3DMarks

Overclocking AMD's A8-3850 Llano APU

Add Your Comments

Previously we ran the full review of AMD’s Lynx desktop chip, the A8-3850, which is their Llano architecture for desktops. In that review, regrettably, I was not able to explore overclocking this platform. Thankfully, after a bit more time, that has been remedied.

Overclocking Llano isn’t all sunshine and rainbows, but it’s not too difficult. Very few voltages need adjusting, but it isn’t without its quirks. This is no unlocked multiplier, insanely easy ‘raise multi, raise voltage, repeat ad nauseam until your cooling gives out’ adventure. Rather, it’s a more delicate process that involves a bit of tinkering. In other words – actual overclocking! That’s not to say I don’t enjoy Black Edition or (Intel) K-series CPUs. It’s just different, and tweaking is always fun for an overclocker.

What Happens When You Overclock Llano?

First, let’s start by discussing what you’re doing when you overlcock Llano. To overclock this CPU, you have to use the base clock. There is no other choice. Neither the CPU nor the GPU have unlocked multipliers. Thus, even if your BIOS allows higher multipliers (in this case, it can go up to 47x; the setting will dutifully report in software via Windows too), you can’t overclock that way.

Same goes for the GPU. If your BIOS allows you to manually raise the GPU frequency, ignore it. You can’t do that either.

Which brings us back to the base clock. Unfortunately, as with Intel’s Sandy Bridge, overclocking the base clock can lead to issues. Because the traditional northbridge functions lie on the APU, overclocking the base clock also overclocks everything else. Everything – including storage. Thus, if you boot at the wrong frequency, it might not boot at all, giving you a ‘cannot find boot device’ error because the storage bus is overclocked too far.

Thankfully, that’s not the end of the road. Some boards do it automatically and some do it manually, but most boards (I say most because I have no evidence to say ALL) will give you a storage, et al bus divider. The Gigabyte A75M-UD2H (also featured in the initial A8-3850 review) does it automatically. Regretfully, there isn’t a manual option. Said divider kicks in at 133 MHz bus clock. For overclocking, it would be preferable to have a storage bus lock option.

What about the IGP?

As mentioned before, the IGP is locked down at 6x (stock = 100 MHz base clock x6 = 600 MHz IGP). When you raise the base clock, it raises the IGP frequency x6 (unless you drop that multiplier, but we’re not going there with this article). So, even if you get stuck at the spot where the storage bus divider kicks down, that’s 133 MHz x 6, for a 798 MHz IGP frequency. I can tell you now, most A8-3850s should run that speed, which is fully a third higher than the stock IGP frequency. Told you this was a strong APU, didn’t we?

Memory – What’s Best?

What have we not talked about yet? Memory! System memory and IGP memory on the Llano platform are one in the same. By default, the system dedicates 512 MB of system memory to the IGP. You can manually change that if you’d like, I believe to a maximum 1024 MB.

Sharing memory like this is a mixed bag on a budget platform. Memory timings are very important. AMD told all reviewers not to exchange high frequency for loose timings. When looking for memory for your Llano-based desktop, shoot for the tightest latencies at the highest frequency you can afford. CAS latency 8 is the highest I’d recommend. In the initial review, we used DDR3-1866 with timings of 8-9-8-24. Those timings would be decent for DDR3-1600 too, but a bit tighter would be preferable.

We didn’t explore tight vs. loose timings and we’ll take AMD’s word for that, but we did look at different frequencies at similar timings. The difference in graphics performance was surprising. You’ll see those results below though. First, some items to note when overclocking Llano.

Tips to Make Your Life Easier

So, tip #1 is regarding the storage bus. If you try to overclock your Lynx CPU and it won’t even get five measly bclk past stock 100, try 133 to force your board past its divider breaking point.

Tip #2 and #3 don’t have lengthy explanations. They just, capital W, Work. Tip #2 is that you must use a digital output from your APU. I don’t know why, so don’t ask…it just is.

Lastly, #3 is that you must run your HDD (or SSD in this case) in IDE mode. Running AHCI (and presumably RAID, though I didn’t try that) will give you the ‘no boot disk’ error regardless of base clock, even after the divider changes. Again, no idea why, it just is.

