Slashdot Mirror
Intel and AMD's 2005 Plans Revealed
Takemedown writes "There's a good article on CTZ that talks about Intel and AMD's plans. Intel, continuing on their 18-month chipset refresh rate, will introduce their Glenwood and Lakeport chipsets for the Smithfield dual core desktop microprocessor in 2005. The chipsets will support SATA II, Matrix RAID and a higher system bus speed for the new Pentium 4 name holder.
As far as Intel's dual core strategies are concerned, they will most likely bring their dual core additions by the very end of Q2 or Q3 this year, so for those waiting for these next generation chips are better off with a due upgrade. Secondly, if you are hoping for a noticeable performance gain in regular computing tasks are in for a disappointment. Dual core microprocessors are for those who like to do multitasking or work on multithreaded applications. For example, if you are gaming and burning a DVD at the same time, dual core chips will come in handy and will definitely give a smooth computing experience."
The real bottleneck for gaming these days is hard drive access. If you are burning a CD while you are playing a game, there is a good chance that the game will need to load something like textures while you are burning the CD (presumably from an ISO on your hard drive). On the other hand, with a 52X CD-R burning a full CD takes less than 3 minutes, so it won't kill your game. Unless you have two hard drives, in which case the above is irrelevant.
Burning a DVD is IO-bound given all the traffic on the PCI bus from the harddrive and to the DVD. Burning a DVD is not CPU-bound, so it doesn't seem like a dual core CPU would actually help that situation.
The reality is most of the server market is their Xeon line and the dual-core Xeons are currently planned for 2006 and maybe even later.
It's exactly the same as SMP, except for two things:
1) Far less 'glue' circuitry is required on the motherboard. This allows cheaper multi-processor systems.
2) Potentially, communication between the processors could be faster.
Mostly, though, the advantage will be social -- if a large fraction of systems have multiple processors, as they will soon, then more and more applications will be written to take advantage of them.
Thad Beier
I love Mondays. On a Monday, anything is possible.
Dual core shares a memory controller, whereas dual processors have seperate memory controllers. AMD's Athlon 64 and Opterons have memory controllers on die, and were originally designed to be dual core. What this means is now two cores on die with one memory controller, communicating through a crossbar (think SGI) architechture. On a side not, imagine where AMD would be if they scrapped 64-bit from the start and released the Athlon 64/Opteron as a dual core from the get go.
AMD is using a technology patented by IBM called SOI (Silicon on Insulator)... IBM is very unwilling to allow Intel to use this technology to solve their heat problems....
Tom's Hardware has some good information about thermal loss. Notice that an idle AMD Winchester (SOI Athlon 64) loses only 3.2 watts, while the more recent P4 chips are losing > 34 at idle.
This number changes at load to 30 watts for the Winchester and 100+ watts for the P4.
Looking back and comparing it to a P2-450 I once owned... the Winchester numbers are close.... and that machine had no fan (just a very large heatsink).
I'm not sure you could have a fully-loaded Winchester without at least some type of active cooling... but certainly the CFM required across a good heatsink would allow you for an almost silent fan.
Memory as in on-die cache... yes, RAM... no:
1) Core based processors have more internal/embedded synchronization built in, especially related to on chip caching. SMP relies more heavily on the O/S for maintaining concurrency.
2) Connection between processors is shorter and theoretically faster. The big gain here is that the MB components for SMP are all integrated on the CPU, so everything is simplified and compressed.
3) Cache in SMP is separate to each processor, core-processors share the cache between the processors. SMP must maintain cache concurrency... this the basis of threading headaches and this takes process cycles to do so. However, sharing the cache in a core processor is often a problem (Intel) if the cache isn't big enough. AMD currently does this better.
4) SMP means higher license costs for multiple processors, core based processors are considered one processing unit (at least to MS).
That last one tends to be the most important to alot of people.
Via has been offering very - if not the most - stable chipset drivers for Linux for ages. There were many occasions in the past when I specifically chose via chipset brand boards with amd cpus because good experience in the past with Linux. And I never had any bad experience. That doesn't mean I don't have and/or use other chipsets/boards/cpus with Linux on them, I do - but mostly @ work. It's just I never see much sane reason in arguing like I-went-intel-because-of-better-linux-drivers.
I am putting myself to the fullest possible use, which is all I can think that any conscious entity can ever hope to do.
