Multiprocessor Gaming - Advantages and Disadvantages

[Rike]

I've taken it as a almost as a matter of faith that a multiprocessor rig holds no real advantages over single CPU systems. That's been challenged and I have to admit I have no data one way or the other. Clearly, cost of a multiproc system is a disadvantage but what else is going on here? Is there a performance advantage in gaming. If so, how much? Those with experience and evidence please post your links.

 

[Hajime]

Almost no games are multi-threaded.

Hence, they use one of the two procs. The parts available for a multiproc rig are far more expensive then equivalent enthusiast parts as a rule of thumb - check out the prices on Reg-ECC ram.

Now, with that setup, you are using slower and more expensive parts for a game that only uses one of the two procs. If anything, I would say there is a performance -disadvantage- to going dual proc for gaming. Try finding FX-55 Opteron equivalents - I dare you. And yet for the price of two top-end Opterons, you can afford a FX-55 easily.

 

[Rike]

Originally posted by: Hajime
Almost no games are multi-threaded.

Hence, they use one of the two procs. The parts available for a multiproc rig are far more expensive then equivalent enthusiast parts as a rule of thumb - check out the prices on Reg-ECC ram.

Now, with that setup, you are using slower and more expensive parts for a game that only uses one of the two procs. If anything, I would say there is a performance -disadvantage- to going dual proc for gaming. Try finding FX-55 Opteron equivalents - I dare you. And yet for the price of two top-end Opterons, you can afford a FX-55 easily.

You've just summed up my own beliefs. But I'm looking for proof. Does anyone know of any benchmark comparisons that back up these beliefs with regard to gaming?

 

[IHYLN]

why do you need proof when the guy said that there are no games that are multithreaded?

 

[Pariah]

There is no advantage for multi-CPU systems running games except you can brag about paying about twice the price for the same performance. If you use your computer mainly as a workstation that does significant amounts of rendering, encoding, calculating, etc, then dual CPU systems can provide a significant speed boost, again at a significant cost. The other advantage of SMP systems is that they can do multiple CPU intensive activities at once with little affect on the other. So while a 3GHz CPU may choke while trying to play Doom3 and encode video at the same time, if you had a 3GHz SMP system, both applications would run as if they were on 2 seperate 3GHz systems. Is an SMP system for you? Well think about it in these terms. What speed CPU are you looking to upgrade to? Then look at whether you can afford 2 Opterons or Xeons running at that speed. If you can't (and most people can't), then it isn't for you. SMP is not for those looking for a bargain. You cannot put together a faster SMP system than a single CPU system for a home user for less money, buying new anyway.

There is no clock speed equivalent to the FX-55 yet, but let's not pretend that in the highend the FX is any bargain compared to a single CPU Opteron system, because it isn't.

A socket 939 FX-53 which is basically the twin of the Opteron 150 with a slight package change is almost $200 more expensive than the Opteron 150 and is about $30 more than the SMP capable Opteron 250. The OEM FX-55 at $899 is a $100 more expensive than the OEM Opteron 250 and is the most expensive CPU AMD sells below the Opteron 8xx series.

 

[Sideswipe001]

To clear up a myth: Dual processor does not ALWAYS mean ECC or Registered RAM. I've been running a dual Xeon for years on regular old Mushkin PC3200. Look at the motherboard. Opterons might require ECC, but Xeons don't.

Is it the fastest gaming machine on the block? Nope. But I do get a good 3.3 Ghz out of my 2 2.8 Ghz Xeons, which makes it good enough to play anything out there. And the fact that I can do anything on this machine that I feel like it made it worth the money to me.

In other words: if I was making a PURELY gaming machine, I would never use dual processors. But for a multi-purpose workstation/computer/gaming machine, show ME a system with a single processor that's as fast or as responsive.

 

[gsellis]

Originally posted by: Rike
I've taken it as a almost as a matter of faith that a multiprocessor rig holds no real advantages over single CPU systems. That's been challenged and I have to admit I have no data one way or the other. Clearly, cost of a multiproc system is a disadvantage but what else is going on here? Is there a performance advantage in gaming. If so, how much? Those with experience and evidence please post your links.

I have a dual Xeon. For almost all games, I could get a faster AMD or P4. Consider that a 3.6 Xeon is up in the high-end of the EE chips on price. Add a mobo, and you have not gained any gaming performance and still spent more money.

