Number of Cores vs. GHz

[Saiki]

@Xellon: as of now there's one (game) for sure which harnesses both 64bit and quad cores. it will take some time, but they'll get to it sooner or later on the rest

yea, I figured as much. But what about for rendering? Is a quad much more effective or only a little? From what Shadow lady said, it seems like a little.

no idea, I don't own a quad, so I couldn't tell you.

[rex1825]

...well, I use 3DSMax & Maya, althou I don't know what usage is from Intel's HT, but on my PII 955BE, measuring stock vs Athlon 64 6400+ also stock, I got around 60%-70% boost in rendering...

Cheers...

[KrazyTrumpeter05]

Informative answer to my question as always, Air.

I have faith that some clever person will come up with an easy/reliable way to code for multithreaded stuff at some point, though. It's just a matter of someone coming up with the right inspiration.

[Air]

More likely you'll see changes to chip architecture instead, because it's just too difficult to be a good parallel-process programmer. There's already the OpenMP stuff for trying to expand the horizons of parallel programming (and I dare you to try to read content on openmp.org and understand any of it), and the tools they provide handle a lot of special and/or common cases with relative efficiency ... but when it's all said and done it still runs into most of the same fundamental issues that any parallel programming task runs into: in real-world scenarios making scalable and efficient parallel solutions is extremely challenging, and typically requires a lot of detail to exchanging specific brands of data dependencies and responding to user input/requests, all while avoiding race conditions and deadlocks. The seer complexity of the content of openmp.org points to the fact that th goal of simplifying parallel processing isn't really much successful yet.

Even in the mundane world of servers and databases, utilizing multi-core systems efficiently has been a challenge. Google ended up doing by far the best job of it by not going multicore at all, and instead using Penium90 server farms. That's in large part because multi-core systems have a lot of inherent disadvantages and performance penalties (many of which Zeydlitz hit on earlier) -- CPU cache contention, ram bandwidth contention, PCI/AGP bus contention, storage device (HDD) contention, etc. Sometimes (many times?) it's just better to have a bunch of closed systems and specialize their tasks, than it is to have one large system that has to spend half its time just managing it's massive breadth of internal resources.

Applying that back to gaming: one of the biggest challenges of using 4+ cores efficiently is memory contention. If the data you're working on isn't in the L2 or L3 cache, you're screwed and all your cores will just sit around and fight over access to your system RAM. The solution that the CELL takes is to have the cache micromanaged by the programmers, to the extent that the SPUs themselves (the worker cores of the CELL) work entirely from CPU cache ram only. As a programmer for the CELL, you have to explicitly upload data into the cache and then run programs to do work on that data. In that way, the CELL's design is not unlike Google's server farms, where each core of the CPU is more or less its own "contention-free" entity that can be afforded the ability of performing it's single isolated task.

The problem, of course, is that it takes the top skilled 1% of programmers in the world to be able to pull off something like Google's server farm, and not have the whole thing come crashing down on them due to race conditions, deadlocks, or other difficult-to-debug problems. If the CELL is to be the future of software design then the programmers of the games of the future will be as uncommon as brain surgeons.

[ChickenLiver]

Yea, I agree completely with you on OpenMP. Currently, it's great for parallelizing code with fairly simply interdependencies with minimal effort, but doesn't help much with much more complicated cases. I've used it in a number of pure number crunching tasks, which I wouldn't have spent the necessary time to implement using CreateThread or the like. Nothing I've used it for would have taken more than an hour or so more of work to use CreateThread() instead.

[KrazyTrumpeter05]

Ah, so the push for better speeds/efficiency will be different architectures?

How do things like scientific models run? Are they multicore processors, or many computers linked together? Er, well, I guess I should ask if supercomputers just use lots of single core processors or have they gone the multicore route as well?

[Air]

Supercomputers are typically multiple single-core processors, but still follow the basic "monolithic" approach as a multi-core cpu in our machines. Multiple single-core CPUs are more efficient because each one has its own unshared caches and fully dedicated socket architecture, but they still have to contend over all other shared resources in the system (ram, pci, storage, what-have-you).

Google was the first to really prove that the supercomputer model is flawed, when they basically surpassed the finest supercomputers in the world with an army of P90's worth about $200k at the time. As far as I know, some of the other recent supercomputer-like jobs have also gone from supercomputers to PC farms, following in Google's footsteps. But I also know that supercomputers and mainframe servers and such are still in use and still being sold -- even though it's been widely proven that their performance-per-dollar ratios are really poor compared to distributed computing solutions. (otoh, you don't need a brain surgeon on staff to effectively use one, so they do still have their upside).

[General Plot]

In reallity, 4-cores/8 thread CPU could not use 8 Mb L3 cache + 4*256Kb in such effective way, than 2-cores, 6 MB L2 cache (plus DDR3 memory suxx). And, 920 does not throttled to 3.4 -- only to 2.93 (3.43 cap -- is i7 860, not 920). Try do testing by yourself, it could be a nice experience. Really, i7 920 is slow on comparison with E8600 at many task, that users usually use.

I meant overclocked I have mine at 3.4. Here's some Sisoft benches. I'll have to do it again later and compare it to the E8600.

[KrazyTrumpeter05]

Supercomputers are typically multiple single-core processors, but still follow the basic "monolithic" approach as a multi-core cpu in our machines. Multiple single-core CPUs are more efficient because each one has its own unshared caches and fully dedicated socket architecture, but they still have to contend over all other shared resources in the system (ram, pci, storage, what-have-you).

Google was the first to really prove that the supercomputer model is flawed, when they basically surpassed the finest supercomputers in the world with an army of P90's worth about $200k at the time. As far as I know, some of the other recent supercomputer-like jobs have also gone from supercomputers to PC farms, following in Google's footsteps. But I also know that supercomputers and mainframe servers and such are still in use and still being sold -- even though it's been widely proven that their performance-per-dollar ratios are really poor compared to distributed computing solutions. (otoh, you don't need a brain surgeon on staff to effectively use one, so they do still have their upside).

Ah, so then how are these new petaflop supercomputers put together?

[mackoyski]

Better IPC for the next Intel / AMD architecture I'd say....