ADFS 2011 Wrapup & AMD Lynx Platform Tests



Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: July 18th, 2011

Llano Test Configuration & Setup

Looking at the hardware and architecture so far, it seems pretty clear that the A8-3850 APU with 6550D graphics should perform comparably to an Athlon II X4 640 and an ATI Radeon HD 5570 with DDR3. The APU has the advantage that it integrates the northbridge into the package as well, meaning the mainboard has only to host the fusion controller hub. Thus, on a pure performance/watt basis, the Lynx platform (APU and FCH equipped mainboard) improves significantly over the Leo platform (AM3 processor with 8-series chipset including southbridge).

Northbridge + X4 CPU + dGPU = APU

Whereas APU + FCH weighs in around 110W (100W APU, 8W FCH, plus ~2W for additional chips), the Leo platform runs around 130W, without any graphics; add 5W for additional chips, and 10W for a HD 4290 or match performance with a HD 5550 with DDR3 for another 40W...; yeah, that's a 175W TDP vs. a 110W TDP. This is where the advantage ends, as (for the moment) it's actually cheaper to buy the Leo platform and a 5550 (~$220, vs. $235 for the Lynx platform). AMD knows this, so obviously the comparison is to competitor platforms and performance for price.

Test Platform

Overclocking is done fully by the base clock, which drives all the busses in the system. Unlike Intel's Cougar Point and Sandy Bridge system, you can make significant adjustments to this clock. AMD are not disclosing anything about future products, and it is unknown if there will be unlocking of CPU or Radeon cores, or Black Edition APUs with unlocked multipliers.

Component Specification
Mainboard Gigabyte A75M-UD2H
Processor A8-3850
Graphics Card APU HD 6550D
  Radeon HD 5550 512MB GDDR5
  Radeon HD 6670 1GB GDDR5
  Radeon HD 6770 1GB GDDR5
  Radeon HD 6850 1Gb GDDR5
Memory 2x2GB Mushkin Blackline PC-16000
Audio Realtek ALC889
PSU Zalman ZM600-HP
Case CoolerMaster HAF 922
Storage Corsair F120 + Seagate 7200.11 1TB
Heatsink/Fan Sunbeamtech Core-Contact Freezer
Display Dell P2210H
Operating System Windows 7 32-bit Ultimate SP1
Driver Catalyst 11.6b hotfix

Our Gigabyte board had no problems at 133MHz and only required a little bump in FCH voltage (+0.025v) for 150MHz base clock. There might be a hidden divider automatically adjusting clocks used for the rest of the buses, which is why some overclocking attempts fail around 120MHz. While overclocking, we found that while the BIOS allowed CPU multiplier adjustment upwards, it had no effect. Our final stable overclock was 3525MHz, achieved at 150Mhz bClk with 23.5 multiplier, and a bump in voltage of +0.150v to both the CPU vCore and NB VID. For the GPU, somewhere between 775 and 800MHz was the limit, so we settled with 775MHz. Some A75/55 platforms have the GPU clock linked to the bClk, but this mainboard does not - the clock you set is the clock you get.

Stock A8-3850 APU
Overclocked A8-3850 APU

The standard memory clocks allow for 1866MHz maximum, but when increased to 150MHz bClk the memory can run well beyond, allowing us to hit 2000MHz on our Mushkin Enhanced Blackline DDR3 PC-16000. We were able to run two sticks at the rated 9-10-9-24 timings of these modules, something that previously we had problems doing on Intel's X58 platform (albeit in triple channel). To run 9-10-9-24 2T at 1866MHz with stock CPU we needed to bump NB VID by +0.050v; interestingly this was not required to run 10-10-10-27 1T. Below you can see the memory speeds and timings we used, along with memory bandwidth as measure by Sisoft Sandra.

Speed Timings Bandwidth Change vs. 1333
DDR3-1333 9-9-9-24 1T 12.88GB/s 0
DDR3-1600 9-9-9-27 1T 14.92GB/s +15.8%
DDR3-1866 9-10-9-24 2T 16.03GB/s +24.5%
DDR3-2000 9-10-9-24 2T 17.11GB/s +32.8%

When we first powered on the Gigabyte A75M-UD2H, it was clear that the BIOS wasn't giving correct values for CPU and mainboard temperature. The CPU was reported some 30 degrees(C) below ambient, and the chipset cooler was very hot to the touch but reported at below ambient temperatures also. Updating to the latest beta BIOS changed nothing, obviously there is some more work to be done on the platform reporting side for this mainboard. We have no empirical temperature results to report, but anecdotal results instead. As we wanted to test overclocking, we used an aftermarket heatsink and fan, in the shape of SunbeamTech's Core Contact Freezer. This alleviated concerns about inadequate cooling for the pursuit of clock speeds; if the APU died a fiery death, it wouldn't be from heat dissipation woes. In testing the CCF never became more than slightly warm to the touch, easily dissipating the thermal workload of the APU. The FCH heatsink felt hot, too hot to touch for more than a few seconds, all the time. Adding FCH voltage didn't change its feel (once you get to 'too hot to touch', it's hard to tell) or affect stability in our testing.

Along with overclocking the APU, we used various discrete GPUs to look for platform performance differences. Below you can see the dGPUs used, and their configurations.

  APU HD 6550D Radeon HD 5550 Radeon HD 6670 Radeon HD 6770 Radeon HD 6850
GPU Core 32nm Sumo 40nm Redwood Pro 40nm Turks XT 40nm Juniper XT 40nm Barts Pro
Engine Clock 600MHz 550MHz 800MHz 860MHz 775MHz
Radeon Cores (SIMD) 400 (5) 320 (4) 480 (6) 800 (10) 960 (12)
Texture Units 20 16 24 40 48
ROP's 8 8 8 16 32
Memory Capacity upto 1GB 512MB 1GB 1GB 1GB
Memory Type System DDR3 GDDR5 GDDR5 GDDR5 GDDR5
Memory Bus 64-bit 128-bit 128-bit 128-bit 256-bit
Memory Speed System 1GHz / 4Gbps QDR 1GHz / 4Gbps QDR 1.2GHz / 4.8Gbps QDR 1GHz / 4Gbps QDR
UVD 3 2.2 3 3 3
Outputs DVI/VGA/HDMI 1.4a/DP 1.1 DVI/VGA/HDMI 1.3 DVI/DP 1.2/HDMI 1.4a DVI/DP 1.2/HDMI 1.4a DVI/DP 1.2/HDMI 1.4a
Multi-Display Dual, or DL-DVI Dual Eyefinity 3 Eyefinity 3 Eyefinity 4

Dual Graphics support allows different AMD Radeon HD 6000 series GPUs to be paired with the APU Vision Engine. Supported configurations are AMD Radeon HD 6450, HD 6570 and HD 6670 with the AMD A8-3850's HD 6550D and A8-3650's HD 6530D.

System Manager - Dual Graphics