If you want I/O or network ops to be freed from burdening the CPU, shouldn't their engines be optimized to more of the work? Intel/AMD control the south bridges, let them get together to create a standard to offload the work--otherwise you'll have a crapload of new api's to coordinate the stream processing.
yeah, and then admire teh awesomeness when you won't even be able to run ms word, let alone crysis. there's a huge difference between fixed function and x86.
@BB: I agree that a stream processor is not the solution. What could be held though from Mick's thoughts is the use of "lightweight" processors for common repeated tasks that require significant less resources. We could be using an asymetric multiprocessor CPU: two to four full featured cores like the ones used in Phenoms and Core's and one or two very simple, very energy efficient processors that could handle common OS tasks and some lightweight processes. This would require a new programming model (how does one define a process as lightweight?) and apparently new scheduling algorithms for the OS, but the gains - especially in the case of energy consumption - could be significant. Imagine using a system that could work with the power consumption of an Atom processor when doing things like email and word processing and the processing capabilities of an i7 while making Matlab computations or encoding tasks.
@Mick: Why would AMD want to put stream processing on their chips? Hardly anybody uses tasks that demand it, and those that can represent a very small segment of the industry. It's hard enough to parallelize code as is, but to have to engineer your code such that it shys away from random access memory, and rely on purely cached data? No thanks. Only very specific tasks can even fit into that paradigm, and on the desktop, they are very few. Most processes can't be "streamed" because they're too complex, and really, what is the point? CPUs are very underutilized as is, so 12% "potential" CPU savings is nothing. I'd rather have them spent the research time try to develop better memory controllers for the tasks we *do* generally run.
CPU's really need stream processors like graphics cards. When system processes (networking, drive access, mouse and keyboard input, audio, registry access, and others things) happen, they take processing time away from the CPU. It may only be 1 or 2 percent, but this can slow down intensive apps more than that simply by the way different takes share CPU time (and the switch time between them). By using stream processors, these small processes can be undertaken separate from the CPU, freeing up performance that is otherwise lost. Say if the processing power of the stream processors add 12 percent performance to the computer, due to the nature of multitasking this could lead to a much greater performance benefit in theory than the 12 percent.
The computer can have dedicated stream processors for the main things outlined above, and any spare processors can be used for programmes that only require a small amount of processing time.
Quad core's been around for a while. It doesn't sound cool anymore. Why does AMD want to close up the gap on Intel when they know it's never going to happen? Just skip the whole damn phenom line and up the number of cores by 2x. If Sun can make a 16 core monster, why can't Arabs?
It looks like AMD has got it's design process *and* it's manufacturing process back on track. Intel probably holds the top end $1000.00 CPU spot for some time but that is fine. Competition in the normal-person market segment is good: if someone wants to buy a desktop CPU for $1000.00, well, more power to her. Paul, have you any insight into how this new model translates into the higher margin Opteron processor market?
If you want I/O or network ops to be freed from burdening the CPU, shouldn't their engines be optimized to more of the work? Intel/AMD control the south bridges, let them get together to create a standard to offload the work--otherwise you'll have a crapload of new api's to coordinate the stream processing.
Larrabeeeee!!!!!!!!!
yeah, and then admire teh awesomeness when you won't even be able to run ms word, let alone crysis. there's a huge difference between fixed function and x86.
@BB: I agree that a stream processor is not the solution. What could be held though from Mick's thoughts is the use of "lightweight" processors for common repeated tasks that require significant less resources. We could be using an asymetric multiprocessor CPU: two to four full featured cores like the ones used in Phenoms and Core's and one or two very simple, very energy efficient processors that could handle common OS tasks and some lightweight processes. This would require a new programming model (how does one define a process as lightweight?) and apparently new scheduling algorithms for the OS, but the gains - especially in the case of energy consumption - could be significant. Imagine using a system that could work with the power consumption of an Atom processor when doing things like email and word processing and the processing capabilities of an i7 while making Matlab computations or encoding tasks.
@Mick: Why would AMD want to put stream processing on their chips? Hardly anybody uses tasks that demand it, and those that can represent a very small segment of the industry. It's hard enough to parallelize code as is, but to have to engineer your code such that it shys away from random access memory, and rely on purely cached data? No thanks. Only very specific tasks can even fit into that paradigm, and on the desktop, they are very few. Most processes can't be "streamed" because they're too complex, and really, what is the point? CPUs are very underutilized as is, so 12% "potential" CPU savings is nothing. I'd rather have them spent the research time try to develop better memory controllers for the tasks we *do* generally run.
I wanna GPU in place of a CPU. o.O
Wouldn't that be faster anyhow..
like quad core gpu 4 a cpu!
Teh awesomeness!!!!!!!!!!
CPU's really need stream processors like graphics cards. When system processes (networking, drive access, mouse and keyboard input, audio, registry access, and others things) happen, they take processing time away from the CPU. It may only be 1 or 2 percent, but this can slow down intensive apps more than that simply by the way different takes share CPU time (and the switch time between them). By using stream processors, these small processes can be undertaken separate from the CPU, freeing up performance that is otherwise lost. Say if the processing power of the stream processors add 12 percent performance to the computer, due to the nature of multitasking this could lead to a much greater performance benefit in theory than the 12 percent.
The computer can have dedicated stream processors for the main things outlined above, and any spare processors can be used for programmes that only require a small amount of processing time.
Quad core's been around for a while. It doesn't sound cool anymore. Why does AMD want to close up the gap on Intel when they know it's never going to happen? Just skip the whole damn phenom line and up the number of cores by 2x. If Sun can make a 16 core monster, why can't Arabs?
It looks like AMD has got it's design process *and* it's manufacturing process back on track. Intel probably holds the top end $1000.00 CPU spot for some time but that is fine. Competition in the normal-person market segment is good: if someone wants to buy a desktop CPU for $1000.00, well, more power to her. Paul, have you any insight into how this new model translates into the higher margin Opteron processor market?