Silent PC Review has a good review of the Zalman Reserator 1, the only product I know of that even comes close to qualifying. It's nearly silent, more or less easy to install (if you built your computer, you can put this together, but I wouldn't recommend it for my mom), and can cool even the hottest processors and videocards simultaneously. It is not cheap, however, at around $240 shipped. I just installed mine, and it's the quietest, coolest-looking computer cooling part I've bought in years. Unfortunately, there really isn't another silent/easy/cheap option available. Seems like it's a "pick any two of three" situation.
This can be achieved on a commodity single-core processor using pure software techniques. The technique is known as Error-Detection through Duplicated Instructions (EDDI), and is implemented as a compilation step between assmbly code generation and object file generation. Stanford has done a bunch of work on this at their Center for Reliable Computing. I don't have any links readily available, but I'm sure that if you Google on EDDI and the ARGOS project you'll find some good info.
Note that IIRC experiments at Stanford showed that when using EDDI on a modern super-scalar processor the EDDI instructions can take advantage of unused portions of the pipeline, resulting in a significant reduction in overhead. You might still experience a slight performance hit, but on the other hand you don't need to add a whole new processor or core.
Not right, this could almost be said about hyper threading but even then it isn't really close. Dual core is two complete cpus put on the same chip. This should also alow them to share cache.
There has recently been a patch the the Linux Kernel about zoning cpus. This helps process migration across cpus. basicly if a cpu is overloaded you have to move some process off of it to another cpu. using cpu zones the migration code can try to pick hyperthreaded or multi core cpus to migrate to first because of shared cache. moving to a diffrent smp cpu has a lot of overhead.
Paying taxes to buy civilization is like paying a hooker to buy love.
Here's the thing with dual AMDs right now: wait for nVidia nForce4-based systems to come out. The current ones use the AMD chipset, which forces you to use expensive registered memory. Once the nForce4-based dual proc boards are available, you'll be able to use regular memory, yay!
The big thing I've been waiting for was PCIe and nForce4 - PCIe is here, nForce4 is here (though very limited), and nForce4 dual proc should be along in just a few months, at most (I hope).
Stay the course!
I'd wait for the nForce4 dual proc socket 940 boards, then get a cheap (ie: obsolete) pair of Opterons.
If you want a machine that uses non-registered memory, is dual proc AMD, then you can forget AGP - it's gonna be PCIe-only with nForce4, from what I've read.
Depending on what you're doing, you may want to get a modern Opteron - there's an updated core coming along soon that has SSE3.
InfoWorld had a nice story about the Power5 multi-core CPU (You'll have to download the report) coming out this year. It may outperform the coming dual core AMD chip, both in raw performance and in lower power consumption.
AMD has a write up on their upcoming dual core processor and what it means to performance. Somewhere I believe there are some published numbers for how an AMD dual core CPU running 5 steps below it's single core counterpart can still outperform dual single core processors. (i.e., a 1.4 GHz dual core CPU will outperform a 2.4GHz dual processor machine)
Meanwhile, Intel's dual core demo was doubted doubted when presented at the same time as the above referenced AMD demo. Also, Intel's dual core will not perform significantly better than a dual processor system, or so the analysis of the two processors stated. (I really need to bookmark these things when I read them! Hopefully someone else will provide that reference.)
The cesspool just got a check and balance.
"AMD chipset, which forces you to use expensive registered memory. Once the nForce4-based dual proc boards are available, you'll be able to use regular memory, yay!"
Not true. The memory controller in Athlon 64 / Operon is on the die of the processor, not a part of the chipset.
Opteron systems will still require registered memory. If a dual-core Athlon 64 is released, it will probably be compatible with NForce3 as well as NForce4.
NForce4 is just NForce3 250GB with a new firewall, SATA with TCQ, and PCI Express.
Memory as in on-die cache... yes, RAM... no:
1) Core based processors have more internal/embedded synchronization built in, especially related to on chip caching. SMP relies more heavily on the O/S for maintaining concurrency.
Not sure what this means. Cache snooping is hardware based and is not controlled manipulated by the OS unless the CPUs allow it to be turned off/on. Non-cache coherent systems are a bear and I don't know of anyone who would build such a system when the CPUs in question (in this case, AMD Opteron or Intel P4) have it built in.