Once you get those three things down, the sky’s the limit. Well, your APU’s the limit. Unless…

Unless you have one of the most annoying BIOS glitches you can think of. You see, evidently (and this is a guess as to its mechanism) the storage bus divider fails to function when you restart the computer. If you’re running, say, 140 MHz base clock and you restart, the system will lose your boot drive. Thankfully there is a solution, however annoying.

In this case, the way to defeat the beast is to restart into BIOS and set the base clock back to 100 MHz. You can leave everything else the same, voltages and all. Boot up into Windows. Then, without logging on even, restart the computer. If it was only when changing something (like pushing the overclock), that would be one thing; but this happens every time you restart and/or shut down the system.

One would assume Gigabyte will address this with a BIOS update, but be forewarned, this may be an issue for anyone that doesn’t leave their system on 24/7 and is definitely an annoyance for anyone testing out maximum overclocks!

Raising the Bar

The results from initial testing of this platform were nothing short of impressive for an IGP. The CPU benchmarks were somewhat unremarkable, but for a Stars core, they did well; especially when you look at the native capability for DDR3-1866 memory.

When pushing this APU and board, first up was the CPU. I raised Vcore 0.1 V from the stock 1.4 V to give 1.5 Vore (idle, vDroop took it down to ~1.475 loaded), which managed a respectable 800 MHz overclock, to 3718 MHz from the stock 2900 MHz. The 26x CPU multiplier was used to bring bclk around the 130-140 MHz mark and the memory multiplier was dropped to keep its speed as close to DDR3-1866 as possible.

3718 MHz LinX Stable

3718 MHz LinX Stable

Unfortunately the IGP didn’t want to go quite that far. At 143 MHz base clock, the IGP was running 858 MHz, which was just a little too high. For the IGP to pass, the base clock had to be dropped two MHz, bringing the final CPU overclock to 3666 MHz and IGP speed to 846 MHz.

3666 MHz - Final Overclock

3666 MHz - Final Overclock

If you didn’t yet, look again at the memory tab in CPUz in the final overclock. That’s right. DDR3-2256, with timings of 8-10-8-24! Simply amazing memory clocking on an AMD chip. This was easier than even Thuban memory overclocking. Whatever they did with the memory controller, let’s hope something with this kind of strength makes it into Bulldozer.

For the curious, here are BIOS shots of the settings used for that overclock. The IGP was left alone, with everything on auto (meaning the default 6x multiplier and default 512 MB memory allocation).

Overclocked BIOS Settings

Overclocked BIOS Settings

Overclocked RAM Timings

Overclocked RAM Timings

We end up with a solid overclock all around. The CPU finished at 3666 MHz with reasonable temperatures on older air cooling (Thermaltake Big Typhoon), a GPU speed of 846 MHz – more than 30% over stock – and memory clocking in at stout DDR3-2256.

Power Consumption

One thing I left out of the initial review was power consumption. At idle, this platform just sips power. The tiniest amount keeps one of these systems going. Loaded, it’s certainly less than I’m used to benching Thuban and Sandy Bridge setups. Anyway, here are the numbers. The CPU was loaded with Prime 95 Small FFTs and the GPU was loaded with Furmark.

Power State Power Consumption
Stock Idle 33 W
Stock Loaded (CPU Only) 131 W
Stock Loaded ( CPU & GPU) 157 W
Overclocked Idle 35 W
Overclocked Loaded (CPU Only) 212 W
Overclocked Loaded (CPU & GPU) 263 W

At stock it doesn’t take much power at all. Overclocked is a different story. Gaining over 100 W with the CPU and GPU overclocked is quite a bit. You’ve got to give something to get something though and overall 263 W isn’t too bad. Comparing that to a higher-end system, an Intel 2600K plus GTX 580 pulls 496 W under load at stock.

Overclocked Performance

When we first looked at the A8-3850, I said the CPU side would perform close to a similarly-clocked Phenom II. The Phenom II x4 965 BE clocks in at 3.4 GHz, so we’ll put that to the test. The Llano APU should be just a smidge better than the 965 BE. The system tested was the same as used previously.