For gaming, get a FX53-55, 3.8 Prescott, or EE if you have the money and have to just have the hottest thing. But remember that the difference between 70fps and 55fps in a gaming system may not be apparent to you.

Gaming apps just are not multithreaded enough to get any benefit.

 

[Rike]

Originally posted by: IHYLN
why do you need proof when the guy said that there are no games that are multithreaded?

As I said above, I already agree with Hajime. What I think and what he thinks are the same. But I have no proof or evidence to back-up what I think and (no offense) nether did he. And because when I say something to others, I want to know I'm right not just believe that I am it. Sometime we say so much (especially "younger" members like myself) without the experience to back it up. I'm just looking for some independent sources. Like some the posts above.

 

[IHYLN]

I think the lack of benchmarks on dual CPU systems should prove a point

 

[gsellis]

Search Tom's Hardware around late 2003 (October?). They did a benchmark of a PC-DL doing gaming.

 

[Rike]

Originally posted by: IHYLN
I think the lack of benchmarks on dual CPU systems should prove a point

lol Point taken!

 

[VirtualLarry]

If you are talking about the tradeoffs between a fast uniprocessor CPU, and a pair of slightly-slower processors in an SMP config - for gaming, all other things being equal, the single CPU should be faster than the pair. Even if they were the same speed, the bus contention between running two CPUs (in an Intel DP config), would make them slightly slower, as would enabling HT on a single-CPU system in many cases, with respect to games.

Now, if you are talking about boards with un-equal hardware, that's something different. Having multiple PCI-X, PCIe busses, onboard workstation/server-quality GigE (with low CPU utilization numbers), etc., those can all lessen the system I/O load, and make it more efficient for the task at hand, which can include games. An onboard PCI-X 64-bit U160 SCSI controller, with a pair of fast 15K RPM drives, is going to most likely be faster than even a WD Raptor setup for gaming.

Also, there is a tradeoff in the RAM, often SMP systems are server boards that require (not always) ECC registered RAM. So they might clock slower, but generally those types of boards allow you to stuff more RAM onto the board in total.

So if you plan to FSB overclock, then a workstation/SMP board is a bad idea, usually because of all of the onboard I/O 'goodies'. However, if you plan to stuff it as full of RAM as possible, then it could be a good idea. Also, if you host multi-player games, and especially if you play on the system system in those games, then an SMP system with plenty of RAM could be quite valuable, as you could run your client on one CPU, while running the server on the other, and get the lowest lag possible.

But other than those possibilities, I wouldn't recommend it. It's usually cheaper to get a stripped-down but fast "enthusiast" uniproc. mobo, add the appropriate components, add decent cooling and a decent PSU, and tweak the living heck out of it, including FSB tweaks. For the same amount that you might spend on a decent SMP workstation board, you could have a much faster single-proc "gaming rig".

 

[Mrvile]

Is this the totally-pwn-mrvile thread?

 

[uOpt]

Not only did games not to bother to multithread themself.

For very high performance applications, if they don't run on a SMP system, it is a slowdown to do threading.

The preferred method to gain concurrency in high-performance games is an event loop. This nails it down to single-thread.

Several years ago, John Carmak was writing that he had tried a Quake3 prototype where he split the pushing of things into the graphics card into a seperate thread from another thread running everything else. It was slower than the event loop driven Quake3 and to the best of my knowledge nobody tried afterwards.

Martin

BTW, I didn't see too many servers requiring ECC and/or registered RAM. Registered is usually required for board with many RAM slots, and ECC is usually not required at all. There are plenty of SMP board running with standard memory. I think this forum is so full of people who have AMD as 90% of their horizon that they assume everybody makes the same mistakes as AMD did with the Opteron (requiring registered, not ECC RAM).

 

[The MACdaddy]

Originally posted by: MartinCracauer

...Several years ago, John Carmak was writing that he had tried a Quake3 prototype where he split the pushing of things into the graphics card into a seperate thread from another thread running everything else. It was slower than the event loop driven Quake3 and to the best of my knowledge nobody tried afterwards....

Carmak ended up releasing a SMP version of Quake 3 for Macs. The performance boost was considerable. This was to help make up for the god awfully slow CPUs Motorola was sticking Apple with.

