Last Thursday, AMD released a 392-page PDF (v1.0 dated March 2009) outlining the Instruction Set Architecture (ISA) native to the R700 family, and display a difference between the Radeon HD 4870 and the Radeon HD 4890.
Along with outlining the R700 ISA, AMD's loaded document also defines the instructions and formats accessible to the programmers and compilers, covering the RV710 (Radeon 4350/4550) on through to the upcoming RV790 (Radeon 4890). However, it isn't for the typical consumer looking to see what the R700 family has to offer, but rather is geared towards the programmer writing application and system software including device drivers, operating systems, system utilities and more. The document actually serves to help these programmers and compiler writers maximize processor performance. Still, it's a good way--especially for those who really enjoy the technical side of GPUs--to see what's under the hood, and the differences between a few of the GPU variants.
According to Chapter 7 in the PDF, the RV790 supports burst memory reads while the RV770 does not. "Burst memory reads are not supported by the RV770; however, the 710, 730, 740, and 790 do support it. Chips after R770 support burst reads in memory-read instructions. This allows up to 16 consecutive locations to be read into up to 16 consecutive General Purpose Registers (GPRs). This adds a new field BURST_CNT to MEM_RD_DWORD0," the document states. "For each iteration of the burst, the DST_GPR is incremented by 1 and the ARRAY_BASE is incremented by 1. The SRC_GPR is not affected." So because the RV770 series (Radeon 4870/4850/4830) does not support the burst memory read function, it's safe to assume that the RV790 (Radeon HD 4890) will work more efficiently.
As defined by AMD, burst mode read is a synchronous operation that is tied in to the rising edge of a clock; the microprocessor/microcontroller supplies only the initial address to the device. Using an architectural approach such as bust mode enables faster access times, as higher density and lower voltage tend to reduce performance in a standard random access memory architecture. With that said, devices using the burst mode architecture offer improvements in system speed and performance by reducing sequential read access times. Additionally, AMD burst mode devices have two different read modes: random read and burst mode read.
There's definitely quite a lot to read within the 392-page document, spanning ten chapters can covering topics such as control flow programs, memory read clauses, data share clauses, and microcode formats. To read more about the R700 family architecture details, download the PDF for AMD right here. Although the document contains 392 pages, the file weighs a mere 1.9 MB.
