Slashdot Mirror
AMD Says Power Efficiency Still Key
Larsonist writes to tell us that even though AMD's new architecture wont be released until mid-2007 they are still letting people in on what some of the new features will be. From the article: "While clock speeds have not been revealed, each of the four cores will integrate 64 KB L1 Cache and 512 KB L2 cache. The native quad-core architecture will also include a 2 MB shared L3 cache, which may increase in capacity over time. The processor will have a total of four Hypertransport links - up from three today - that provide a total bandwidth to outside devices of 5.2 GB/s. AMD is also thinking about integrating support for FB-DIMMs 'when appropriate.'"
The article is very light on details but the one picture implies power control at the core level. For example if core-1 is running a 100% workload and core-2 has a 50% workload, core-3 and core-4 can be halted resulting in a power load of only 45% the total 4-core max load.
MS licensing is for currently for physical CPU's not for cores. Right now a dual xenon (two CPU's) counts as two CPU's in MS licensing terms but a dual-core (two CPU's within a single die or processor socket) is one CPU under MS licensing terms.
In other words, MS counts sockets, not cores.
Microsoft will not charge per core anytime soon. They might have a dominating position on the desktop, but when it comes to servers, they're fighting tooth and nail to get a position of respectability, let alone dominance.
So, let's look at the two markets seperately.
Desktop, the users are likely to not care too much, provided the "Per core" cost is low enough. When we start seeing 4/8/16 core CPUs, a $10 per core fee will add up quick, but most home users will be using OEM copies and won't see that cost. Most business will have site licenses and won't care. But, some home users will, and some businesses will care, and they'll seriously consider alternatives (Maybe, maybe not switching). Microsoft would much rather "lose" money by not fleecing people, than have them even CONSIDER switching, so management's going to ditch the idea for desktops and workstations.
Server market... they need any advantage they can get. They main competition is Solaris, BSD, and Linux. Linux and BSD are *free*, and Solaris has a bunch of good features which are pretty much Solaris only, even still. Charging per core would be suicide in this market, too.
So, what market would charging per core be a good idea for Microsoft? None. Say what you will about their software writing abilities, but nobody should doubt their marketing prowess.
AMD now provides a TSC (cycle counter) that doesn't vary in speed when the core speed changes. This greatly helps with timekeeping.
As for race conditions: that is pretty well taken care of already. SGI has Linux on a 2048-way system now.
Well, at least if you are in a data center.
There are two huge concerns in a typical data center enviornment: Heat and Power. These two areas are key because of the density of servers today. We're cramming so much processing and storage into 48U that people 10 years ago couldnt have even dreamed of even existing. Delivering enough power to run 48 servers can be difficult if each server is pulling 4 amps each (thats 192 amps). Considering most circuits are 20 or 30amps, thats alot of circuits to fit in one rack.
This was always the biggest reason why Dell servers were not as popular with the companies that I have worked with. Quite simply, AMD was kicking Intel's ass with heat and power. I heard many people say they'd start ordering Dell servers by the pallet if they sold AMD processors (looks like they finally listened).
Most SMP code is tested on CPUs of equal clock speeds
You're kidding right? I can't imagine any software which depends on the timing of coperative CPUs.. MPI and general divide-and-qonquer work-clusters could care less about the performance level of peer threads/co-processes. Hell process interrupts due to pre-emptive multi-tasking is enough to guarantee lack of symmetry.
Now perhaps you're referring to scheduling problems in the kernel.. I'm sure that AMD would be generous enough to provide kernel patches as are necessary.
-Michael
I work for a very LARGE bank. I guarentee you we have more boxes running MS Server 2003 than all others combined.
We have some HPUX, IBM, and SUN sprinkled in there but several of the vendor apps I've worked with lately has dropped all support
for SunOS and is now requiring Windows Server for their apps.
We use to run Novel for our file servers but that was dropped for MS Active Directory.
The next leading server OS we use is probably zOS. No Linux yet but I know the powers-that-be are looking at it.
(BTW, I'm no MS fanboy, I'm just making a point.)
I'm sure there are many industries in which MS does not have the majority of the server market but
large financial groups are not among them.
$7.95/mo, 200 GB disk, 2TBxfer, MySQL, PHP, RoR.
SMP means the processors are similar (i.e. can run identical binary code). They do not need to run in lock-step synchonization to be SMP. Indeed, it is currently possible to halt a single processor in a dual processor system (two sockets) so a similar case already exists in current SMP systems.
I was giving an example to counter the GPs claim that MS doesn't have respectability in the server market.
Not only do they have respectability, I see more and more vendors requiring them and most custom apps we write is on MS (desktop and server), although a lot of our web stuff is IBM (websphere, etc).
I work in telecommunications; Aspect, Avaya, Cisco (not cisco routers/switches) have ALL migrated from other OSs to MS in the past 5 years for their server and desktop apps.
