|
Page 2 of 10
Radeon HD 5900 architecture
We covered AMD's flagship architecture in some detail in our Radeon HD 5800 series technology preview, and that article is well worth reading as a full evaluation of AMD's dual-GPU "Hemlock" offering and the "Cypress" core upon which its GPUs are built.
Starting out with the basics, the Radeon HD 5800 series boasts a second-generation 40 nanometre GPU from AMD (with the Radeon HD 4770 their first 40 nanometre part) featuring a die size of 334 mm² against the 263 mm², 55 nanometre die used by RV770, a 1.27x size increase. This pales into comparison to the increase in transistor count however, with the Radeon HD 4870's 956 million transistors dwarfed by the 2.15 billion transistors employed by the Radeon HD 5800 series.
A full Radeon HD 5800 series core features double the number of Stream Processors as RV770, from 800 processors on that architecture up to 1600 Stream Processors here. This change also sees the number of SIMD units employed by the graphics core doubled, from ten SIMD units holding eighty Stream Processors in RV770 to twenty SIMDs which again house eighty Stream Processors each here. All of this equates to a vast amount of processing power altogether, giving a full Radeon HD 5870 configuration 2.7 TeraFLOPS to play with (against 1.2 TeraFLOPs for the Radeon HD 4870) in normal single precision, and 544 GigaFLOPS when handling 64-bit double precision data.
Perhaps unsurprisingly, as per the architecture's Stream Processor count, the number of texture units on show here is twenty, exactly double the number seen in RV770 and matching the number of available SIMDs. This gives "Cypress" a hefty boost to theoretical texture bandwidth, offering up to 68 billion bilinear filtered texels per second and up to 272 billion 32-bit texture fetches a second via a total of 80 texture address processors and 80 FP32 texture filtering units, together with 320 texture samplers.
Moving on to the Radeon HD 5800 series ROPs (or render back-ends, as AMD prefer to call them), again doubling everything is the name of the game here - Thus, the Radeon HD 5800 series in its full configuration is equipped with 32 ROPs split into eight ROP partitions (compared to 16 ROPs and four partitions with RV770). This allows for the Radeon HD 5800 series to handle 32 pixels per clock with anything from no AA through to having 4x multi-sample anti-aliasing enabled, and sixteen pixels per clock in 8x MSAA mode, although these numbers are halved when handling 64-bit colour data. When it comes to processing depth or stencil data alone, Cypress can handle up to 128 pixels per clock cycle.
Given the success of the Radeon HD 4870, it's no surprise to see GDDR5 memory coming to the fore once again here, with 1GB of the stuff coupled with a 256-bit wide memory bus. This memory controller uses the same distributed hub design as the Radeon HD 4800 series first sported, using controllers dispersed around the edge of the GPU core near memory bandwidth intensive parts of the GPU, and connected to a hub which handles all of the lower bandwidth traffic required by the core.
A reference Radeon HD 5970 finds itself clocked lower in terms of both core and memory speeds than the Radeon HD 5870 despite otherwise making use of two full Cypress cores with all of their Stream Processors, texture units, ROPs and so on intact - This leaves us with a core clock speed of 725MHz, coupled 1GB of GDDR5 memory per GPU (2GB in total) sporting a 1000MHz clock speed. The reason for this choice of clock speeds (which also explains the reason for the very modest factory overclocks you might see on some boards, such as Sapphire's offering discussed here today) is that this is the only way that the Radeon HD 5970 could be kept below the 300 Watt power envelope specified by the PCI Express specification, despite the board itself supposedly having been designed to handle a 400 Watt load - A compromise we may start to see more and more often from graphics boards unless that specification is changed in the near future.
Finally, with two GPUs on board the Radeon HD 5970, naturally these two cores need to communicate with one another, and this duty is performed as usual for AMD's dual-GPU cards via a PLX PCI Express bridge chip with support for three 16x links - Two to connect the GPUs and one to connect both cores to the system's PCI Express interface.
|
