Page 2 - AMD 3870 Architecture
Manufacturer: ATI
Series: Radeon HD 3000 Series
GPU: RV670
Interface: PCI-E 2.0 x16
Core Clock (Stock): 775 MHz
Shader Clock (Stock): 775 MHz
Memory Clock (Stock): 1125 MHz (2250 DDR)
Memory Bandwidth: 72 GB/sec
FLOPS: 496 GFLOPS
Pixel Fill Rate: 12400 MPixels/sec
Texture Fill Rate: 12400 MTexels/sec
Framebuffer: 512 MB
Memory Type: GDDR4
Memory Bus Type: 64x4 (256 bit)
DirectX Compliance: 10.1
OpenGL Compliance: 2.0
PS/VS Version: 4.1/4.1
Process: 55 nm
Shader Processors: 320
Pipeline Layout: Super-scalar MADDx5
Texture Units: 16
Raster Operators: 16
Data above courtesy of GPUReview.com.
As far as naming scheme goes, AMD/ATI took a completely different approach than NVIDIA. The whole NVIDIA 8800GTS/8800GT/8800GTS (G92) naming scheme is enough to confuse most users -- basically, we are saying that, to less informed individuals, it may seem illogical that an 8800GTS 512MB is better than a 8800GTS 640MB. Of course, we understand that the architecture of G92 and G80 is a completely different story both on paper and in performance, but hey -- it sure doesn't look like that on the box.
Back to the AMD style of naming, all suffixes are dropped as we've been so used to the whole SE/vanilla/GT/Pro/XL/XT and even XT Platinum Edition in the past couple generations. The new naming scheme completes the relative performance between cards with the first number indicating generation, second number is family, and the last two indicating the variant. This means that the 3870 we are look at today is a third generation (The "3"), family "8", as almost universal of a higher midrange card from both NVIDIA and ATI, and the "70" basically saying that it is an "XT" card. This means that a Radeon HD 3850 is really a "Pro" variant; so in classic naming scheme the card we are looking at today is essentially a '3800XT' and a 3850 is really a '3800 Pro'. Interestingly, however, the pricing of the 3870 puts it in the spot of the previous generation X1950 Pro -- when its predecessor is actually the X1950XT.
Under the hood is the new 55nm AMD RV670 GPU core. The die shrink, like the G92, is not a simple fabrication process change -- the die shrink added new features (But also took some away, we'll discuss that later) as well as fixed inefficiency issues faced by the ill-fated ATI Radeon HD 2900XT. The R600 in the original 2900XT based on the 80nm architecture has a die size of 400 square millimeters with 700 million transistors; while the RV670 core has a die size of only 192 square millimeters housing 666 million transistors. This devil is going back to haunt NVIDIA. Muhahahahahahahahaha.
Lame pun aside, what it really means is that it uses nearly half as much power and costs roughly half as much as the Radeon HD 2900 while attempting to deliver similar performance in the benchmarks. With a 105W peak power draw according to specifications, it only requires a single 6-pin PCIe power connector. ATI PowerPlay is also introduced as they've ported over from their mobile segment -- similar to Intel SpeedStep, clocks and voltages will be reduced automatically to conserve power. Much unlike Intel SpeedStep, however, it is independent of the processor state -- so programs that take advantage of GPU calculation power will no longer be restricted to 2D clock speeds only.
Being that, the RV670's basic root architecture still remains somewhat the same as the R600 -- with new features, obviously; but with a deal of bandwidth removed. As we can see from our detailed GPU information list above, the RV670 used on the Radeon HD 3870 has its external bandwidth reduced from 512 bit to a 256 bit interface. Interestingly, AMD still retained eight memory channels, which means that each channel's original 64 bit interface is modified to become 32 bit each.
In terms of internal ring bus, the 1024 bit interface is now 512 bit -- this significant reduction of bandwidth internally and externally is reflected in the drop of transistors of roughly 40 million. This is accounted for even with the introduction of Universal Video Decoder hardware and ATI Powerplay; as well as optimization of the hardware in terms of DirectX 10.1 compatibility.
As far as the Universal Video Decoder (UVD) goes, it is the widely implemented AMD technology that mostly does the dirty work for video decoding -- which supports, obviously, decoding, deinterlacting, and post processing during video playback. VC-1 and H.264 decoding is natively supported. DX10.1 is not widely implemented yet, but it's nice to see its implementation into these cards. Personally, it's probably not going to be the deciding factor in buying a card just for DX10.1 support, but it's nice to see its inclusion.
Remaining unchanged between the R600 and RV670 with its DirectX 10.1 unified shader core are 320 stream processors, 16 texture units, 16 render back-ends, and a programmable tessellator unit.
As tradition goes, the Asus TOP variant of the ATI Radeon HD 3870 is one of the fastest factory overclocked variants of graphics cards based on this core. It almost tops in terms of clock speed; and memory is close to being the fastest available in the market -- only a few other manufacturers have faster factory overclocked memory out of the box. The engine is clocked aggressively at 851MHz and memory at 2.28GHz effective (DDR4). We'll put the Asus EAH3870 TOP to our tests in just a moment.
Page Index
1. Introduction, Specifications, Bundle
2. AMD 3870 Architecture
3. A Closer Look, Test System
4. Benchmark: FEAR
5. Benchmark: Prey
6. Benchmark: Half Life 2: Lost Coast
7. Benchmark: CS:Source HDR
8. Benchmark: 3DMark06
9. Power Usage, Overclocking
10. Noise Factor and Conclusion