2) Connection between processors is shorter and theoretically faster. The big gain here is that the MB components for SMP are all integrated on the CPU, so everything is simplified and compressed.
Because the logic is built on-die does not necessarily make it more simple. It can be basically the same logic, just moved from the external chipset onto the die.
3) Cache in SMP is separate to each processor, core-processors share the cache between the processors. SMP must maintain cache concurrency... this the basis of threading headaches and this takes process cycles to do so. However, sharing the cache in a core processor is often a problem (Intel) if the cache isn't big enough. AMD currently does this better.
Except in the cases we are talking about (AMD Opterons and Intel P4s), neither have a shared cache at any level. There may be later versions of the CPUs that have it, but we won't be seeing them this year (or maybe next either).
4) SMP means higher license costs for multiple processors, core based processors are considered one processing unit (at least to MS).
Yes, this is purely artificial. There's functionally no difference between a dual core and a dual CPU (SMP) machine.
The main thing about multi-core vs. multi-chip SMP (both are SMP, btw, just different implementations of it) is computational density. With multi-chip, you have to have a socket and traces for each CPU on the motherboard. This takes area and having a bunch of CPUs makes you have a big motherboard. Two CPUs = 2X the area on the motherboard, for example. A dual core CPU takes the same space as a single core one. With multi-core, you cut the number of sockets and traces down, thus making the thing cheaper to make.
Have you ever seen a dual socket microATX board? A dual-core microATX machine is very doable because it's really not much different from the single socket board.
Single Opteron motherboard
Dual Opteron motherboard
Quad Opteron motherboard
With dual-core CPUs, that single board turns into a dual, that dual turns into a quad and that quad turns into an octo? as if by magic. Basically, this gives you 2X the computational density by using dual-core CPUs. Similar calculations can be made using quad core CPUs to show 4X the computational density.
Although most /. readers probably won't care, dual core CPU's are already on the market in the form of the UltraSPARC IV CPU from Sun Microsystems. Sun also happen to be sporting the most ambitious multi-core project going in the form of Niagara, which although initially an 8-core system has apparently been seen running Solaris 9 with 32 independent CPU cores.
In addition to this, the POWER 5 CPU is also available with multiple cores, fully supporting Linux.
Also of note is that the Opteron dual-core CPU's from AMD are apparently going to be pin-compatible with the current Opteron processors (by current,I mean, the latest socket 939 (I think) systems, not the original Opteron 2xx or whatever).
This is really of most use for the data center right now, but as more applications wrap their heads around paralelizing themselves, multi-core CPU's on the desktop will become more popular.
That said, developers really have no excuses for not having blazing fast "dual-core aware" apps... a multi-processor system purchased today provides about as much performance as a dual core system... so it's not like a wild new technology where application developers have to wait for SDK's or test hardware. Multiple cores, HyperThreading CPU's or multiple physical processors are all just additional CPU's from the operating systems perspective, and are developed for using the same tried and true thread libraries (pthreads, etc).
Multi-thread those apps people! There are so many instances, especially when writing GUI apps, where an extra thread or two thrown in the right direction can really improve the user experience.
Of course, a big problem is just how developers learn to program. Everyone learns their "Hello World!", then goes from there... but this is all very linear in approach. Finding good programmers who can think of an application in terms of what many parallel threads should (or shouldn't) be doing isn't easy... but I digress.
bash-3.00$ uname -a
SunOS panda 5.10 Generic sun4u sparc SUNW,Ultra-2
Linux drivers?
AMD provided full driver support for their chipsets, and even optimized BLAS libraries (these are scientific computing libraries).
To get this from Intel you have to buy their compiler suite.
How about next time you study up on some facts before you blindly repeat the Intel party line?
Maurice W. Hilarius Voice: (778) 347-9907
err, I beg to differ. I don't know what chipset you've had trouble with, but even the laptop chipset of my athlon XP-M gives gentoo no trouble. Almost all chipsets are now supported by linux, at least most of the VIA I've tried, and the kernel lists a big bunch, which are not for Intel processors. I've had linux on 3 computers with all a different motherboard, but all AMD CPUs: Duron 1.3 GHz, Athlon XP 1800+ and Athlon XP-M 2400+. No trouble from linux with those chips.
---- I am certain of only one thing : I know nothing else.