At APU13 today, AMD announced a full suite of new products and development tools as part of its push to improve HSA development. One of the most significant announcements to come out the sessions today-- albeit in a tacit, indirect fashion, is that Kaveri is going to pack a full 512 GPU cores.
There's not much new to see on the CPU side of things -- like Richland/Trinity, Steamroller is a pair of CPU modules with two cores per module. AMD also isn't talking about clock speeds yet, but the estimated 862 GFLOPS that the company is claiming for Kaveri points to GPU clock speeds between 700 - 800MHz depending on whether or not the CPU's performance is considered within that 862 GFLOP figure. With 512 cores, Kaveri picks up a 33% boost over its predecessors in terms of core count, but memory bandwidth will be essential for the GPU to reach peak performance.
AMD has also confirmed that Kaveri will support both the Mantle API and its new TrueAudio technology, which means these technologies will be anchored at the high-end with the Radeon R9 290/290X and the low-end with the Radeon R7 260X and Kaveri.
AMD's Kaveri GPU Will Pack 512 GCN-Based GPU Cores
To showcase its performance, AMD showed Kaveri up against the Haswell-based Intel Core i7-4770K paired to a low-end GeForce GT 630 discrete GPU. There are three variants of this particular card -- it's not clear which AMD utilized -- but the overall performance figures are rather stark. In the intro scene to BF4's single-player campaign (1920x1080, Medium Details), the AMD Kaveri system (with no discrete GPU, using only on-die graphics) consistently pushed frame rates in the 28-40 FPS range. The Intel system, in contrast, couldn't manage 15 FPS. Performance on that system was solidly in the 12-14 FPS range -- meaning AMD is pulling 2x the frame rate, if not more.
This, of course, is just one game and one configuration, but AMD is clearly feeling confident about what Kaveri can offer on the desktop. And keep in mind, performance in this game is only going to get better once Mantle support is added in a patch next month.
Other new features include the launch of AMD APP 2.9, a new development SDK that adds a Visual Studio plugin for authoring OpenCL, supports the CMake utility, and includes code samples for OpenCV and Bolt. AMD is also launching its own Media SDK (appropriately titled Media SDK 1.0), which is allows developers to use AMD's APUs for various forms of pre-and-post process filtering and a specific library for encoding low latency video.
AMD is also announcing CodeXL 1.3, which is significant for its addition of Java support. GPU-accelerated Java has been a talking point of the HSA initiative for several months, so actually pushing that option into standard tool sets is a major achievement for the company.
Sunnyvale is expected to unveil more details about Kaveri as the event progresses, so we'll have updates on the new chip later today or tomorrow after the various keynotes. The broad push and emphasis of today's event is to position APUs as the natural evolution of PCs. That's rather different than Intel's approach, which has continued to emphasize the CPU as the center of the development environment, but with GPU support and capability ramping up rapidly across generations. AMD's approach emphasizes the HSA Foundation and broad coalitions of developers using OpenCL.What we see over the next few days should give us a much better feel for just how effective the multi-pronged HSA Foundation will be at challenging companies like Intel.