Oh dear. Our review of the Radeon HD 5700 series is late. AMD keeps pumping out Radeons at such a ridiculous rate, they broke me. I just couldn't keep up with it allran out of snark and iffy jokes, or something like that.
Just a few short weeks ago, the Radeon HD 5870 debuted, the first of a new generation of graphics processors that packs improved performance, new capabilities needed to support DirectX 11 and emerging GPU-computing standards, higher image quality, improved power efficiency, and multiple other marketing bullet points of varying worth. We liked it, to put things mildly. My wife keeps trying to convince me not to keep mine under my pillow at night, but I know better. Oh, yes.
New generations of GPU technology tend to make their way from higher end products into more affordable ones over time. This trend is often called the waterfall effect, because, well, I dunno. Water falls down, and stuff. Spray goes everywhere. It's noisy, and sometimes there are rainbows.
Where was I?
Oh, right. Anyhow, the technology built into the Radeon HD 5870 is propagating to a second GPU in record time. AMD has just introduced a pair of graphics cards based on a new mid-range graphics processor code-named Juniper, bringing DirectX 11 and a whole slew of other buzzwords into price ranges heretofore reserved for older buzzwords followed by lower numbers. The bottom line for you and me, roughly speaking, is better graphics cards for less money. Let's see how well the new Radeons deliver on that expectation.
Juniper takes a, erm, bough
Yep, the GPU behind the Radeon HD 5700 series is code-named Juniper, part of the "Evergreen" naming scheme AMD is using for these chips. The big daddy in the family, Cypress, is the chip inside the Radeon HD 5800 series. Juniper is classic example of a chip company firing up its design tools, including the vaunted World's Smallest Chainsaw, and essentially chopping its high-end part in half to serve a broader market.
In truth, it's not quite that simple, because Cypress and Juniper were essentially co-developed. The Juniper silicon came back from the fab about a week after Cypress, and AMD had both chips in validation at the same time. If a bug was found on one chip, AMD would then attempt to replicate it on the other one and, if needed, apply a fix to both. That co-development strategy is what allowed the firm to crank out multiple new Radeons in an unprecedently tight time window.
Still, if you look at the block diagram above, you can see what I'm saying about that tiny chainsaw, a tool Canadians are renowned for using well. Architecturally, in many key respects, Juniper has half the resources of Cypress. That reality is reflected first in the graphics engine at the top of the diagram, which includes only a single rasterizer, not two like Cypress.
In its shader core, Juniper has 10 SIMD cores, each of which has 16 superscalar execution units. Each of those has five ALUs. Multiply it all out, and Juniper is sporting a total of 800 ALUs, or "stream processors," as AMD likes to call them, a bit immodestly. Each SIMD core has a texture unit associated with it, so Juniper includes 10 of those, giving it 40 texels per clock of texture filtering power.
Also halved on Juniper is the number of render back ends, which now stands at four. As a result, the chip can produce 16 color pixels per clock and 64 Z/stencil pixels per clock. Again, that's half of Cypress' capacities.
Interestingly enough, Juniper turns out to be quite similar to the RV770 GPU that powers the prior-gen high-end Radeon HD 4800 series. Both Juniper and RV770 have 800 SPs, 10 texture units, four render back-ends, and a single rasterizer. But Juniper is a smaller chip intended for less expensive graphics cards, so it has only two 64-bit memory controllers or an aggregate 128-bit memory interface. Both Cypress and RV770 have 256-bit memory interfaces. Still, the use of higher-clocked GDDR5 memory in the Radeon HD 5700 series will at least help bridge the deficit versus the RV770.
Across the nascent Evergreen lineup, AMD has increased processing and graphics power disproportionately compared to memory bandwidth. That trend may largely be driven by cost considerations and the bandwidth per pin limitations of available memory types. Still, AMD claims the tradeoff makes sense, asserting that the RV770 actually had more memory bandwidth than it needed and that chips like Juniper are simply more balanced, not bandwidth-starved.
Of course, Juniper varies from the RV770 in many other respects. The chip's hardware supports many capabilities exposed in new software APIs like DirectX 11, DirectCompute, and OpenCL 1.0. AMD has improved its image quality via superior texture filtering methods, as well, and revamped its display output block to support up to six four-megapixel displays simultaneously. I suggest you read our Radeon HD 5870 review for a full rundown on these features. The many incremental improvements are considerable, when taken together.
Juniper has dropped one major feature from Cypress, though: the ability to process double-precision floating-point math. Both Cypress and RV770 can process double-precision datatypes at one-fifth the rate they do single-precision numbers, with varying degrees of capability and internal precision. Cypress is more or less fully IEEE 754-2008 compliant. But if that string of letters and numbers doesn't mean anything to you, you probably won't miss DP support in Juniper. The ability to handle DP math is crucial for certain GPU computing markets, but its value for a consumer product is shaky. DP math doesn't matter a whit for real-time graphics, for one thing; its only real use is for GPU computing. Even for GPU computing, the first consumer applications aren't likely to need double-precision datatypes. Things like image processing, video compression and effects, and even physics simulations for gaming get along just fine with integer or single-precision FP datatypes. Omitting DP support reduces the size of the chip and thus cuts AMD's costs, which is why the company chose to leave it out of this very consumer-focused GPU.