4.3bn is the magic number
Ho, ho, ho
The AMD Santa has a present to give to all you well-behaved enthusiasts. Christmas is coming early this year as AMD announces its next-generation graphics architecture. Ladies and gentlemen and boys and girls, say hello to the Radeon HD 7970 3GB graphics card. We're giving you a first-look at the architecture and benchmarks while AMD gets its act together and officially launches the card on January 9, 2012.
A brand-new GPU... honest
AMD and NVIDIA have in recent history been guilty of launching new series of GPUs that largely resemble what's currently available. Heck, why not call a Radeon HD 5770 an HD 67x0, or a GeForce 210 a GeForce 310? No-one is going to notice, right? Well, we do, and a jump in nomenclature should be accompanied by a genuinely new GPU. Thankfully, AMD has designed the Radeon HD 7970, codenamed Tahiti and headlining the 'Southern Islands' architecture, from the grounds-up, and it's all rather exciting.
One card now, more to follow
The best way to describe what AMD has done with Tahiti is to roll out the vital specs and then show how the underlying architecture has been manipulated to affect such a change.
| GPU | Radeon HD 7970 (3,072MB) |
Radeon HD 6970 (2,048MB) |
Radeon HD 6990 (4,096MB) |
|---|---|---|---|
| Codename | Tahiti | Cayman | Antilles |
| DX API | 11.1 | 11 | 11 |
| Architecture | GCN | VLIW | VLIW |
| Process | 28nm | 40nm | 40nm |
| Transistors | 4.3bn | 2.64bn | 5.28bn |
| Die Size | 365mm² | 389mm² | 2 x 389mm² |
| Processors | 2,048 | 1,536 | 3,072 |
| Compute units | 32 | 24 | 48 |
| Texture Units | 128 | 96 | 192 |
| ROP Units | 32 | 32 | 64 |
| GPU Clock (MHz) | 925 | 880 | 830 |
| Shader Clock (MHz) | 925 | 880 | 830 |
| GFLOPS | 3,789 | 2,703 | 5,099 |
| Memory Clock (MHz) | 5,500 | 5,500 | 5,000 |
| Memory Bus (bits) | 384 | 256 | 2 x 256 |
| Max bandwidth (GB/s) | 264 | 176 | 2 x 160 |
| Power Connectors | 8+6 | 8+6 | 8+8 |
| TDP (watts) | 250 | 250 | 375 |
| GFLOPS per watt | 15.15 | 10.81 | 13.6 |
| CrossFire Support | 4-way | 4-way | 2-way |
| Release MSRP | $549 | $369 | $699 |
Let's sift through the table and compare the single-GPU Radeon HD 7970 directly against the Radeon HD 6970, which is the fastest single-GPU offering in AMD's 2011 line-up. We'll go down the table and highlight the salient points, and make a few observations against the twin-gun Radeon HD 6990 along the way, too.
DX11.1
A jump in the graphics API usually enables developers to speed-up rendering by using hardware-based shortcuts made possible by advances in Microsoft's gaming version of DirectX. DX11.1, due to ship with Windows 8, is a relatively minor update that supports v1.2 of Windows Device Driver Model. From what little is known right now, amongst other programming features, DX11.1 features Microsoft-backed stereoscopic 3D support - a blow to NVIDIA, perhaps? - and larger buffers for shaders, meaning more code can be moved away from the CPU to the GPU for efficient processing.
Of course, the new API requires games developers take advantage of it, highlighting that its worth is likely to be felt many months after launch. Still, it's only right that AMD supports a forward-looking DirectX update from Microsoft.
28nm process
GPU boffins tend to throw a mix of more transistors and cleverer design for brand-new launches. AMD wanted to shift down to a 32nm process for the Radeon HD 6-series GPU, but was stymied by problems at manufacturing partner TSMC. Skipping a generation and going straight down to 28nm carries implicit risk, yet AMD's content to launch its flagship single-GPU card with these itty-bitty transistors.
Shifting down to 28nm enables AMD to fit in a whopping 4.3bn transistors into a die size that's a touch smaller than present-generation Cayman's. Make no mistake, Radeon HD 7970 is a gargantuan GPU that's only feasible with a 28nm process. And it is this relatively low-power process that enables AMD to ramp up frequencies and pure GFLOPS ratings while keeping noisome power in check. The one takeaway from this section is that the new Radeon card is, quite easily, the most transistor-heavy consumer GPU ever made. And transistor-heavy usually means fast.
Frequencies, memory and power
Leading on from the above, AMD crams in 2,048 processors, 128 texture-units and 32 ROPs into the Tahiti behemoth. Frequencies are handsome, with the core clocked in at 925MHz and AMD intimating there's much more to come. Such top-end power needs to be fed with lots and lots of juicy memory bandwidth.
Rather than introduce a new memory standard and therefore add even more complexity into the design, AMD runs with tried-and-trusted GDDR5. This time around, however, the backend has been boosted by a further two 64-bit memory partitions, now totalling 384-bit, and providing an extra 50 per cent peak bandwidth.
Put simply, there is a lot more of everything when compared to the Cayman-derived HD 6970, which, let's remember, is no slouch in its own right. In fact, an argument could be made that the topology is rather NVIDIA GeForce GTX 580-like: a crazy number of transistors allied to an 'expensive' 384-bit memory bus. The similarities don't end there, folks, so let's now drill down into what makes Tahiti tick.
