The end of 775 socket (LGA775)

[tuanming]

Sure it is useless; it useless because we have useless programmers.

They must be pretty badly optimized to require 8 cores....

No, they don't have to be and 8 cores is even out yet... Simply put, if you have no use for "multi-core" processor then don't buy it, end of story. Don't judge a processor when it is not even on the market, unless you have a grudge against it

[Squall-Leonhart]

Having physics an graphics done on the gpu will be more beneficial then sticking the physics data through 3 bus interconnects just to send it to the video card anyway.

[cottonvibes]

it seems here we have a contradictory ideas.


first someone mentions that there are no applications that take advantage from 4-8 cores.

then later when someone mentions there are apps that take advantage of 4-8 cores, and they're called "un-optimized."


so it seems no matter what is said, some people just won't accept that multi-cores are a good thing...


i don't understand why theres so much hate about multiple cores.
it dosn't matter if the multiple cores will only be used for 1% of the time, they're still a good thing.

if you had the choice between a 2-cored processor @ 4ghz, and a 8-cored processor @ 4ghz, you'd be a fool to choose the 2-cored processor.

obviously with the law of diminishing returns, the more cores a processor has, the less performance increase it gets.
But thats not to say that 8-cores are "useless."

and the faster we get 8-cored processors on the market, the faster coders can start experimenting with the benefits of the extra cores.

no-one is making you people buy the 8-core processors.
if you're happy with your single and dual-cored processors, stick to your core2duos and Pentium 4s,

[runawayprisoner]

Originally Posted by tuanming

Sure it is useless; it useless because we have useless programmers.



No, they don't have to be and 8 cores is even out yet... Simply put, if you have no use for "multi-core" processor then don't buy it, end of story. Don't judge a processor when it is not even on the market, unless you have a grudge against it

Dual quad xeons are 8-core setups... And it's true programmers were taught with everything in mind. The implementation is another thing. What sucks so much is not that coders are not skilled enough, but rather, they (the processor manufacturers) made it harder for us. Which also explains why most current coders suck at making multithreaded apps. They don't suck because of their skills.

Originally Posted by cottonvibes

it seems here we have a contradictory ideas.


first someone mentions that there are no applications that take advantage from 4-8 cores.

then later when someone mentions there are apps that take advantage of 4-8 cores, and they're called "un-optimized."

But they are. Only a lazy coder would thread things out and try to sync them up in hope of gaining some performance. And well... he DOES gain some performance boost, but the application should be written so that it squeezes out as least resources as possible. A programmer's goal is not to stress hardware to their most extremes.

if you had the choice between a 2-cored processor @ 4ghz, and a 8-cored processor @ 4ghz, you'd be a fool to choose the 2-cored processor.

Realistically, we have no choice, though. That aside, I'd still choose 2-core @ 4GHz over 4-core @ 4GHz. Why? The latter is hotter... always.

and the faster we get 8-cored processors on the market, the faster coders can start experimenting with the benefits of the extra cores.

Cell BE on the PS3 is more interesting to mess with than a Core 2 with more cores. Seriously.

[cottonvibes]

Originally Posted by Squall-Leonhart

Having physics an graphics done on the gpu will be more beneficial then sticking the physics data through 3 bus interconnects just to send it to the video card anyway.

like i said b4,
the coders will then have to learn nVidia's API for physics on their cards.
and then they'll have to learn ATI's as well...

or else their physics would only work on 1 brand of cards.


having it done with the CPU is better IMO.

the main thing that takes up speed isn't the transaction between CPU to GPU, its really the transformation of physics operations for all the different particles.

for example.

you wish to move 1000 particles of dirt in an explosion.

you'll need some complex math to be preformed for every particle, so ur program will run an algorithm 1000 times to transform 1 pixel of dirt, to its next location on the screen.

it then saves those coordinates, and transfers them to the GPU to be drawn.

i guess it could be faster to do it all on the GPU, but its not worth learning 2 different Card's Libraries.

but like i said,
if a standard library like DirectX can handle physics for both cards, then i guess it would be nice.

[runawayprisoner]

Originally Posted by cottonvibes

like i said b4,
the coders will then have to learn nVidia's API for physics on their cards.
and then they'll have to learn ATI's as well...

or else their physics would only work on 1 brand of cards.

