AMD finally launches the Kaveri APU, its most important chip launch in years

This launch has been a long time coming. AMD took the wraps off its Kaveri APU today, officially putting the chip on pre-order — for delivery in the next few weeks — and launching a few benchmark results to demonstrate its performance. We’ve written a great deal about Kaveri in the past, so I’ll mostly recap today. This is AMD’s most important chip launch in years; Kaveri is the first CPU to truly fulfill the Fusion vision AMD sketched out as its justification for buying ATI nearly eight years ago.

Kaveri is the first core with full HSA support, the first PC processor to integrate a GCN variant on-die, and the first true architectural update for Bulldozer since that core was released more than two years ago. AMD’s previous Piledriver update fixed a few of the core’s more notable flaws and improved clock speeds, but didn’t address the architectural shortcomings. Kaveri’s Steamroller CPU core has promised to dramatically improve IPC (AMD’s own documentation implies a 20% IPC uplift in certain workloads).

The new A10-7850K is coming in at the clock speed and TDP targets we expected, with a base clock of 3.7GHz, a Turbo Mode of 4GHz, and a 95W TDP. That’s slightly behind the older Richland, which clocked in at 4.1GHz base, 4.4GHz Turbo. The GPU’s clock speed is also lower — the A10-7850K will clock its GPU at up to 720MHz, compared to 844MHz for Richland. Where Richland was fabricated on GlobalFoundries’ 32nm node, Kaveri is 28nm bulk silicon.

A kit from AMD actually arrived today with appropriate testing hardware in it, but we thought we’d take a look at the figures AMD is pushing for the new APU, as shown below.

The first set of three tests here are standard PC tests. They show the AMD core outperforming Richland modestly in PCMark 8, substantially in 3DMark, and by a huge margin in Basemark CL, an OpenCL-based test. This implies that the CPU multithreaded gains from Kaveri are significantly offset by Richland’s higher clock speeds, but that the GCN-based GPU has picked up a more substantial performance margin.

As for the much-discussed performance advantages of HSA? AMD is offering test results from a single program to illustrate those, as shown below:

Performance in LibreOffice Calc really has increased enormously.

Now, an 8x speedup in application performance is nothing to sneeze at, but our understanding of HSA indicates that this is something application developers will have to go back and specifically target. Just having an application that runs in OpenCL, for example, isn’t enough — it needs to be updated to target the APU’s HSA capability. Given AMD’s small market share, driving widespread HSA adoption is going to be difficult. Moreover, while Kaveri is the first HSA part, it also appears to be the only chip with that capability coming to market in 2014. None of the HSA Foundation’s partners with mobile CPUs seem to be in any great hurry to integrate HSA capability into shipping designs.

At this point, we think Mantle is more likely to drive APU sales for AMD than HSA throughout 2014, and Mantle’s own sales-improving potential will depend entirely on how much additional game performance it can deliver. Clearly AMD had hoped to have BF4′s Mantle version shipping before Kaveri went out the door, but the bugs and problems in BF4 have prevented that. Still, we should have our first Mantle game ready in the near future, not long after launch.

AMD is advertising Kaveri as available for pre-order with multiple OEMs, with expected delivery in the next few weeks. Will the chip live up to the expectations placed on it? We’ll be able to tell you in the not-so-distant future.