Today we’ll be comparing the performance of AMD’s first and only desktop Carrizo part, the Athlon X4 845, to its Godavari predecessor. The goal here is the see how the performance profile of AMD’s chips has improved from generation to generation and set the stage for AMD’s upcoming APU launches. It’s worth pointing out that Carrizo was designed as a mobile-first product. To that end it’s helped AMD regain some marketshare in laptops, but as Anandtech showed OEMs still don’t respect AMD.
In recent years the power management capabilities of modern processors have grown by leaps and bounds. Traditionally when different architectures are compared reviewers like myself will construct two systems that share almost all of their hardware components like SSDs, power supplies, and graphics cards. Where things get tricky is when you attempt to compare processors that don’t share the same socket, memory type, or manufacturer. In this article we’re comparing chips that luckily share all of those elements. But that’s where the similarity of these two chips come to an end. At a silicon level we’re comparing to very different products. They have different CPU cores, GPU cores, memory controllers, caches layouts, and key to this particular discussion: different power management schemes.
Back when power management was simple we would disable the power saving and turbo states in BIOS and then match each chip’s memory and core clock speeds. This would allow us to understand the performance a given chip offered in a specific benchmark at a specific clock rate. I called this metric PPC or performance per clock. They are other names for this metric but for consistency and clarity we’ve always used PPC. With Kaveri and Godavari we could still more or less do these kinds of PPC comparisons. But with our Carrizo-based Athlon X4 845 things have changed. The primary issue is that our 845 is not a clockspeed multiplier unlocked K-edition chip. This means that clocking it to a specific frequency is complicated especially if we want to reach a frequency that is above 3.5Ghz.
Traditionally we’ve compared CPU cores at 4 Ghz partly because AMD’s best APUs are usually clocked at about that frequency and partly because we like round numbers. We won’t be doing that in this comparison.
But we tried to. Our X4 845 sample didn’t seem to take all that kindly to BLCK increases so we disabled all of the power management and TDP caps in BIOS but left the turbo mode enabled so that under load it would stick to its highest turbo multiplier, 38, and then bumped the BLCK to 106 Mhz to hit a shade over 4 Ghz. Then we ran our benchmarking suite. Performance in this overclocked state was significantly lower than what we saw at stock clockspeeds. For example our Speedometer benchmark scored about 40 runs per minute at stock speeds. We were getting 32 runs per minute at 4 Ghz or a 20 percent performance hit for what was ostensibly a 12 percent clockspeed boost. To be clear these results do not make any sense inside of our old performance per clock testing methodology.
More to the point they shouldn’t make sense. Meddling with power management schemes as a user is border line masochistic at this point.
Thinking more generally our PPC metric is myopic because while modern chips are designed around balancing goals for frequency, architecture, thermal envelope, die space, and power consumption; PPC only considers frequency and architecture. To meaningfully make generational comparisons between chips we need to design our testing methodology in a way that explicitly or implicit considers all of those aspects.
Given that frequency has become such a squishy variable how else can we can attempt to create an even playing field that doesn’t hide the advantages of aggressive power management, good uncore design, and feature-level advancements from generation to generation. I believe that comparing chips with a cap on resource consumption is the best way to recognize all of these improvements. More specifically comparing chips with equal TDPs and the same cooling solution is the best way to understand generational advancements. Comparing chips with the same resources both in terms of power consumption and cooling, in addition to using the same non-core components like SSDs and power supplies is the way forward for making meaningful comparison between chips.
All of this is a long way of saying that we’re going to be pitting the 65 Watt TDP Athlon X4 845 up against the A10-7860K which is currently AMD’s best 65 Watt quad-core chip. With matched peak power consumption profiles we can also put them on equal thermal footing by strapping AMD’s fancy Wraith cooler on both chips and kitting them out with 16 GBs of DDR3-2133 RAM. To be clear, we expect the Carrizo-based X4 845 to match or win in our benchmarks. The questions we aim to answer are where does it win and by how much?
What we see here is an interesting split between multithreaded and single threaded performance. Carrizo comes out clearly ahead in multithreaded performance. In some cases like H.264 encoding and wPrime 1024M it’s 26 to 35 percent faster. But in other cases like PeaZip and particularly WinRAR it’s no different or slower than Godavari. We were a bit concerned by these results so we reran these offending benchmarks a number of times. The results were the same. But we did notice something rather odd: Carrizo’s clocks dropped into the low 2 Ghz range when we ran our WinRAR benchmark. Why the X4 845’s power management was throttling it under these conditions we don’t know.
Looking at single threaded performance we see a slight regression compared to Godavari most of which is due to a low score in our WinRAR benchmark. Ignoring that result single threaded performance is essentially the same. Ignoring both of our WinRAR benchmarks puts Carrizo ahead of Godavari by 6 percent overall and 14 percent in multithreaded performance. In sum and excluding WinRAR the only version of Carrizo to fit into socket FM2+ is between -1 and 35 percent faster than a Godavari chip with the same cooler and TDP.
It’s also worth noting that the Athlon X4 845 doesn’t appear to be Carrizo’s finest bin. Given that it’s already had its GPU portion and FCH controller disabled and apparently wasn’t power efficient enough to go into a laptop it’s found its way onto the desktop for a mere $70. In fact in all honesty the X4 845 is probably a dumping ground of a SKU where AMD can try and monetize the piles of defective Carrizo dies it has sitting out back with the spare Llano chips it had to write down.
On the other hand it’s quite impressive how well the X4 845 does given its handicaps. More to the point matching Godavari in single threaded performance and winning by 10 percent in multithread performance is a good result for a chip that has 100 Mhz slower base clock and 200 Mhz slower turbo clock and half the L2 cache.
More than anything else playing with the X4 845 has made me pumped up for the upcoming Bristol Ridge launch. It’s exciting to consider how moving the chipset off the motherboard and switching to DDR4 could improve the performance profile of AMD’s APUs.S|A
Latest posts by Thomas Ryan (see all)
- AMD Launches its Radeon Instinct Server Accelerators - Jun 22, 2017
- A Look at AMD’s EPYC Line-up - Jun 21, 2017
- AMD Details EPYC’s Power Efficiency - Jun 20, 2017
- Intel’s Skylake-X and Kabylake-X Meet Our Scale - Jun 19, 2017
- MIPS for Self-Driving Cars with the I6500-F CPU - Jun 19, 2017