And who cares then? Either way, we still don't know for sure if there will be a DirectPhys in the next DirectX version. If there will be, that fixes your problem. All current graphics card manufacturers for the mainstream market still have to conform to DirectX standards. Even if they play the different API game, that just means... whoever has the easier and faster API wins. It also means you don't have to pay for an 8-core processor. Current quads are still going for $300, while their duals equivalence are $150 - 200. How much do you think an 8-core processor will be when it comes out? (And by the way, the estimated price tag scares me...)

the main thing that takes up speed isn't the transaction between CPU to GPU, its really the transformation of physics operations for all the different particles.

for example.

you wish to move 1000 particles of dirt in an explosion.

you'll need some complex math to be preformed for every particle, so ur program will run an algorithm 1000 times to transform 1 pixel of dirt, to its next location on the screen.

it then saves those coordinates, and transfers them to the GPU to be drawn.

Transfering some GBs of data isn't going to be fast, either. Besides, a brute force physics processor with 128 physics units beats an 8-core processor any day.

[cottonvibes]

Originally Posted by runawayprisoner

And who cares then? Either way, we still don't know for sure if there will be a DirectPhys in the next DirectX version. If there will be, that fixes your problem. All current graphics card manufacturers for the mainstream market still have to conform to DirectX standards. Even if they play the different API game, that just means... whoever has the easier and faster API wins. It also means you don't have to pay for an 8-core processor. Current quads are still going for $300, while their duals equivalence are $150 - 200. How much do you think an 8-core processor will be when it comes out? (And by the way, the estimated price tag scares me...)

so either your CPU is going to be expensive, or your GFX card is going to be expensive...

price is irrelevant, as either way, its going to be expensive.


and another reason its better to have an 8 cored-processor vs GPU-based physics is because the extra cores don't have to be used specifically for physics, whereas having the GPU preform physics calculations would limit it's power specifically for physics.

Originally Posted by runawayprisoner

Transfering some GBs of data isn't going to be fast, either. Besides, a brute force physics processor with 128 physics units beats an 8-core processor any day.

you won't be transfering GBs of data

if there was that much data being transformed, it would be slow regardless.

and data still has to be transfered within the GPU as well...
The GPU has its own Bus to transfer information between it's processors and memory.
its faster than from CPU -> GPU, but its still not instantaneous.


and a physics processor would be faster than a general-purpose processor at physics, but thats all it would be good for.

[runawayprisoner]

Originally Posted by cottonvibes

so either your CPU is going to be expensive, or your GFX card is going to be expensive...

Nah, more like... your GFX card is ALWAYS going to be expensive, and pay for your CPU at your own risk. I mean... who among us did not pay at least $200 for his high-end graphics card? Anyone?

price is irrelevant, as either way, its going to be expensive.

It's another issue you need to take into account, if you are to analyze this situation to its fullest potential. I mean... we are talking about the benefits and losses of dual-core and quad-core, or octo-core processors, not just about multi-threading technology?

and another reason its better to have an 8 cored-processor vs GPU-based physics is because the extra cores don't have to be used specifically for physics, whereas having the GPU preform physics calculations would limit it's power specifically for physics.

Indeed, but what else do you need to use the processors for? Mmm? Because in a game, it's either A.I., or physics, and A.I. doesn't need more than a dual-core processor for some crazy conditional branches.

you won't be transfering GBs of data

Really now? Because coordinates are stored in floating point precision values, or better yet, double floating point precision values. Only the graphics card can determine where the particles are in the space. Up to 1000 particles with each coordinates stored as floating point precision values, and their other parameters to identify them as particle objects in the game, as well as their lighting coordinates and a whole slew of other values ought to take up at least 1MB each. That's almost 1GB for 1000 particles. Realistically, a matrix would be best, but since our processors don't have that big a cache, they are going to have to store everything into RAM, which means... the GPU has to read everything from RAM to its own integrated memory units, which is way slower than direct CPU -> GPU transfer. You can send coordinates on the fly if the processors are fast enough, but if not, latency is going to waste bandwidth, and/or bog bandwidth, causing tremendous delays.

if there was that much data being transformed, it would be slow regardless.

