AMD has been busy of late as it has started to roll out its Radeon HD 5000 'Evergreen' series of graphics cards. We've already looked in depth at the architecture behind the new Evergreen chips - particularly focusing on Cypress - and we've evaluated the gaming experience delivered by the Radeon HD 5870. For reasons completely out of our control, we haven't said an awful lot about the Radeon HD 5850 yet, but the good news is that there'll finally be a full bit-tech review of that online tomorrow. Today, AMD is launching the ATI Radeon HD 5700 series with two new graphics cards that sit below the Radeon HD 5850 and 5870 in the performance mainstream segment. We've been told to expect the Radeon HD 5770 to sell for around £120, while the cheaper Radeon HD 5750 will likely sell for about £100.  Juniper's Architecture Flow Diagram Both of the new GPUs are based on the Juniper core, which is effectively half of Cypress with a few bits lopped off based on the differences in transistor count (1.04 billion vs. 2.15 billion). What isn't clear, is whether Juniper is the Turkish or Greek half of Cypress, though. (Groan - Ed.) The Radeon HD 5770 uses Juniper in its full form with all 10 SIMD cores enabled, clocked at 850MHz - the same frequency as the Radeon HD 5870's engine. Each of these SIMD cores has 16 five-way superscalar shader processors, four texture units, a 32KB local data store and an 8KB L1 texture cache, resulting in a shader core not too dissimilar to the Radeon HD 4870's (RV770). There are some changes inside, as the shader units learned some new tricks in the Radeon HD 5000 series - we discussed these at great length during our Radeon HD 5870 Architecture Analysis. The global data store has remained unchanged and still stands at 64KB - AMD still doesn't talk about the bandwidth between local and global data stores. The Graphics Engine hasn't remained exactly the same, as the dual rasteriser and hierarchical-Z units are no longer required - Juniper is half the size of Cypress and it doesn't suffer from the same bottlenecks in this area. The back end of the chip has also been cut in half, as there are only two 64-bit memory controllers, each with two ROP units that can each handle four colour writes or 16 Z/stencil writes per clock. The Functionally, L2 caches haven't changed from Cypress and there are just half as many of them (two instead of four) - they're still 128KB each and the theoretical peak bandwidth between L1 and L2 still stands at 435GB/sec. To continue the similarities with the 5870, the 5770 shares the same memory clock too at 1,200MHz (4,800MHz effective), which means it has 76.8GB per second of bandwidth because of its 128-bit memory interface, meaning it has precisely half the memory bandwidth of AMD's current flagship part. The beans are spilled on the final specifications of AMD's "Juniper" mid-range GPU architecture, and the two SKUs based on it, namely the ATI Radeon HD 5770 and Radeon HD 5750. Leaked alleged company slides on XtremeSystems Forums, reveal exactly how Juniper is a 50% downsizing of Cypress' machinery, in having half of nearly all its vital-statistics, if you may. Juniper features 10 SIMD engines physically, amounting to 800 stream processors. It holds 40 texture memory units (TMUs), and 16 raster operation units (ROPs). With a 128-bit memory interface churning out up to 76.8 GB/s of memory bandwidth by utilizing 4.8 GT/s GDDR5 memory chips, the fact that Juniper is half of what Cypress stands for specs-wise, becomes clear. In the next two slides are pictured and detailed the two SKUs based on the GPU. The ATI Radeon HD 5770 is the top-end implementation, which makes use of all the components present on the GPU. With a core clock-speed of 850 MHz, and memory speed of 1200 MHz, the GPU churns out a shader compute power of 1.36 TFLOPs. The idle and max board powers are rated by AMD at 18W and 108W, respectively. It becomes clear that AMD is eying the US $150 price point with this part. The Radeon HD 5750 on the other hand, has one SIMD engine disabled, and eds up with 720 available stream processors. It comes with 1 GB or 512 MB of GDDR5 memory, and clock speeds of 700 MHz core, and 1150 MHz memory (4.6 GT/s). With these, the Radeon HD 5750 manages a shader compute power figure of 1.008 TFLOPs. The rated board power consumption figures have also come down on this one, with 16W idle, and 86W max. This one targets the $109 (512 MB) and $129 (1 GB) price points. Both accelerators are DirectX 11 compliant, and support ATI Eyefinity technology supporting up to three 2560 x 1600 pixel display heads. According to the slides, they should be unveiled this Tuesday the 13th. |
