AMD’s Richland APUs: Setting the stage for the arrival of Steamroller

As we reported earlier today, AMD has just officially launched its new Richland series of desktop APU parts. If you’re familiar with the Trinity-based hardware that AMD launched a year ago, you’ve already got an idea of how these chips stack up against their Intel counterparts.

The new processors are debuting as the AMD “6000” family, and each core is more-or-less a straight refresh of the part that came before it, with slightly higher clock speeds.

The A10-5800K adds support for DDR3-2133, which may help nudge its graphics performance a bit higher, but by and large, this is a straightforward refresh.

Richland still moves the ball forward for AMD in several important ways. Firstly, power consumption tests show that the new A10-6800K uses only about 4.5 per cent more power at peak load than the A10-5800K, despite being clocked 8 per cent higher at stock. That’s a significant gain for AMD, particularly given that the chips we’re testing are sitting at the top of its product stack. The gains here suggest that Richland yields have improved over the past 12 months across the board.

Another positive development – for top-end hardware that’s already pushing clock speed hard – is that there’s more headroom in the A10-6800K than existed in the A10-5800K, as shown below:

This is rather different to what we saw when the A10-5800K first debuted. We re-tested that chip, just to confirm it – our CPU wasn’t stable above 4.4GHz without massive voltage increases. The chip’s effective maximum speed, in other words, was just 200MHz, or 4.5 per cent, above its Turbo Core. The A10-6800K doesn’t blow the doors off the room, but we were able to run all four cores at 5GHz on our single sample – 16 per cent faster than the chip’s typical quad-core Turbo frequency of 4.3GHz.

These improvements help explain why AMD has raised its prices slightly with the introduction of the 6000 series; the A10-5800K debuted at $129 (£83), while the A10-6800K is coming in at $149 (£96). Is this enough to fundamentally change the competitive situation with Intel? In a word, no. Here’s the same Cinebench 11.5 comparison, only with Intel’s Core i5-3550 – a $199 (£129) chip – and the just-launched Haswell Core i7-4770K thrown into the mix.

Richland’s higher frequencies and better overclocking potential don’t change the current status quo. But they do give us hope that AMD’s next-generation Steamroller is coming along smoothly. New roadmaps from Sunnyvale, quietly updated last week, imply that the third iteration of the Bulldozer architecture could drop quite soon – by the beginning of the fourth quarter, in fact.

Kaveri is the Steamroller-based, GCN-equipped APU that will hopefully kick AMD’s single-thread performance out of the doldrums. In its Hot Chips presentation last year, AMD CTO Mark Papermaster showed off a number of design changes being baked into the new core. Steamroller has two decoders capable of dispatching four instructions each rather than a single four-issue decoder serving its integer pipelines. The two decoders work in parallel rather than a staggered formation, and either can dispatch code to the FPU.

AMD is forecasting a 30 per cent increase in decode operations per cycle, a 5-10 per cent improvement in scheduling efficiency, and improved branch prediction. Some of these benefits are going to primarily impact the server market, others will favour desktop, but the bottom line is the same: Steamroller should offer a significant improvement over Piledriver. There’s been a great deal of discussion over whether or not this purported photo of the new die is actually legitimate (click to enlarge the image):

If it is, it shows a Steamroller design that’s doubled up in a number of key areas compared to Piledriver. The L1 caches are larger, the registers are wider (this could mean that AMD has gone for 256-bit registers early rather than waiting for Excavator). There’s debate, in fact, over whether this is Excavator or Steamroller, but either way, it points to significant improvements coming in the not-too-distant future.

The Richland – Steamroller Express

There’s a reason we’ve segued into Steamroller from the Richland launch. The A10-6800K is proof that AMD has improved yields and improved on the base Trinity design. The move from 32nm SOI (Bulldozer, Piledriver, and Trinity) to 28nm bulk silicon (Steamroller) shouldn’t have necessitated a major revamp, given that AMD is building Kaveri at GlobalFoundries, and GF used gate-first for 28nm.

Richland’s strong showing relative to where Bulldozer started from in 2011, in other words, implies that Steamroller is coming along swimmingly. Obviously that could change. Roadmaps have slipped before, and I was surprised to hear AMD claiming that Steamroller would show up in the back half of this year. Historically, AMD has had a hard time hitting its performance targets the first time – but its follow-ups have often been markedly better.

Faint praise? Maybe. But I think Richland is more exciting for what it says about AMD’s mid-term roadmap than for what it offers right now, today. We already know Kaveri will include a new GCN-based GPU. It’ll also be the first conventional x86 CPU with HSA support, and it should have a wide range of clock frequencies and TDP targets.

Should you buy Richland today? It’s not a bad offer for a reasonably priced CPU+GPU, and the overclocking could be fun. No, it’s not going to be the best performing solution, even when overclocked, but we’re glad to see the technological advances baked into this core. With Kabini already shipping and Kaveri on the way, the next two quarters could be some of the strongest AMD has had in years.