APU AMD A8-3850
Motherboard Gigabyte A75M-UD2H
RAM G.Skill F3-17000CL9D-4GBXL
Disk Drive Patriot Inferno 100G
OS Windows 7 Professional x64
PSU Corsair TX450

For overclocked 2D runs, the APU was clocked at 141 MHz base clock with a 26x multi, for 3666 MHz on the CPU. Memory was using the 6.66x multiplier, clocking in at DDR3-1878 with 8-9-8-24 timings.

2D Benchmarks

Again starting with AMD’s less than favorite benchmark, let’s see how the single core integer performance stacks up overclocked.

SuperPi 1M

SuperPi 1M

SuperPi 32M

SuperPi 32M

We were given healthy increases in both SuperPi benchmarks for sure. Both 1M and 32M gained just under 20%. Not quite a clock percentage for time percentage decrease (the overclock here is 26.4%), but pretty close. If we had clocked the RAM up for these, I’ve no doubt they would be closer to overclock-to-time parity.

WPrime 32M

WPrime 32M

WPrime 1024M

WPrime 1024M

WPrime is always better for the AMD camp. They focus a lot of energy on multi-core performance and it pays. The overclocked Llano performance is on par with what you’d expect from a 965 BE clocked up a couple hundred MHz.

Cinebench R10

Cinebench R10

Cinebench R11.5

Cinebench R11.5

Cinebench results are impressive for sure. Both of these benchmarks reached overclock-to-score parity, which is a good feat. Being the more real-world of our 2D testing, this bodes well for everyday use. You know, if you plan on rendering or encoding with a Llano-based build.

3D Benchmarks

Moving on to 3D, there are two overclocks displayed in these and the game test graphs. The base clock was the same as above, at 141 MHz, which means the IGP was running at 846 MHz using the default 6x multiplier. The two different results were the result of changing the RAM multiplier. Results in the graphs designated “A8-3850 – OC” were using the RAM as above in the 2D benchmarking, at DDR3-1878 / 8-9-8-24. To test out just how much strong RAM would help benchmarking, I also threw in a stronger RAM kit, the G.Skill DDR3-2400 kit we looked at over a year ago now. For the results designated “A8-3850 – OC-RAM”, the RAM was run at DDR3-2256 with timings of 8-10-8-24.

Starting us off, we’ll look at all four 3DMarks run in one fell swoop.

3DMarks

3DMarks

This APU is very impressive. The IGP truly does have lower-end discrete performance and when you overclock, it just gets better. Score improvements up to 34% are nothing to sneeze at. What really stunned me was how much of a difference the RAM made when you overclock.

Moving on to Heaven, we see big gains in both DirectX 11 and DirectX 9 versions. Nothing dissappointing here.

HWBot Heaven Benchmark

HWBot Heaven Benchmark

Now we’ll move on to the real-world APU performance – how can this platform really perform as an HTPC / home-made console gaming system.

Gaming Performance

Starting off, we’ll have a look at three iterations of Stalker: Call of Pripyat. The bench was run using the DirectX 11, Enhanced Full Dynamic Lighting setting at three different detail settings.

Stalker: Call of Pripyat - High Detail

Stalker: Call of Pripyat - High Detail

Stalker: Call of Pripyat - Medium Detail

Stalker: Call of Pripyat - Medium Detail

Stalker: Call of Pripyat - Low Detail

Stalker: Call of Pripyat - Low Detail

There are definitely strong gains overclocking this platform, especially when you use strong RAM. That’s the real takeaway from a lot of these results – RAM makes a big difference. If you want to use the high detail setting, you’ll want to consider sticking with 720P and enjoying the increased framerates. Both medium and low settings have playable framerates when overclocked.

Aliens vs. Predator was a tough benchmark. Even overclocked, it struggled here because we were seeing how it performed with AA and tessellation on. The low detail setting turned all of that off and the results reflect the change.

Aliens vs. Predator DirectX 11 Benchmark

Aliens vs. Predator

There are definite gains, but not enough to make AvP really playable at 1080P. 720P started off playable and would only get better overclocked.

Two more game benches to go – HAWX 2 and Dirt 2.

HAWX2 and Dirt2

HAWX2 and Dirt2

There are obvious gains throughout in both of these benches. I mentioned in the original piece that Dirt 2 looked playable even at mid-20’s frames per second. The gains from overclocking made it even more so. HAWX 2 started playable and only got better, getting close to matching the stock 720P numbers.

