Radeon HD 6970 And 6950 Arrive
In the world of high-end hardware, waiting for the next great thing means reading about a lot of technology and never actually buying a new piece of gear. That’s no way to enjoy your favorite games though, is it?
Sometimes the wait is worth it, though. Last week was a perfect example. Nvidia launched its GeForce GTX 570, but we already knew AMD’s Radeon HD 6970 and 6950 boards had shipped out, FedEx Next Day Air. How could we not recommend waiting to see how these boards would compare? The very same morning, AMD launched a refresh on its Phenom II and Athlon II processors, incrementing clock rate. But we have Intel’s Sandy Bridge-based CPUs up and running in the lab. Again—why wouldn’t you wait a couple of weeks to see how they stack up?
Anticipating Greatness
I was particularly interested in the performance of the Radeon HD 6970 and 6950, the two cards AMD had shipped so urgently. Whereas the Barts GPUs used to build Radeon HD 6870 and 6850 centered on the same VLIW5 architecture that earned Radeon HD 5870 a place in infamy, the Cayman GPU consolidates functionality into a VLIW4 design, incorporating fewer ALUs per thread processor, but improving performance per square millimeter of die space. How would that translate to the experience you get when you plug a 6900-series card into your system? Well, since Cayman is larger than Cypress, we have to assume it’s also going to be faster.
A market populated by GF100-based boards like GeForce GTX 480 and GeForce GTX 470 would have made this launch so much easier for AMD. But Nvidia has this habit of kicking into gear when it really needs to. The resulting GeForce GTX 580 and 570 demonstrate that, even if Fermi remains a power-hungry architecture, it is possible to contend with more heat using a better sink and finely-tuned fan control. Suddenly, AMD had a much more daunting competitive landscape in front of it.
We’re not sure if that was the real reason AMD’s Radeon HD 6900-series cards were delayed for nearly a month. However, the boards are here now (two of each, actually), and AMD claims that channel availability will be plentiful right out of the gate.
New GPU, Familiar Face
AMD’s GPU team isn’t known for its bold, brash architectural moves—that honor belongs to Nvidia—so it’s hardly a surprise that much of Cayman looks familiar. In fact, there was no fancy press day hosted on an aircraft carrier to herald the improvements, nor were there conversations with chip architects. After the fanfare ahead of Radeon HD 6800, today's 6900-series launch was preceded by surprisingly little commotion. It was almost…refreshing. That gave us plenty of time to dig into the details for ourselves. Fortunately, we know enough about AMD's Cayman GPU to discuss what changes.
Fully loaded, the chip features up to 24 SIMD engines (up from Cypress’ 20). Each SIMD still includes 16 thread processors. Except now, each thread processor consists of four ALUs rather than five. So, while Cypress sported as many as 1600 ALUs, Cayman incorporates up to 1536 (that’s 24 * 16 * 4). Each SIMD engine is still tied to four texture units, totaling 96 on a complete Cayman chip (versus 80 on Cypress).
| Header Cell - Column 0 | AMD Radeon HD 6970 | AMD Radeon HD 6950 | AMD Radeon HD 5870 | Nvidia GeForce GTX 580 |
|---|---|---|---|---|
| Manufacturing Process | 40 nm TSMC | 40 nm TSMC | 40 nm TSMC | 40 nm TSMC |
| Die Size | 389 mm² | 389 mm² | 334 mm² | 520 mm² |
| Transistors | 2.64 billion | 2.64 billion | 2.15 billion | 3 billion |
| Engine Clock | 880 MHz | 800 MHz | 850 MHz | 772 MHz |
| Stream Processors / CUDA Cores | 1536 | 1408 | 1600 | 512 |
| Compute Performance | 2.7 TFLOPS | 2.25 TFLOPS | 2.7 TFLOPS | 1.58 TFLOPS |
| Texture Units | 96 | 88 | 80 | 64 |
| Texture Fillrate | 84.5 Gtex/s | 70.4 Gtex/s | 68 Gtex/s | 49.4 Gtex/s |
| ROPs | 32 | 32 | 32 | 48 |
| Pixel Fillrate | 28.2 Gpix/s | 25.6 Gpix/s | 27.2 Gpix/s | 37.1 Gpix/s |
| Frame Buffer | 2 GB GDDR5 | 2 GB GDDR5 | 1 GB GDDR5 | 1.5 GB GDDR5 |
| Memory Clock | 1375 MHz | 1250 MHz | 1200 MHz | 1002 MHz |
| Memory Bandwidth | 176 GB/s (256-bit) | 160 GB/s (256-bit) | 153.6 GB/s (256-bit) | 192 GB/s (384-bit) |
| Maximum Board Power | 250 W | 200 W | 188 W | 244 W |
As with the Cypress and Barts GPUs, Cayman is a product of TSMC’s now-mature 40 nm manufacturing process. It probably wouldn’t have been, but TSMC canceled its 32 nm node back in 2009, leaving both AMD and Nvidia to rethink their strategies. The aforementioned specs allowed AMD to keep its die size manageable, while still improving performance, though. Cayman is a 389 mm² piece of silicon composed of 2.64 billion transistors, while Cypress was 334 mm² part made up of 2.15 billion transistors. What AMD didn’t want to do was follow in Nvidia’s footsteps, creating a 500+ mm² behemoth that it’d need to power and then cool. From all angles, Cayman looks to be a compromise based on the hand TSMC dealt.
It seems that the company was able to turn those lemons into something more palatable, though, by getting inventive with a feature called PowerTune, which balances TDP to facilitate higher shipping clock speeds. The two resulting board models, Radeon HD 6970 and 6950 handily outperform AMD’s last at-bat on the 40 nm process. The real question is, can they stack up to the more recent and renewed competition from Nvidia?
Ten points, by the way, if you can figure out the reference in this story’s title.
