Frequency matters
Hitting NVIDIA where it Hertz?
There's always something oh-so nice about being the best at a particular task. You know, I want to play golf like Rory McIlroy - let's forget the last few tournaments, shall we? - I want to be better at chess than Viswanathan Anand, and I want to play football like Andrés Iniesta. Sadly, for me, these pipe dreams are just that. Coming back onto more familiar ground, the huge intellectual property requirements for continual graphics processing unit (GPU) development means that only two companies, NVIDIA and AMD, can reasonably claim to be the best in the business.
According to our findings, AMD claimed the 'world's best gaming GPU' when it came out with the Radeon HD 7970 just before Christmas 2011. NVIDIA rolled up its sleeves and, precisely three months' later, presented the GeForce GTX 680. Drawing on one-too-many sporting analogies, we reckon that the GTX 680 - a refined, power-efficient GPU designed almost exclusively for PC gaming - wins on points.
Fast forward to today and the £350 Radeon HD 7970 is a little slower in our benchmarks than the £400 GTX 680. AMD points out that the 'Tahiti' card's memory-bandwidth and size advantage translates to better performance once plugged into three screens, for multi-monitor gaming, but our analysis begs to differ. AMD and NVIDIA's GPUs have historically polarised opinion to such an extent that technology aficionados ardently support one side or the other, and to be seen to have the fastest GPU is more than a mere bragging right; it's reason enough to convince potential purchasers to choose sides and lay down their hard-earned on any red Radeon or green GeForce.
AMD can still make a tenuous case for the Radeon HD 7970 to be called the fastest GPU around. Carefully-selected benchmarks, disseminated by the Austin outfit, attempt to inveigle the press into popularising the head Radeon as the supreme GPU, but fervent price-cutting in recent weeks would indicate otherwise. AMD would like to sell the cheapest Radeon HD 7970 for £400-plus; NVIDIA's GeForce GTX 680 stands as the obvious impediment. What's needed, then, is for another Radeon GPU to take the fight to NVIDIA and win the hearts, minds and, most importantly of all, wallets of the well-heeled enthusiast.
Supercharged Radeon HD 7970: the GHz Edition
Manufacturing genuinely high-end GPUs is a case of striking an uneasy balance between frequency and yield. Increase the default clocks and fewer GPUs from a given wafer will pass the certification process; lower the clocks and while a greater proportion of GPUs can be used, the retail cards will need to be cheaper, as performance is lower. This is why pre-overclocked HD 7970 cards can attract a significant premium - one needs to screen for the best-yielding GPUs and use better-than-reference cooling.
To this end, AMD's partners currently retail factory-overclocked Radeon HD 7970s. We've taken a look at Sapphire's monster, PowerColor's beastie, and Gigabyte's powerhouse. The trio boost the GPU clock from 925MHz to at least 1,000MHz and, depending upon card, inch up the memory from an effective 5,500MHz to, say, 5,700MHz. The lean overclocking indirectly confirms that the Tahiti GPU doesn't have a whole heap of headroom, right?
Now, some six months on from the first release of AMD's finest graphics processor, silicon yields have improved enough for AMD to start shipping higher-clocked Tahiti parts to its roster of partners. Engineers have found ways to streamline the process to tease out extra frequency without negatively impacting on yield and voltage characteristics. This isn't a re-spin of the silicon, per se, but general improvements made possible by the passage of time. The sum of their efforts is a new GPU called the Radeon HD 7970 GHz Edition. Let's roll out the Trusty Table™ and reveal all.
| GPU | Radeon HD 7970 GHz Ed. (3,072MB) |
Radeon HD 7970 (3,072MB) |
Radeon HD 7950 (3,072MB) |
GeForce GTX 680 (2,048MB) |
GeForce GTX 670 (2,048MB) |
|---|---|---|---|---|---|
| DX API | 11.1 | 11.1 | 11.1 | 11.1 | 11.1 |
| Process | 28nm | 28nm | 28nm | 28nm | 28nm |
| Transistors | 4.3bn | 4.3bn | 4.3bn | 3.54bn | 3.54bn |
| Die Size | 352mm² | 352mm² | 352mm² | 294mm² | 294mm² |
| Processors | 2,048 | 2,048 | 1,792 | 1,536 | 1,344 |
| Texture Units | 128 | 128 | 112 | 128 | 112 |
| ROP Units | 32 | 32 | 32 | 32 | 32 |
| GPU Clock (MHz) | 1,000 (1,050) | 925 | 800 | 1,006 (1,058) | 915 (980) |
| Shader Clock (MHz) | 1,000 (1,050) | 925 | 800 | 1,006 (1,058) | 915 (980) |
| GFLOPS | 4,096 | 3,789 | 2,867 | 3,090 | 2,459 |
| Memory Clock (MHz) | 6,000 | 5,500 | 5,000 | 6,008 | 6,008 |
| Memory Bus (bits) | 384 | 384 | 384 | 256 | 256 |
| Max bandwidth (GB/s) | 288 | 264 | 240 | 192.3 | 192.3 |
| Power Connectors | 8+6 | 8+6 | 6+6 | 6+6 | 6+6 |
| TDP (watts) | 250 | 250 | 200 | 195 | 170 |
| GFLOPS per watt | 16.38 | 15.15 | 14.34 | 15.84 | 14.46 |
| Release MSRP | $499 | $549 | $449 | $499 | $399 |
We're one, but we're not (exactly) the same
Let's be clear at this juncture: the Radeon HD 7970 GHz Edition and Radeon HD 7970 are almost identical. They share the same GCN Compute underpinnings, architecture, silicon, UVD video trickery and are pin-for-pin compatible, meaning that it's straightforward for a partner to switch to the 'newer' GPU without any board redesign.
