@Mr A . Redhouse. Just a minor point.
In addition the AmigaOne has a major inbuilt bandwidth benefit compared to normal PC architecture. This is because the transparent synchronous design of the Articia S allows direct memory-AGP transfers without suspending cpu-memory operations as is required in the x86 world. Just ask any user who has installed the same version of, say, PPC Linux on an 800 MHz AmigaOne and a PC of 3x that clock speed as to which is the most responsive. The A1 wins every time.
- Mr A . Redhouse, Nov 2003.
To quote nVidia's IGP TwinBank Memory Architecture for the AMD K7(X86-32) platform.
III. The TwinBank Architecture
Crossbar Memory Controller
TwinBank consists of two independent 64-bit DDR-266 memory controllers (MC0 and MC1) to deliver a
whopping 4.2GB/s peak memory bandwidth**. This is four times the memory bandwidth of PC133 SDR
memory and over two-and-a-half times of single RAC 800MHz DRDRAM. The radical crossbar memory
controller enables CPU and GPU to concurrently access the two 64-bit memory banks and is optimized for
64-bit CPU and GPU accesses to ensure near perfect bandwidth utilization. The two memory controllers
are interleaved so that consecutive CPU memory requests can be started before the previous one is
completed, reducing CPU read latency. The TwinBank architecture allows the two independent 64-bit
memory controllers to access 128-bits of data on each clock cycle using DDR memory, effectively fetching
256-bits of data total on each clock cycle. Since the high-performance CPU and GPU data types are
optimized for 64-bit access, both can access the two memory banks simultaneously and independently,
fully utilizing available memory bandwidth. The average read latency of the CPU is now greatly reduced,
which increases both graphics and system performance. Without this type of architecture, there would be
tremendous bottlenecks in the system with the high-performance CPU and GPU both struggling for
access to valuable system memory bandwidth.
- (1)
the end-user has the option of using an even more powerful
external AGP GPU, such as NVIDIA’s GeForce3, which also takes full advantage of the TwinBank dual
independent 64-bit memory controller architecture for dramatic increases in performance.
- (1)
This optimization is on top of NVIDIA's "dynamic adaptive speculative pre-processor" (DASP)(2).
Applying generalization in relation X86 markets is not recommended in the light of intense competition in the X86 market.
Reference.
1.
NVIDIA nForce IGP TwinBank Memory Architecture, Pages 5,7.
http://www.nvidia.com/object/LO_20010528_5545.html2.
http://www.nvidia.com/object/dasp_tech_brief.html**Minor updates; Corresponding bandwidth increases with PC3200.
Can one realistically apply "The A1 wins every time" (in regards latency) for the Opteron/AthlonFX/Athlon64 X86 based systems?
This is because the transparent synchronous design of the Articia S allows direct memory-AGP transfers without suspending cpu-memory operations as is required in the x86 world.
I don’t think one could fit both CPU’s bandwidth (e.g. 3.2GB/s) and AGP’s bandwidth (e.g. 2.1GB/s for AGPX8) on a single channel 64bit bandwidth (e.g. 3.2GB/s). The Northbridge’s bandwidth must at least equal the aggregate bandwidth of CPU and AGP for concurrent access. NVIDIA's dual channel controller enables this for AMD K7’s case. Bad luck for VIA K7 KTxxx users.
Intel Pentium VI’s QDR800 (6.4GB/s) pretty much consumes the dual channel DDR400 bandwidths (e.g. 6.4GB/s).