Yeah, hence why a monstrous processor dedicated only to calculate is better than 4 processors dedicated to doing a lot of other things.

and data still has to be transfered within the GPU as well...
The GPU has its own Bus to transfer information between it's processors and memory.
its faster than from CPU -> GPU, but its still not instantaneous.

It's not just faster than from CPU -> GPU, it's significantly faster. PCI-E x16 bandwidth is obviously at most half that of the transfer rate between the GPU and the integrated memory units. Actually, it's up to about 4x faster than PCI-E x16 with GDDR4.

and a physics processor would be faster than a general-purpose processor at physics, but thats all it would be good for.

And again, what do you need that many processors for? Do you play Crysis while encoding and watching a BD movie at the same time? Or do you also include PCSX2 into the bunch? That's the only way I can see a typical user using an 8-core processor. Quite honestly, with a quad-core, I can already play PCSX2 while watching and encoding a BD movie. It's uncanny how people just keep upgrading blindly, without knowing how much they need. This is why I have a business, too. I advise people and address what they need. I don't just blindly sell a computer for a high price.

[Phil]

Spoken like a true dealer

[cottonvibes]

Originally Posted by runawayprisoner

Really now? Because coordinates are stored in floating point precision values, or better yet, double floating point precision values. Only the graphics card can determine where the particles are in the space. Up to 1000 particles with each coordinates stored as floating point precision values, and their other parameters to identify them as particle objects in the game, as well as their lighting coordinates and a whole slew of other values ought to take up at least 1MB each. That's almost 1GB for 1000 particles.

i doubt 1 particle will end up taking 1MB, infact all 1000 particles won't even take up 1MB.

On traditional 32-bit systems, Double Precision values take up 8 bytes (32bits).

if you store X,Y,Z coordinates as Doubles, then you'll only be using 24 bytes (192bits)

then we can also add Lighting info (not sure if lighting would be applicable, since we were only talking about transforming their location with physics, but i'll include it anyways)

anyways, i'll consider lighting and various other information about the particles as being an extra 4 doubles.

so, so far we have 7 doubles of information per each particle.

that adds up to 56 bytes (448 bits).


so if we have 1000 particles each storing 7 doubles of information (7doubles = 56bytes = 448bits), we end up with 56,000 bytes worth of data.

and all that only equates to 54.6875 kB! (1kB = 1024bytes)
not even close to 1GB, and not even 1mB, and thats for all 1000 particles!




anyways, i guess to get back on topic, 8-cores for the average user might be over-kill.
but i'm sure people will find ways to utilize the extra cores.

maybe just run folding@home if you really can't find a use for them...

[Fadingz]

I want clock speed, not number of cores
Duo Xeon is 8 cores already...and it sux for anything but rendering...

[cottonvibes]

Originally Posted by Fadingz

I want clock speed, not number of cores
Duo Xeon is 8 cores already...and it sux for anything but rendering...

yeah i rather have clock-speed rather at this point, since we're already past 4-cores.

i wish intel focused on a 5ghz quad-core, instead of a 3ghz 8-cored CPU. (well not sure what the clockspeed is going to be, but i guess its around there...)


anyways, i guess intel is really working on what they said a while back, something about having 86-cored cpus in 4 years... (something like that...)

[runawayprisoner]

If you want to run folding@home, buy a PS3. Or a cabinet with about 10 Cell blades do, too. Each blade has two Cell BE, and quite honestly, I bet that thing would still wipe the floor with Nehalem. Until Nehalem becomes reasonably priced, I might change my opinions, but right now, you wouldn't even want to know the "estimated price" of a true 8-core, single Nehalem processor system... a 4-core Nehalem seems more reasonable, but is still quite out of reach for the mid-end users.

[tuanming]

With each nanometer shrink or new micro-architecture there will always be performance gains. Each generation of new micro architecture has always boost performances compared to the previous generation clock-for-clock, more IPC (instruction per clock) or OPC (operation per clock) and bring along increase overclocking capabilities.

Don't be surprise if a 2.4Ghz Nehalme is as fast as the 3.0Ghz Core 2 Duo/Quad

[skoreanime]

Errr, that's a given . Thanks for stating the obvious though .

As long as my games can look decent and have decent fps, I doubt I'll have to upgrade to a Nehalem to do that with games coming out this year and next. Like RAP generally said, why upgrade when you don't need to?