Convergence Of CPU + GPU?
When they were first introduced, 3D graphics processors were simple, fixed-function devices. Their processing pipelines worked like assembly lines, with each station performing a simple, fixed function. To ramp up their performance, designers merely needed to add as many pipelines as necessary or as manufacturing constraints allow.
Over time though, they started hitting the limit on the number of pipelines they can cram into the die at realistic power and thermal levels. That's when designers changed the graphics rendering pipeline from simple, fixed-function designs to far more complex, programmable designs that are more like CPUs than graphics processors.
According to Intel, this trend of GPU becoming more complex and programmable parallels the current trend in CPUs of increasing parallelism to quickly improve performance without added complexity. In other words, there appears to be a convergence of CPU and GPU architectures. The GPU was becoming more and more like CPUs while the CPUs were becoming more and more like CPUs.
Intel believes that hybrid architectures like Larrabee will offer the best of both worlds - a CPU with the capabilities of the GPU. In essence, CPUs based on the Larrabee architecture are designed to obviate the need for separate graphics processors. That's certainly frightening news for graphics processor companies like ATI and NVIDIA.
Why Many Cores?
At a time when most applications (including games) are not even optimized for quad-core processors, Intel is certainly ambitious in developing a "many-core" design for desktop processors. However, unlike current Intel Core 2 processors, Larrabee-based processors are optimized for graphics and other throughput-intensive applications, and nothing provides a better boost in throughput than lots of processor cores.
You can think of these cores as the processing pipelines in graphics processors. The more pipelines there are, the greater the throughput, the greater the amount of data that can be processed simultaneously. Such a design is not useful for general computing applications that are still mostly single-threaded and have lots of dependencies in their code, but it is ideal for processing graphics data that are by nature highly parallel.
If you are wondering just how effective the "many-core" approach of Larrabee is compared to the current Intel Core 2 processor, Intel gave us an idea with a "design experiment" involving a theoretical 10-core Larrabee processor and a standard dual-core processor of the same die area and power consumption :
|
Current |
Theoretical 10-Core |
|
CPU Cores |
2 |
10 |
|
Pipeline Type |
Out-of-order |
In-Order |
|
VPU Lanes Per Core |
4-lanes wide SSE |
16-lanes wide |
|
L2 Cache Size |
4 MB |
4 MB |
|
Single-Stream |
4 per clock |
2 per clock |
|
Vector Throughput |
8 per clock |
160 per clock |
With such a high vector throughput, Larrabee processors would be much more suitable for graphics processing than current CPUs.
Support Tech ARP!
If you like our work, you can help support out work by visiting our sponsors, participate in the Tech ARP Forums, or even donate to our fund. Any help you can render is greatly appreciated!
Page |
Topic |
|
1 |
||
2 |
||
3 |
||
4 |
• The
Larrabee Texture Sampler |
|
5 |
||
6 |
||
7 |
||
8 |
<<< A Paradigm Shift?, Larrabee's Key Features : Previous Page | Next Page : Inside The Larrabee, The Basic Core Design >>>