Final Thoughts & Conclusion

Overclocking takes an already strong integrated graphics platform and makes it even better. The gains are very real and palpable.

Llano clocking was certainly a very quirky experience on this A75M-UD2H. With any luck they’ll fix those quirks with a BIOS update. It’s still a very, very young platform so we won’t hold it against them.

The A8-3850 / UD2H combination initially garnered an Overclockers Approved and shows now that it definitely deserves it, not just for its strength as a platform but for its overclocking ability too. It clocked farther and gained more benchmark and gaming performance than expected, which is always a good thing. Thus, if there was any doubt, overclocking Llano is also Overclockers Approved!

– Jeremy Vaughan (hokiealumnus)

Leave a Reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>

Discussion
  1. Did you buy it as a single stick?

    It's probably worth it to have dual channel but I never tested and can't tell you for sure. As RAM-dependent as the IGP is, it's probably a safe assumption that it will help at least a little bit.
    ^^ Dual channel is definitely preferred, particularly when overclocking. But it wont make a significant difference. But ram sticks are cheap enough where it would be justified, just be sure to get the same stick(but not required).
    Keep them where they are; you don't want to loosen timings on this platform. The IGP will suffer performance loss that you won't get back with MHz increase.

    Why is it in single channel?
    Hi,just wondered what timings you would advise....i have corsair vengeance 1600 single channel 4gb...standard ones are 9-9-9-24 i believe....useful advice by the way...first time i've attempted an overclock...i have the a75-ud2h m/b...thanks in advance
    beganovicc
    Sorry to bump into an old thread but i am interested if i do this with my a8, with a high end cooler will the overclocked igp stay at 700+ when i drop in a 6670 for asy hyb. crossfire? And if i use a 2000mhz ram cas9, can i set it at 1600 in bios so it works at ~2100mhz? Or do i need a 2400 ram?


    As per the ram question, most 2000 MHz sticks will do 2100 without much of a fuss. You may just have to loosen timings a bit. But it will depend on you sticks like hokie said.
    I have no idea about crossfire with the 6670, I'm sorry. I presume the IGP would stay the same and you could clock the 6670 manually (you can clock individual cards manually with normal video card based crossfire, even though you want them the same).

    Re: RAM - If you are asking, when you put FSB at 134(ish) and drop the dividier to the one that would be 1600 at stock, yes, it will run ~DDR3-2100. Whether your RAM will get there or not is up to your sticks. You definitely don't need 2400 RAM; there are lots of 2133 kits on the market now, with 4GB kits relatively cheap.
    Sorry to bump into an old thread but i am interested if i do this with my a8, with a high end cooler will the overclocked igp stay at 700+ when i drop in a 6670 for asy hyb. crossfire? And if i use a 2000mhz ram cas9, can i set it at 1600 in bios so it works at ~2100mhz? Or do i need a 2400 ram?
    If that's a direct quote, trust the quote. My memory is (very admittedly) faulty and it is likely I remembered incorrectly.

    Archer - No bus adjustments in BIOS. It's automagic with the UD2H and splits at 133 bus. There may be boards with manual adjustments, and those would allow you to go farther than this on the CPU (by reducing the IGP multiplier), but I don't know what boards have it.
    Hokie are there any bus multiplier adjustments? I remember having these issues on older pentium boards (jumper controled) as well as many socket A boards. I cooked more than one Conner drive because of bus issues.
    so you can't dedicate a full gig to video? this is a direct quote from the article:

    "System memory and IGP memory on the Llano platform are one in the same. By default, the system dedicates 512 MB of system memory to the IGP. You can manually change that if you’d like, I believe to a maximum 1024 MB."

    I want to make sure I'm clear on this, because I just ordered a 3850/mobo/ram today and want as much information as possible before it gets here.

    great article by the way, the best one I've read so far on apu's.
    I can't help but wonder why(if you're able to) you don't dedicate a full gig of ram to the video. Theoretically if you're using dual channel ram, you dedicate 1 gig, oc the ram to 2000 and the gpu to 800, you've literally got 6670(well, minus 80 stream processors) and the almost exact equivalent of a 5570.

    is there some piece of information I'm missing?