
Not cores, but memory!
Submitted by dmpase on 01/09/2009 - 4:58AM
As a long time system developer for HPC, it strikes me that it isn't the number of cores, per se, that causes the problems. It is and always has been the access to memory.
Consider that the ratio of memory throughput to CPU throughput has gone down year after year for decades. Many programs run successfully on clusters with thousands of cores. Why not a desktop system with a few tens of cores? Because the aggregate memory bandwidth of the cluster is many times that of a desktop system.
With every generation memory and CPUs both get faster, but CPUs get faster at a faster rate than memory. It isn't too much parallelism and too little parallel work that hurts us. It's our inability to feed the raging beast.
For those old enough to remember the FPS T-Series, it died a dismal failure for very similar reasons. It had FPUs capable of extremely fast processing, but far too little bandwidth to get data into the FPUs.
It is not the number of cores, nor is it that cores are too fast. It is that the balance is shifting inexorably towards the processing end away from memory and current software can't take advantage of it.
Post #1
Platform HPC Workgroup Manager
Platform HPC Workgroup Manager integrates all the cluster productivity tools you need to deploy, run and manage your HPC environment.
Michael Feldman is the editor of HPCwire.
More Michael Feldman
We think by watchesuk
Re: IBM and HPC by truly64
HPC = servers but a lot more by lawries
Lena by Nastyanna
Lena by Nastyanna
Multi core deployment becomes a memory game by truly64
Re: Venture Capital Drought? Not So Much. by Ron Van Holst
Re: AMD Confirms 12-Core Opteron Production by Nastyanna
Re: Cray Corrals Big Defense Deal by Nastyanna
Re: Podcast: Cray Awarded Defense Deal; SGI Makes Storage Buy; IBM Invents New Algorithm by Nastyanna
Painful Truth by jeffrey.mcallister
SGI = graphics + HPC by johnbarr
HPC = servers but a lot more by truly64
Oracle SPARC != Fujitsu SPARC by Alan M. Feldstein
Sun & HPC != Oracle & HPC by Merblich
a third vendor for lossless low latency 10GbE fabric by lee.fisher@hp.com
Response to GAH by KevinButerbaugh
Response to KevinButerbaugh by GAH
Response to KevinButerbaugh by GAH
Response to GAH by KevinButerbaugh
Response to bdrupp by KevinButerbaugh
Climate Crisis and Exaflops by bdrupp
Climate Crisis and Exaflops by John Hules
Climate Crisis and Exaflops by GAH
Climate Crisis by KevinButerbaugh
IBM "Brain Simulation" article is not properly presented. by Merritt
563 out of 1206 by vvolkov
Little Iron by gadunk
At least it's not "cloud" by KevinButerbaugh
Native QPI Interface? by commike
Mmmmmm by hellcats
New transistorized IC chip scales. by symmecon
Itanium at IDF by Alan M. Feldstein
Communication time by jnapper
"The financial meltdown and computing" by donpellegrino
Human Models by mdgabriel
High-End SPARC Chip for Scientific Applications by Alan M. Feldstein
RapidMind by Mr LolO
Rapidmind by dminor
Longer run times by JohnWest
re: Algo trading Angst by jshore
Results of Testing by in_the_crease
The ACM Turing Award goes to the creator of the modern personal computer; and Voltaire announces a mid-range InfiniBand switch and new technology that accelerates distributed applications. We recap those stories and more in our weekly wrapup.
Read More...
The prospects for virtual SMP technology got another boost last month when Florida State University announced it had installed a new HPC system from 3Leaf Systems. The servers are being housed at the university's HPC facility and will be used across a range of scientific disciplines.
Read More...
For the first time in 62 years, the four-man Olympics bobsled team from the US captured the gold medal, setting a course world record in the process. The winning bobsled had some state-of-the-art engineering behind it, including CFD software from Exa Corporation. As it turned out, that software may have proved to be the margin of difference in the race.
Read More...
Mar 11 | Linux Magazine | CUDA may be the rage, but OpenCL is a standard that has some features you may need. Read more...
Mar 09 | Free Software Magazine | Data-driven computing will need open software. Read more...
Mar 09 | Bio-IT World | Tahoe Informatics founder eyes GPUs, CUDA software. Read more...
Mar 08 | Sporting Life | Formula One engineers differ on benefits of CFD. Read more...
Mar 08 | InfoWorld | AMD offers up 48-core server prize. Read more...
Jan 12 | | In-depth look at vSMP Foundation server virtualization technology, technical implementation, use cases and capabilities. The technical whitepaper provides an architectural overview and details on the three vSMP Foundation products: vSMP Foundation for SMP, vSMP Foundation for Cluster and vSMP Foundation for Cloud.
Jan 18 | | This white paper discusses Gore’s copper cable assemblies, and how they continue to exceed the standards for providing reliable, cost-effective solutions for high-performance computer applications.
Join this online panel discussion for live Q&A with leading industry experts, analysts, and end-users to discuss the latest innovations, best practices, barriers to implementation, and measurable benefits of server virtualization with a particular focus on today's real world solutions.
Learn about scalable fault-tolerant architectures and examples of energy efficient and scalable supercomputing clusters using dual QDR InfiniBand to combine capacity computing with network failover capabilities with the help of programming languages such as MPI and a robust Linux cluster management package.
LIVE@SCO9: The IBM team discusses new innovations in hardware, software and services that help clients better understand their workloads and get insight from their R&D efforts. Technology demonstrations include the soon-to-be-released Power7 HPC processor, the DCS990 system with 2.4 petabytes of storage, the xCAT management tool, secure HPC cloud computing and more. Winners of two HPCwire Readers' and Editors’ Choice Awards! Take the IBM virtual tour at SC09 or more information go online to: http://www-03.ibm.com/systems/deepcomputing/sc09.html