AMD Radeon HD 7750 & HD 7770 Launch Review



Product: AMD Radeon HD 7700 Series
Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: February 14th, 2012

Cape Verde Architecture

The Cape Verde design is based on AMD's new Graphics Core Next architecture, first seen in AMD's ultra-enthusiast Tahiti GPU (Radeon HD 7900). The configuration here is 640 stream processors arranged in 10 compute units. There are two 64-bit memory controllers for a combined 128-bit memory bus, with 40 texture units and 16 ROPs. In Tahiti we saw one RBE per four CUs, but that math doesn't work for Cape Verde - there's a remainder left over dividing 10 by 4; leaving the obvious configuration as 2 render back-ends, matching the memory controllers.

Cape Verde's die size is a tiny 123mm2, down from AMD's Juniper 166mm2 core size and half the size of Tahiti. This in turn drops other key metrics like power and thus TDP rating. AMD has two products based on Cape Verde, created by selectively disabling parts of the chip. The Cape Verde XT configuration has the full specification, while Cape Verde Pro is scaled back to hit AMD's targeted performance and power limits. This allows Cape Verde to compare favorably with Juniper, with XT power dropping from 108W to 80W and Pro from 86W to 55W. Where both the Juniper variants required a 75W 6-pin PCIe power connector, only Cape Verde XT does now - and only just at that.

Cape Verde Specifications

The core engine clock hits 1GHz in the Cape Verde XT specification, something AMD is branding as GHz Edition and is claiming as an industry first. AMD supports this claim with their assertion that the hot-clocked shaders of their competitors GPU aren't the majority of the chip, and so most of the chip doesn't run at that over-GHz speed, whereas most of the Cape Verde XT chip does. Hmmm, let's not dwell on that too much then, despite the fact that the shaders are the important bit ... and just enjoy the way ATI got back on the branding ;-)

AMD AT 1GHz Edition

AMD gives the number of transistors as 1.5Bn, up nearly 50% over Juniper's 1.04Bn. Courtesy of the architecture change, Juniper's 800 VLIW5 Stream Processors in 10 SIMD engines is now 640 GCN Stream Processors in 10 Compute Units, and interestingly compute performance is down - Juniper's 1.36TFLOPS isn't matched by Cape Verde, which clocks in at 1.28TFLOPS. However, gaming performance appears to increase with Cape Verde's pixel fillrate given as 16.0 Gigapixels/sec, and texture fillrate at 40.0Gigatexels/sec. This is a nice jump up from Juniper's 13.6 GPixel/s and 34.0GTexel/s. This increase is thanks to Cape Verde's higher clocks - 1GHz trumps Juniper's 850MHz. The fact that it does this in 80W TDP instead of 108W TDP is due to the new process and the more efficient GCN architecture. Oh, one other thing - Cape Verde supports double precision floating point natively in the hardware. Who's got less compute performance?

Cape Verde Architecture

Memory bandwidth is down from Juniper, with Cape Verde using 1125MHz memory clocks. This is 75MHz down from Juniper XT (1200MHz/4.8Gbps/76.8GB/s), and 25MHz down from Juniper Pro (1150MHz/4.6Gbps/73.6GB/s). Interestingly, both Cape Verde products share the same memory clocks, using an effective 4.5Gbps rate and offering 72GB/s bandwidth.

The specifications indicate the HD 7770 should outperform the HD 5770 / 6770 and come close to the HD 6790; Cape Verde XT has more pixel and texture fillrate than Barts LE, but the 6790 has double the memory bandwidth. The HD 7750 seems to be splitting the difference between the HD 5750 / 6750 and the HD 5770 / 6770. If you're only considering the performance aspect of Cape Verde, this seems like a fail, but that's missing a major factor: power and features.

GCN brings several important optimizations to gamers, including the improved RBEs with upgraded AF and a newly improved 9th generation tessellation unit. Both of these will offer significant upgrades in IQ and performance from Juniper-based cards, and a decent upgrade from Barts based hardware, too. GCN also adds the capability to run Enhanced Quality AA and SSAA/AAA in DirectX 10 and 11 - features both Juniper and Barts lack.

As part of the true Southern Islands family, Cape Verde gets all the compute, media, and power saving technology introduced with Tahiti. AMD's PowerTune technology is implemented along with ZeroCore Power, which is a significant feature set in this segment. With both hard power caps and truly low idle power consumption, Cape Verde is very attractive for a number of use cases.

For media consumption use, it features the UVD 3 engine used in many of the HD Radeon 6000 series products, adding hardware acceleration for DIVX playback and BluRay 3D streams. The Audio capabilities are upgraded; now featuring positional audio with per display audio streams that follow applications as they move from display to display (assuming HDMI or DP audio capable displays are used). AMD's AVIVO branding is now deprecated and replace with Picture Perfect HD, but the Color and Quality controls remain available, along with AMD Steady Video if you have an AMD platform. These combined controls offer unparalleled enhancement and improvement for all kinds of media, including Internet streamed.

The new Video Codec Engine adds faster than real-time encoding to H.264, and GCN architecture features new instructions for improving video processing, too. The output of UVD or a video stream can be piped into the VCE for media codec transcoding either on the fly for webcam video transmission or saved to disk. Video can be processed both post-UVD and as part of VCE using hybrid mode operation to add effects or improve video quality.

A very impressive part of GCN architecture is its power saving features. Zero Core Power is a mobile technology now implemented in the desktop realm, leveraging Bus Alive Chip Off technology to power down as much of the GPU as possible while appearing to the system as being still there and functional. Whenever the OS enters long idle state for the display, the GPU powers off and delivers instant and significant power savings. Where the Juniper and Barts competitor cards use 18-20W at idle, Cape Verde (and its bigger brothers Pitcairn and Tahiti) drop to sub 3W in long idle. GCN's complementing feature is PowerTune, which uses on-chip workload analysis to adjust core clock speed to keep in boundaries.

Currently, AMD sets Powertune limits higher than nearly all workloads will commonly consume. Unlike PowerPlay states which set GPU core voltage, clock and memory clock for specific operations (like UVD, or Eyefinity desktop), PowerTune operates inside a tenth of a second and adjusts clock speed very finely.