 

[pm]

It's always seemed to me to be a bit of a chicken and egg thing. Not many people own dual-processor systems that they use to play games, so the developers don't bother writing dual-processor games that take advantage of two CPU's. Especially since have one code base that is multi-core will slow down single-core systems, and splitting the code base to run single and multi core versions depending on the hardware configuration requires much more validation work.

In terms of the applications that I use at work on our dual-CPU systems, I never seen one that didn't gain at least a small speed-up by changing them to use multiple threads/cores so that the extra CPU that's otherwise largely idling actually does something. Of course, this is not exactly a valid test since CAD programs for designing microprocessors are probably not valid comparisons versus something like a game.

There can be no doubt that the future of the microprocessor server, workstation and desktop market is multi-core CPU's. It's only a matter of time until the applications follow. I am sure that developers will find a way to use the extra processor for improving performance.

 

[Rike]

That's a good point. It looks multi-core is the way Intel and AMD are going to get the next round of big speed bumps. As that base builds, programers will finally have a reason to write multithreaded code for all types of programs. An interesting prospect.

 

[VirtualLarry]

That's true, actually. The entire gaming industry is probably at a crossroads over this issue right now, especially given that nearly all future popular desktop PC/gaming-system CPUs are going to be multi-core, and even upcoming consoles will be based on a heavily multi-threaded or multi-core architecture. ("Cell" for PS3, rumored multi-core PPC for XB2, etc.)

Another similar class of applications - game emulators running on PCs. They share many of the same performance/behavior traits that high-performance PC games do, and are primarily single-threaded.

 

[dullard]

There are dual CPU gaming benchmarks. They are just harder to find.

Not too dated review of dual gaming performance. Used Ati 9800 Pro 128 MB video card. 70-90 fps in UT2003 Botmatch at 1024x768 resolution.

Not too dated review of single CPU gaming performance.. Used Ati 9800 Pro 256 MB video card. 70-100 fps in UT2003 Botmatch at 1024x768 resolution.

Ok there was a slight difference in video card memory. Little if any difference in UT benchmarks between the two video cards.

The results as you can see are basically the same with one CPU vs with two CPUs. Heck even if you compare basically the same two chips the results are within 1 fps or so (2.8 GHz Xeon vs 2.8 GHz P4 for example).

 

[pm]

I was looking at the Steam Hardware survey results page that I found while registering my copy of Half-Life 2. It's here.

Anyway:
1 CPU systems: 978,576 (99.8%)
2 CPU systems: 1,972 (0.2%)

Clearly a very small market to attempt to tune performance for.

 

[obeseotron]

Dual CPU's are notoriously hard to program games for. Xbox and Gamecube are much easier to design for than PS2, and going back a generation, the Saturn was harder than the PS1 and n64. While the Saturn and the PS2 couldn't be further apart in terms of success, the progression graphically on both systems was huge over the lifespan of the system as developers got a better feel for it. Look at the the first Madden version verus the current one or ESPN football on the PS2 or in the dated saturn example, a buggy virtua fighter 1 followed by a nearly perfect VF2.

Maybe they'll get it down eventually but I don't really think so. PC games these days tend to be FPS based on one of 2 (now 3 with source), console ports, or stategy/rpg games that usually aren't as dependant on horsepower. None of those engines see benefit from dual CPU, nobody porting console games will bother, and performance isn't usually a big deal with the rest. I think Dual CPU on the desktop will only help accelerate the end of serious PC gaming which has been clearly looming for some time by slowing down game performance, or at least it's progress.

 

[Hajime]

Actually, of late, dual CPUs have gotten radically easier to program for, Obeseo.

I'm speaking from experience here, btw. The tools have improved radically.

However, it's -not- easy to make a program multithreaded unless you plan on it from the start. It's not easy at all to add it as an after-thought.

 

[Rike]

Well here's to the multithreaded, multicore world of tomorrow. :beer:

Time to load HL2 on to my two year old rig.

 

[Hajime]

Rike: I can't wait.

There is one major benefit to multiple CPUs - and the exact reason why I'd rather develop on a dual P3 700mhz instead of an A64 FX-55.

Responsiveness. I can have a program compiling the bejeebus out of itself while I'm working on stuff in another window and not see even the -tiniest- bit of lag. If I try that on the A64, ewww.

It's like Hyperthreading on steroids . Doesn't do much for modern games, but for any sort of multi-tasking, it is great.