AMD FX-8350 – Piledriver – CPU Review

Today AMD is releasing their new FX processors. They are an evolutionary step from AMD’s oft-maligned Bulldozer processors released just over a year ago. How large of a step forward is Piledriver? We’re here to show you!

Vishera Explored – What’s New in Piledriver?

Like Bulldozer before it, the Piledriver CPU (code named Vishera) is composed of two, three and four module CPUs. A module consists of two integer cores and one floating point core, so when AMD references a “core”, that mean an integer core. Traditionally, before Bulldozer (and currently on Intel CPUs), a core was one integer core and one floating point core. Things are different on the AMD side of the fence now.

AMD basically took a little bit here and a little bit there, with optimizations over pretty much the entire CPU. It’s not a die shrink, nor is it a new architecture. As I mentioned, it’s an evolutionary step.

Another enhancement not mentioned yet is the clockspeed increase. Piledriver weighs in at 4.0 GHz base clock on all cores and 4.2 GHz boost in lightly threaded loads. The boost clock is the same as Bulldozer, but that base clock is a 400 MHz gain, 11% over Bulldozer. In heavily multi-threaded loads, that clockspeed increase is where a large chunk of Piledriver’s estimated performance gain of (up to) 15% over Bulldozer comes from.

The good part for people that already have 9-series chipset boards is that they will be keeping the same chipsets (and AM3+ socket)  for Piledriver. All you’ll need is a BIOS update. Indeed, we’re reviewing Piledriver on the same 990FX-based Crosshair V Formula Bulldozer was reviewed on.

We’ll explore performance obviously (that’s why you’re here after all!), but AMD of course supplied some numbers of their own for us to check out.

AMD has long focused on multi-threaded computing, mostly because single threaded performance is where Intel dominates currently. If you can find an application that does take advantage of multi-threading though, AMD does well for itself.

The FX-8350 even shows pretty strong FPS gains over the FX-8150 when paired with an HD 7970 GPU.

Here we start to see what part of the market AMD is aiming for. I don’t think anyone expected it to compete with the top Ivy Bridge CPU (i7 3770), but it seems to match up quite well with the i5 3570K. You also get your first glimpse of the pricing and it is actually quite nice, coming in at only $195 for the FX-8350. So far so good AMD.

In this slide they’re comparing CPU and GPU cost added together; in AMD’s testing, they come out ahead but for less money.

Interestingly, AMD mentions performance benefits of overclocking to 5 GHz on water cooling. Now, that will be heavily dependent on your particular CPU. We couldn’t quite make 5 GHz, but if you can, these are some of the gains you can expect across the FX line.

The FX-8350 isn’t the only CPU with pricier Intel competition. Down the line, AMD is pricing the new FX series CPUs very aggressively against their competition.

We’re past the company line now; it’s time to get down to the nitty gritty.

Meet the FX-8350

There aren’t many ways to photograph a CPU but we’ll try to make the FX-8350 presentable for you.

There is definitely more to look at when you put it in a good looking motherboard.

Of course, no CPU review would be complete without a die shot or two.

Yep, it’s a CPU all right, complete with a pretty die inside.

Power Consumption

The first thing we’ll look at is power consumption. Frankly, it’s not a pretty sight for those that are looking for super efficient computing.

So, um, yea. Power consumption was not a strong point of Bulldozer. Add some MHz to that equation (which of course requires some voltage) and things don’t get any better. If you are into distributed computing (Folding@Home, SETI, Rosetta, etc.), you’ll probably want to go Intel at this point. FX processors are just too inefficient to be of worth to you.

Overclocking for Stability

AMD mentioned 5 GHz in their reviewer’s guide and press deck (the latter surprised me). Time was limited for this review, so the only overclocking I really did was add voltage and increase the multiplier. The chip made 4.9 GHz quite easily, but once I got to 5 GHz, it just wouldn’t stabilize. Once the voltage went past 1.5V, I stopped there.

The good news is that Piledriver obtained a 4.9 GHz stable overclock! That’s 100 MHz north of Bulldozer. It easily booted at 5 GHz and ran benchmarks, but it was not 24/7 stable.

In other good news, look at those temperatures. It took a fair bit of voltage to get to 4.9 GHz stable (1.476 V loaded) and temperatures were less than Bulldozer on the same water loop (Swiftech MCR-320 radiator, Swiftech MCP-35x pump, EK Supreme HF CU block). With increased voltage and frequency, Piledriver clocked in ~5°C lower than Bulldozer. Not bad. Not bad at all.

Test System, Opponents & Methodology

We’ve got a solid range of CPUs tested for your viewing pleasure today, from the venerable Thuban hex-core 1100T all the way up to Intel’s 12-threaded monster i7 3960X and plenty in between. Most important to this review is the inclusion of the Ivy Bridge i5 3570K, which is precisely the chip at which AMD is aiming with the FX-8350.

Speaking of market targets, I’ve also included current prices for those chips still on the market (prices as of 10/22/12 obtained at Newegg).

Great competition top to bottom today. Now we must pause a second to give a shoutout to our esteemed editor EarthDog, who did us all a solid favor by testing his i5 3570K so we would have results from that CPU in this review. He actually also tested the 2500K results from last year for the Bulldozer review. Thanks Joe!

Performance

Performance graphs are a little different this go-round, so please take note of what I’m about to say!

AIDA 64 graphs are graphed with percentage only. I can’t name one person that runs AIDA all the time and can easily decipher what the results depict, so we’ve gone with percentages only. The percentages expressed are how that particular CPU does compared to the stock FX-8350 – a higher percentage means the CPU scores better (except for memory latency, where lower is better) and a lower percentage means the CPU scores worse. All results are relative to the stock FX-8350.

