OpenCL performance of next-gen AMD mid-range GPUs hit the intertubes

Saturn, or Bonaire for that matter…

Radeon AMD will be releasing Radeon HD 8000 series (Sea Islands family for desktop and Solar family for notebook) later this year, along side the original Radeon HD 7000 series. The OpenCL performance of the mid-range GPU in both these lines appeared on this CLBenchmark results page.

CLBenchmark is a benchmark for testing OpenCL performance on each device supporting OpenCL 1.1, and includes several tests of different aspects of OpenCL programming techniques and applications. Included are SPH Fluid Simulation, Raytracing, Optical Flow, Sobel Image Filter, Separated Gaussian Blur Image Filter, and Median 3×3 Image Filter along with some programming principles such as Bucketing, Reduction, Scanning, Summing, Bitonic Merge Sorting, and Tree Searching. Each item will be given a score, reflected as the relative performance of the device for comparison with other OpenCL-capable devices.

Two sets of results were posted, the tested GPUs shared the same codename, Saturn, and all of them are based-off the same ASIC codenamed Bonaire. The codenames suggested that the GPU is part of the Solar notebook GPU family, and the ASIC is part of the Sea Islands (C.I.) GPU family, so now you know the codename of its desktop counterpart. The GPUs have two different device IDs, 6640 and 6641 respectively. [Author’s note: for ease of reading, the two GPUs will be referenced as Saturn 6640 and Saturn 6641.], the Saturn 6640 uses a beta version of existing production OpenCL drivers (Version 1124.2), while Saturn 6641 uses a newer pre-production driver (Version 1169.1), both GPUs were tested using Windows OS.

The included OpenCL device information shows the Saturn GPU is having 12 compute units running at a fixed frequency of 1075 MHz, they are provided with 2 GB of graphics memory on-board and OpenCL 1.2 is supported on the GPU. With the GCN architecture, each OpenCL compute unit means exactly a compute unit of 64 shaders and 4 TMUs, that means the Saturn GPU has a total of 768 shaders and 48 TMUs, running at 1075 MHz. [Author’s note: However, do note that this is supposed to be a notebook GPU meaning tighter limits on thermal and power envelope, while the drivers used to produce the CLBenchmark results don’t implement any power saving technologies such as PowerTune. Thus the fixed frequency instead of multiple values for the CL_DEVICE_MAX_CLOCK_FREQUENCY parameter as shown for Radeon HD 7970 GHz edition, so the above specification isn’t necessarily equivalent to the final hardware specification found on the next generation entertainment notebooks.]

With a bit of data gathering, we could compare the results of Saturn 6640 and Saturn 6641 with Radeon HD 7770 GHz Edition (Cape Verde XT GPU with 640 SPUs, 40 TMUs) and Radeon HD 6770 (Juniper XT GPU with 800 SPUs, 40 TMUs). First off we have the comparison with Radeon HD 7770 GHz Edition:

Saturn6640_1Saturn 6640 compared with Radeon HD 7770 Ghz Edition, showing a 10% advantage

Saturnvs_SaturnSaturn6641_2The two Saturn GPUs compared to each other, showing each of the drivers has its own advantages over the other

With the extra frequency and compute units, the new GPUs show 8% to 10% better OpenCL performance over a Radeon HD 7770 GHz Edition with early drivers. However it’s worth noting that the newer OpenCL driver for Saturn 6641 has clearly a different approach to optimizing compute workloads, it shines on areas where memory usage is being highlighted and regressed on the rest. And next up we have similar comparison to the older Radeon HD 6770 with 1 GB graphics memory:

Saturn_3Since the GCN architecture puts heavy focus on graphics compute performance, and drastically improved the per unit efficiency, we are seeing an average 2.9x OpenCL performance for Saturn GPUs over a Radeon HD 6670. That would translate to a great value for those who wants decent compute performance from an affordable upgrade of graphics card over an older platform from few years ago.

Of course, all the Saturn results we are seeing today are based off pre-production OpenCL drivers, there will be many more optimizations to come. We will hopefully see the score go up further when the GPUs are closer and closer to launch. At least that’s what we believe will happen… S|A

The following two tabs change content below.

Leo Yim

Author at SemiAccurate
Leo Yim is our correspondent from the far flung reaches of East Asia. Fluent in Mandarin, Cantonese and English he'd rather be talking about computers no matter what language. A true detail man he dreams of building gaming rigs from workstation class parts.