Another Tahiti GPU cometh
AMD hurried the press into launching Radeon HD 7970 reviews just before Christmas 2011, but the company has been moving at a more leisurely pace for the release of the Radeon HD 7950.
Fortunately for us, the second-rung 'Tahiti' card shares heck of a lot with its bigger brother, the HD 7970, so feel free to trawl through the launch-day review to understand just what a GCN is and how these new 7-series GPUs put equal store in compute and graphics.
For those that just want to know speeds and feeds and how the Radeon HD 7950 stacks up against its peers, let's bring forth the updated specification table:
| GPU | Radeon HD 7970 (3,072MB) |
Radeon HD 7950 (3,072MB) |
Radeon HD 6970 (2,048MB) |
Radeon HD 6990 (4,096MB) |
|---|---|---|---|---|
| Codename | Tahiti XT | Tahiti Pro | Cayman XT | Antilles |
| DX API | 11.1 | 11.1 | 11 | 11 |
| Architecture | GCN | GCN | VLIW | VLIW |
| Process | 28nm | 28nm | 40nm | 40nm |
| Transistors | 4.3bn | 4.3bn | 2.64bn | 5.28bn |
| Die Size | 352mm² | 352mm² | 389mm² | 2 x 389mm² |
| Processors | 2,048 | 1,792 | 1,536 | 3,072 |
| Compute units | 32 | 28 | 24 | 48 |
| Texture Units | 128 | 112 | 96 | 192 |
| ROP Units | 32 | 32 | 32 | 64 |
| GPU Clock (MHz) | 925 | 800 | 880 | 830 |
| Shader Clock (MHz) | 925 | 800 | 880 | 830 |
| GFLOPS | 3,789 | 2,867 | 2,703 | 5,099 |
| Memory Clock (MHz) | 5,500 | 5,000 | 5,500 | 5,000 |
| Memory Bus (bits) | 384 | 384 | 256 | 2 x 256 |
| Max bandwidth (GB/s) | 264 | 240 | 176 | 2 x 160 |
| Power Connectors | 8+6 | 6+6 | 8+6 | 8+8 |
| TDP (watts) | 250 | 200 | 250 | 375 |
| GFLOPS per watt | 15.15 | 14.34 | 10.81 | 13.6 |
| CrossFire Support | 4-way | 4-way | 4-way | 2-way |
| Release MSRP | $549 | $449 | $369 | $699 |
Same architecture, different focus
Go down the two left-hand columns and it should be apparent that Radeon HD 7950 is a slightly cut-down version of the chief card. While based on the same 4.3bn-transistor die, 28nm process and GCN architecture, the HD 7950 is segmented by having fewer compute units - 28 vs. 32 - and, consequently, fewer processors.
And because the texture units are all linked into compute units, which is now a common approach, there are commensurately fewer units, too. AMD leaves the backend unchanged, that is, 32 ROPs and a 384-bit pathway to GDDR5 memory, and this is a sensible move given just how much oomph is packed at the front-end of the GPU. Actually, we were a little surprised to see AMD use just 32 ROPs on the daddy HD 7970.
But a snip here and there on the architecture front probably ain't gonna be enough segmentation, especially if pesky partners get their way and pre-overclock the Tahiti Pro core to way beyond base speeds. This is why AMD is looking for some model-to-model insurance by dropping the core speed to 800MHz and memory to an effective 5,000MHz.
Carving out small parts of the architecture and reducing frequencies is a by-the-playbook method that's been used countless times before. One benefit of less power is shown by a lower TDP - 200W vs. 250W - and, in a physical sense, by the use of two six-pin PCIe connectors instead of an 8+6 arrangement.
Yup, the Radeon HD 7950 contains enough architecture hamstringing to make it a second-rung GPU. The obvious benefit of a slower card is a lower street price, and we believe that the initial batch of Radeon HD 7950 3GB cards will retail for $449 (£350) or so. This 20 per cent-plus reduction should correlate well with an equivalent drop in performance, assuming AMD's got its modelling right.
Now, we know what you're thinking. Enthusiasts are asking themselves whether the second BIOS on the Radeon HD 7950 can be used to flash the card into a full-fat HD 7970 by activating these missing compute units and raising clocks. It was successfully done on select HD 6950 cards that, once flashed, turned into HD 6970s. The answer in this case is no, because AMD is hardware-disabling these HD 7950 cards such that they cannot be BIOS-upgraded to a full complement of 32 compute units. Boo-hoo, we say.
And there was one more
AMD's not finished there, folks, as there will be another version of the HD 7950 board that cuts price further. These high-end monsters use a 3GB frame-buffer, configured through 12 2Gbit memory devices. A simple way to reduce the cost is to use less RAM. AMD and its partners will also release a 1.5GB part, dropping the density to cheaper 1Gbit chips, much like how the current NVIDIA GeForce GTX 580's memory topology is arranged. The suggested price for this version is $419 (£325) and may well be a good bet for gamers who play at a 1,920x1,080 resolution, where the framebuffer load is much lower than, say, at 2,560x1,600 or, going further, at an Eyefinity3 5,760x1,080.
A little slower, a lot cheaper
So there you have it. Stating the bloomin' obvious, the Radeon HD 7950 is a little slower and quite a lot cheaper than the HD 7970, for the reasons explained above.
