This new Sun processors that feature CoolThreads is very curious from the perspective of how many threads can be processed at once with at a notably low power wattage...
I've read many articles about this processor and in one of these articles said about the implementation of that CPU. It's very interesting, because the Sun Niagara's power isn't raw power (aka high clockspeeds), but throughput. In this article which I read, it has a part about the underutilization of CPU in a Technical, Desktop and Server workloads. In the technical enviroment (that utilizes floating-point instructions more frequently), about 48% of all clocks are not utilized, due to memory latency.... In a desktop enviroment (that utilizes integer instructions more frequently), the percentage climbs to 52% and in a server enviroment, this percentage go to 85%...
Well, Sun is aware of this and created this Niagara CPU with a goal of keep it busy at the maximum levels for a server workloads... If the underutilization in this enviroment is too high, the CPU can process a new thread, because the other thread is stalled waiting to be feeded with data or writing it's results in memory, right? Another important thing in this enviroment is that Out-of-Order Execution or extremely deep buffers or pipelines aren't too useful... We can see this in a Intel Xeon Processor.. It isn't as efficient as Opteron, due to these... Xeon has a large OOO engine and a deep buffer because of HyperThreading. But, due to it's complicated circuitry, achieves extremely high wattage levels. Sun Niagara is a simple CPU, it hasn't all of the power of Xeon's OOO engine, it hasn't a 31-stage pipeline like Xeon, which it turn in a relatively simple CPU (and a small core!) and can be grouped with other cores to have a multi-core server, because the cores are so simple and are low-power, in order to achieve high throughput, which is the most important performance measuring in a server enviroment.
Intel probably is aware of this too, of course... It hasn't so much importance in put a Desktop/Technical enviroment CPU in a Server enviroment CPU... it was like to put a Ferrari engine in a VW Beetle structure... Or drive that Ferrari in a autobähn with the handbrakes activated... The Xeon has too horsepower (but Opteron has more yet...), but this horsepower is underutilized, why? Because Xeon is more a Technical/Desktop CPU than a Server CPU. In a IT enviroment, it won't be so useful how well Xeon performs in 3DMark CPU test, but how many threads can be processed at once. Intel could create market-segment-driven CPUs: Thin&Light (PentiumM LowPower), Notebook (Regular Pentium M), Desktop (Pentium 4), Performance Workstation (where Xeon and P4XE fits), (all of these already has), and Server (that will be a multicore CPU, it will be more-or-less be like a Cell CPU (a group of SPE's...)).
I've read many articles about this processor and in one of these articles said about the implementation of that CPU. It's very interesting, because the Sun Niagara's power isn't raw power (aka high clockspeeds), but throughput. In this article which I read, it has a part about the underutilization of CPU in a Technical, Desktop and Server workloads. In the technical enviroment (that utilizes floating-point instructions more frequently), about 48% of all clocks are not utilized, due to memory latency.... In a desktop enviroment (that utilizes integer instructions more frequently), the percentage climbs to 52% and in a server enviroment, this percentage go to 85%...
Well, Sun is aware of this and created this Niagara CPU with a goal of keep it busy at the maximum levels for a server workloads... If the underutilization in this enviroment is too high, the CPU can process a new thread, because the other thread is stalled waiting to be feeded with data or writing it's results in memory, right? Another important thing in this enviroment is that Out-of-Order Execution or extremely deep buffers or pipelines aren't too useful... We can see this in a Intel Xeon Processor.. It isn't as efficient as Opteron, due to these... Xeon has a large OOO engine and a deep buffer because of HyperThreading. But, due to it's complicated circuitry, achieves extremely high wattage levels. Sun Niagara is a simple CPU, it hasn't all of the power of Xeon's OOO engine, it hasn't a 31-stage pipeline like Xeon, which it turn in a relatively simple CPU (and a small core!) and can be grouped with other cores to have a multi-core server, because the cores are so simple and are low-power, in order to achieve high throughput, which is the most important performance measuring in a server enviroment.
Intel probably is aware of this too, of course... It hasn't so much importance in put a Desktop/Technical enviroment CPU in a Server enviroment CPU... it was like to put a Ferrari engine in a VW Beetle structure... Or drive that Ferrari in a autobähn with the handbrakes activated... The Xeon has too horsepower (but Opteron has more yet...), but this horsepower is underutilized, why? Because Xeon is more a Technical/Desktop CPU than a Server CPU. In a IT enviroment, it won't be so useful how well Xeon performs in 3DMark CPU test, but how many threads can be processed at once. Intel could create market-segment-driven CPUs: Thin&Light (PentiumM LowPower), Notebook (Regular Pentium M), Desktop (Pentium 4), Performance Workstation (where Xeon and P4XE fits), (all of these already has), and Server (that will be a multicore CPU, it will be more-or-less be like a Cell CPU (a group of SPE's...)).
Facebook
Twitter