Refinements in the manufacturing process, as alluded to above, enable AMD to increase the default GPU clock to 1,000MHz, up from 925MHz, and the memory to 6,000MHz, up from 5,500MHz. GDDR5 memory manufacturers, too, have been refining their processes, and the move to 6Gbps memory ICs, first implemented by NVIDIA, is both sensible and needed. The upshot of such a move is that the GHz Edition, equipped with the same 3GB framebuffer as on the basic card, now enjoys nine per cent extra bandwidth, or, putting it in a way AMD would prefer, a 50 per cent uptick compared to the GTX 680. The higher memory speed will play well at ultra-high resolutions and/or with multi-monitor gaming.
We'll have some of that GPU Boost malarkey, too
Notice the brackets for the GPU clock? The HD 7970 GHz Edition implements a performance-boosting feature somewhat similar to NVIDIA's GPU Boost. Called AMD PowerTune Technology with Boost (PTWB) it's an evolution of the clock-setting tech already present in high-end 7-series GPUs.
This AMD-supplied slide gives a high-level overview of how it works. The middle stack is representative of how the regular HD 7970 performs. AMD is able to qualify that GPU at a 925MHz core because behind-the-scenes monitoring ensures that, when required, the GPU is maximising in-game frequency; it's running close to the wire with respect to frequency potential. In a nutshell, PowerTune provides a means of increasing baseline (guaranteed) performance without having to spend significant time qualifying every board, as has been the case pre-7-series GPUs.
The HD 7970 GHz Edition, meanwhile, uses a Boost P-State. Here, the GPU takes control of the PowerTune settings and dynamically increases frequency and voltage if there's scope to do so, all the while still keeping within the board's 250W TDP. The better-yielding nature of the newer GPUs enables such performance-boosting behaviour to take place without breaching the specified board power. One can, of course, override the TDP by adjusting sliders in the PowerTune control panel.
As shipped on the reference card, this Boost has only one state: +50MHz. The GPU jumps from 1,000MHz to 1,050MHz if temperatures and board power permit. Interestingly, AMD takes a swipe at NVIDIA's GPU Boost by stating that every single Radeon HD 7970 GHz Edition's Boost is predictable; you may know that two GTX 680 cards, for example, may GPU Boost in different ways, leading to 'Golden Sample' press cards' performance that isn't replicated in retail cards. AMD uses what it calls Digital Temperature Estimation (DTE), first seen in mobile Trinity APU parts, to accurately determine temperature based on input power, not on sample-specific qualities, per GTX 680. The long and short of it is that every GHz Edition card will boost in precisely the same way, according to AMD.
So how can AMD implement this new feature onto a die that's the same as incumbent Radeon HD 7970? The answer is that PTWB is a combination of firmware and driver upgrades. There is no technological reason why it can't be made available non-GHz HD 7970s, but AMD may well choose not to do so in order to differentiate this product. C'mon, AMD, have a heart and give folk who already own an HD 7970 some PTWB love, will ya?
Isn't this just an overclocked HD 7970, then?
Yup, in the main, the new GHz Edition is just an AMD frequency-boosted version of the card that many of you know all about. Advancements in process manufacturing and subtle tweaks give rise to higher clocks, paving the way for add-in board partners to indulge in their own qualifying and release super-fast special-edition cards, say, 1,200MHz core and 6,500MHz memory.
AMD doesn't need to, or cannot, reinvent the GPU wheel just six months after the HD 7970's launch. What it can do, and patently has done, is pull more performance from an existing design. Given that regular Radeon HD 7970 cards are available for £350 one would suspect the GHz Edition to chime in at £400, or very much on GeForce GTX 680 money.
We'll examine the reference card and then plough straight through to the benchmarks. Question is, can a faster-clocked HD 7970 give the GTX 680 a good lickin' and kickin'?