Yes, Novel, Sun, HP, and IBM are huge in the server business. Except for IBM, they will steadily lose share to MS over time.
One of the main reasons this hasn't happened faster (like in the desktop arena) is because there are a LOT of home-grown apps that companies have built over the past 15 years that still require a particular OS to use.
At the bank I work for, we have a home-grown desktop app that is "critical". It was originally written for OS2. Around 1997 the team was forced to port it to MS windows NT.
The conversion was brutal and it took them forever to get all the bugs worked out.
Cut to 9 years later. WERE STILL USING THAT APP! (and no one is happy about it.)
They've tried 4 times in the past 5 years to either have it rewritten or replaced with a vendor product, but every project has failed.
It's so bloated and patched it's a nightmare to maintain, but even tougher to replace.
Now apply that story to the thousands of custom server apps in thousands of companies around the world that only run on SunOS, HPUX, Novel, etc.
The number of custom web apps alone that have been written on Sun servers could keep that company going another 20 years.
We run a lot of MS server OS at work because we've been busy migrating stuff over the past 5 years. (a LOT of stuff.)
So, my point is, MS Server OSs are not only respectable, but many large companies are actively moving towards them because they see it as streamlining. (Large companies like to be consistent and stick to "industry standards", even if it doesn't make good business sense.)
MS knows it's going to take time; it's not like replacing a desktop OS.
$7.95/mo, 200 GB disk, 2TBxfer, MySQL, PHP, RoR.
Nah. Here's what Intel's 45nm page says: This important milestone demonstrates that we are on track for 2007 to manufacture chips on 300mm wafers using the new 45nm (P1266) process, in accordance with Moore's Law.
Intel shipped a few 65nm processors in 2005, but didn't really get started until 2006, and full conversion might not have happened yet, although all the important plants should have migrated by now.
Even if true, AMD has yet to ship a SINGLE 65nm processor. By this measure alone, I'd say the claim that they are a year behind is quite adequete. But by the speed at which AMD is producing fab plants, I'd argue that they are or soon will be an entire chip generation behind.
AMD have been behind on the process node, but that's not the only issue when it comes to making chips, although it is the most major. SS + SOI are other technologies that AMD is far ahead of Intel on, and they help reduce power significantly - hence AMD's low power 90nm processors compared to Intel's 90nm, and even Intel's 65nm P4s, and AMD aren't doing too badly in terms of performance/Watt right now either.
Traditionally Intel has won the absolute performance title. As you said yourself it is the biggest factor when it comes to performance/Watt statistics. If been following any of the Core 2 Duo reviews, Intel is now dominating in that arena too.
Symmetric doesn't really have any deeper implications. It just implies the processors are similar. There's no underlying implications about synchronization. When writing code for SMP machines, not only is it not possible to accidentally depend on the processors being synchronized, it's impossible to explicitly depend on synchronization.
Spin-locks or not, cores running at different clockspeeds aren't going to expose any more race conditions than regular usage. Even on a current SMP system, the processors will almost always be asymetrically loaded, so threads of an app won't really see the two processors running at the same speed anyway.
A deep unwavering belief is a sure sign you're missing something...
SMP means the processors are similar (i.e. can run identical binary code). They do not need to run in lock-step synchonization to be SMP.
The Symmetric in SMP refers to the fact that each process can run the same tasks.In an assymmetric setup, there may be a processor dedicated to the kernel or other tasks.
It most definitely has nothing to do with speed.
Now perhaps you're referring to scheduling problems in the kernel.. I'm sure that AMD would be generous enough to provide kernel patches as are necessary.
I find that the two processors in my dual-core Athlon X2 run at slightly different speeds (according to AMD, this is expected). That in fact did cause the Linux kernel some problems, since it was trying to balance handling the interrupts between the two. The problem happened when the timer interrupt bounced between the two, as within an hour or two of startup their tick counts became significantly separated. This made the system clock start running forward at a rapid rate. A Linux patch fixed this issue. So I can definitively say that Linux does run on SMP cores at different speeds.
I'm not sure how Windows will do it, but they'll probably figure it out if they haven't yet. The real challenge is a new scheduling algorithm for variable CPU capabilities (although we do have that to some extent with frequency scaling on single CPUs).
Customers are wanting better power efficiency. Due to the rising costs of energy a lot of data centres are now charging based on energy consumption and heat output (needs more AC -> needs more energy). It's becoming a real problem, especially for some data centres who for varying reasons cannot increase their electricity supplies. That's why google, microsoft and others have started building data centres near a hydro-electric dam in the states, cheap and plentiful electricity.
I am NaN
Actually the new low power X2-3800+ (ADD) now has the performance/Watt crown.
It has idle power of 8 watts and full load of 25 watts.
They measured the performance of various benchmarks and also the energy required to perform the benchmark to get a figure for performance/Watt and the
X2-3800+ came out miles ahead of the Conroe (or any other chip). http://www.lostcircuits.com/cpu/low_e/6.shtml