Cooking with Chorizo. (Overclocking Carrizo cores in the Athlon x4 845.)

(No images or videos yet, sorry. This was basically my morning project.)

For those who have not heard, or have not been paying attention, AMD decided to release some Excavator cores to the market. They are basically cannibalized/salvaged cores that are still good, but the rest of the chip failed to make the binning process, so the integrated GPU and Southbridge are disabled, and the die is slapped onto the appropriate PGA format. The goods news is that IPC has been improved, along with some of the cache issues that plagued the Bulldozer family of processors. The bad news is a limit of x8 PCIe lanes, and 2MB of L2 Cache. (Also, potentially a slightly gimped memory controller, but more on that later.)

When the Athlon X4 845 was first released in late February, I snagged one up, perhaps too hastily. Although it arrived by March 10th, it was just last week that Gigabyte finally released their (beta) BIOS update that allowed for Carrizo support for my particular motherboard, the F2A88X-UP4. Some of their other products, like the newly released FM2+ boards with USB 3.1 support, already had updates, along with their G1.Sniper board, but, for whatever reason, the UP4 was put on the backburner. Most annoying! (For those that care, ASUS has allowed for Carrizo support for all their boards for roughly a month, and ASRock followed suit a couple weeks ago. MSI, like Gigabyte, is still mixed.)

Anyway, let's get to the meat and potatoes.

Test set-up:
Athlon x4 845
Gigabyte F2A88X-UP4
G.skill DDR3 2400 Trident X
Gigabyte R9 285

Thermalright Silver Arrow IB-E. (Why? Because I could.)

Meat and Potatoes

My tool of choice was IntelBurnTest, and from this program I'm going to share some numbers (from the standard test). Not an end-all-be-all metric, but it's what I have to offer, currently.

At stock speeds, with its 3.8ghz boost, the Athlon x4 845 managed to hash out 22.5 GFlops, which matched my previous CPU, the Athlon x4 760K, when it was overclocked to 4.7ghz.

With a 108 base clock, I obtained a ~3.8ghz base clock, with a ~4.1ghz boost clock, which hashed out roughly 24.5 GFlops. This required 1.462V on the Vcore, and 1.2 volts to the NB. At 110mhz on the baseclock, I managed to boot, but could not successfully pass any tests with IBT, regardless of the voltage I crammed into the componentry - all the way up to 1.55 volts.

Side Notes and Stream of Consciousness

Overclocking on the FM2+ platform, without an unlocked multiplier, is a giant pain in the dick. All of the I/O of the Fusion Controller Hub (FCH) is attached to the Base Clock, with no way of separating it. This means that only a small boost is possible before your rig starts throwing fits at you, because the clock rate is affecting how the USB, SATA, and PCIe operate. SATA is the most finicky, and those who have the patience can try switching to IDE Mode (as opposed to AHCI) to have some more leeway.
Furthermore, overclocking with AMD on any of their newer CPUs that were released after their Piledriver/Trinity line-up has its own set of problems. Since they have more telemetry devices installed that attempt to provide a more smooth-grained approach for performance optimization and thermals, you can get some wonky throttling results when you still have plenty of leeway on both the Core and Socket temperatures. These weird results are mostly regulated to heavy stress testing programs, like Prime95 and IBT, and light day-to-day tasks usually don't have this effect.

To confound matters, the 845 refused to boot on either of the G.skill RAM's overclocked profiles. I had to dial back the frequency to 2133, from 2400. Apparently, others have run into this issue, and it has been theorized that the cause is because the product is essentially a salvaged mobile component with a lesser memory controller.

While there are other (more indepth and professional) reviews that show the small, but measurable, affects that having access to only x8 PCIe lanes and less L2 cache, the perceived difference is negligible. Even after I run through a test suite of gaming and benchmarks, I don't think I'll notice much of a difference, and the effort will be spend mostly in the pursuit of bigger, higher numbers.


Would be interested in seeing clock for clock comparison of the 845 and the 760 if you are interested.

1 Like

Awesome read, nice to see someone else on FM2 haha.
Looking forward to some delicous pics (pun intended).

1 Like

Seems like if they put out a fully featured Carrizo chip it'd probably do pretty well at least well in terms of AMD chips, though I think they were saving that for AM4

1 Like

This might happen sooner than expected, but it really depends on how quick Gigabyte is to put out a non-beta BIOS.

But, also, I finished playing some Fallout4. You know how some people complain how performance tanks in the cities? For whatever reason, my CPU frequencies are dialing back to roughly 3.0ghz. Doesn't seem to matter how much each of the cores are being utilized, it just seems to happen when I am in the cities. (CPU Core and Socket temperatures do not broach 40C.) I know the Fallout4 engine is bloody awful, but I don't recall my previous CPUs choking up like this.

Anyway, I'd also like to do a short little comparison video of this new heatsink they shipped with the CPU. I have some old AMD heatsinks readily available. Before that, I still have to tinker with this benchmarking program and the Tek Syndicate mouse.

EDIT - @1920.1080p.1280.720p - Some Russian site beat me to it, with the inclusion of Steamroller cores as well. Most of the benchmarks were while running them at 3.0ghz each. The gaming benchmarks, however, did not have any restrictions.

Some 3Dmark results, Skydiver and Firestrike:



Lastly, here is a complete Firestrike comparison with another rig, sporting an FX-8320 OC'ed to 4.0ghz:

1 Like

Hopefully, later tonight I will have the benchmarks of my 760K.

Even since I noticed that weird issue with Fallout 4 and the CPU declocking itself for no apparent reason, I've been tinkering. It does not throttle during any stress tests, only certain games. Most notably, Fallout 4, Planetary Annihilation, and Company of Heroes 2. I thought this might be a GPU driver issue, but I've done clean installs with a couple sets of drivers, to no avail.
I have posted about this issue on the AMD forums, and the replies I received suggested that it may have something to do with the beta BIOS - which I kind of figured, but I wanted to make sure that it wasn't because of some hardware power-saving feature that decided to work against the CPU's own performance.

I'm currently awaiting reply from Gigabyte about this interesting little issue for their input, and, otherwise, I'm waiting for the release of a non-beta BIOS.

I came here expecting a bed of CPUs with sausages being cooked on them. Great info btw!

1 Like

If that CPU "bug" persists, even after a proper BIOS, I might just do that. :P

1 Like

The following comparisons take some explanation.

There are three columns: the one of the left is the 760K, while the other two are the 845. The right-most column of the 845 is a test run with a base overclock of 3.8ghz without any Turbo Boost, and, for whatever reason, it has a higher graphics score than the stock and overclocked tests. However, the stock and overclocked tests show similar results in graphics score, while displaying the difference in phsyics. There is always a margin of error in tests, but I wanted to display all the data on hand to show the consistency.

Stock clocks (3.8ghz) compared:

Overclock (no frequency caps) compared:

To help show the margin of error, here is a third set of data points:
Maximum results - maximum CPU and GPU overclock
The physic score for my 845 shot up - probably because achieving a 4.1ghz overclock required the use of Turbo Boost, which can be inconsistent.

All of this said and done, the x8 PCIe lanes do marginally, but measurably, affect scores. Even in the worst case, it is less than 5%, or a couple frames per second. Computationally, according to these benchmarks, a Carrizo core is roughly 12-15% faster than a Richland core.