The rest of the graphs are based on actual score, time and/or FPS. Percentages are useful for graphing multiple metrics in one graph but our CPU list has grown such that it isn’t practical to graph more than one result per graph. Thus, we’re going with actual scores, so you can see how the CPUs compare directly rather than confusing things with percentages and actual scores in parenthesis.

Then, at the end of the performance section, when we evaluate the FX-8350 vs. the FX-8150, we’ll go back to percentages to show how much Piledriver improved over Bulldozer.

AIDA 64 Benchmarks

First up, the synthetic benchmarks, specifically the AIDA 64 CPU tests. Here, the FX-8350 comes out ahead of the FX-8150 by an average of 13.7%. As an average, that’s not bad at all; it even got up to 20.5% in the Zlib test.

With regard to the competition, it is a solid win over the 3570K, even besting the 3770K in two of the five CPU tests. Interestingly, the 2600K is now behind AMD’s offering – and still priced considerably more.

FPU tests are where AMD is struggling at this point, even with heavily threaded loads. Because there is only one FPU core per module, there are only four FPU cores in this “eight core” CPU. That hindrance is shown here. While it does beat the FX-8150 by a healthy 9.5%, it does lose out to its Ivy Bridge competition. However, it is certainly a step in the right direction – even with its lack of FPU capacity, it still beats out the 2600K.

Ahh, memory. This is where you still see AMD struggling a bit. The memory on the FX-8150 was running at DDR3-2400 / 10-12-12-31. The memory run today is worse than the FX-8150, which had an unfortunately deceased memory kit to run, spec’ed at DDR3-2000 / 7-9-7-24, but even with that, the FX-8350 comes out ahead in bandwidth testing, so there were solid improvements to the memory subsystem for sure.

However, Sandy Bridge still has more bandwidth to offer than Piledriver and Ivy only increases that gap. Latency is looking solid though, with only the 3770K (and the better memory on the FX-8150) beating the FX-8350.

Overall, this is a solid win for Piledriver. Priced where it is, we’re starting to see a decent value here.

(If anyone wants to see the actual AIDA 64 results, simply have a look at this chart.)

3D Benchmarks & Games

For our 3D benchmarks, we still use an HD 6970, mostly because we’ve got data from our past reviews. Regrettably, I don’t have an i5 3570K to test and EarthDog doesn’t have an HD 6970, so we can’t see how that fares. However, as you’ll see, the FX-8350 does pretty well for itself.

One very important note in these tests – I lost my old driver used for the previous tests and used Catalyst 12.8. There are increases here due to the CPU and there are increases due to the driver. I will remedy this disparity in a future article though and test the FX-8350 versus the 3770K on a strong HD 7970. For now, this is the best we could do. Please accept my apologies for the accidental driver deletion!

First up, we’ll do synthetic tests. Starting with 3DMark06, this test is very CPU bound; even the graphics scores are heavily tied to the CPU. What I like here is the very strong increase over the FX-8150. What I like even more is that it’s approaching the 2600K’s performance level, especially when overclocked. Take four threads away from the 3770K and you’re probably looking at near-parity with the 3570K.

Vantage again shows healthy increases for the FX-8350 over the FX-8150 and a much closer score to the 2600K.

3DMark 11 shows healthy gains. When you get to DX 11, that driver difference starts to show itself, so remember that caveat. In any case, the Physics test doesn’t change with GPU driver and the FX-8350 is doing great with huge increases over the FX-8150. Overclocked it’s also approaching the 2600K.

Heaven shows great gains, with the FX-8350 actually coming out ahead all around.

Our game suite for CPU reviews is long in the tooth. One of our games (the HAWX 2 demo) can’t even run any more. Again, I’ll be doing a follow-up testing an HD 7970 with our new game suite.

In the three games we did test, the FX-8350 is doing well for itself, with gains over the FX-8150 and relative parity with the Sandy and Ivy Bridge chips.

Once more – this CPU will be retested on a newer GPU with identical drivers. Until then, I think it’s safe to say the FX-8350 does a solid job over its predecessor, based especially on the CPU tests in 3DMark benchmarks.

2D Benchmarking

We’ll get the worst thing out of the way first. AMD does not do well with SuperPi. They haven’t been competitive in this bench for a long time against Intel. AMD would note correctly that this is an antiquated benchmark and does not measure real-world performance, with which we would agree wholeheartedly.

SuperPi is, however, a favorite of competitive benchmarkers. Thus we still include it as a solid subsection of our readership enjoys such things. Very strangely, the FX-8350 performs worse in SuperPi than the FX-8150 did. I believe the RAM is the biggest problem here, as SuperPi is heavily dependent on RAM, especially when you’re talking the length of time it takes AMD CPUs to complete this benchmark.

You see, IMOG killed my super awesome, super rare G.Skill Flare kit (that’s right, I’m calling you out boss man!), which would have most likely shown improvements akin to what you’ll see below with the WPrime results. This is a good thing for the reader, because you can see how this CPU will perform with a kit currently available on the market. It’s a bad thing for comparison graphs, because you can’t get a direct 1:1 comparison with dead, no longer available RAM.

The disparity due to RAM only gets worse when you move up to SuperPi 32M. Note the scale at the bottom of the graph is in seconds; the actual times are next to the bars.

Now we’re back to showing some gains. WPrime doesn’t care one bit about RAM (within reason) and Piledriver shows healthy gains over Bulldozer in both benchmarks. You can also see it competes very closely with its 3570K competition at stock, then comes out quite a bit ahead when overclocked.

RAM problem notwithstanding, we continue to see solid gains for the FX-8350. Things are looking up for AMD.

Rendering, Video & Compression

Ahh, now we get to the real world benchmark comparisons. All of our real world tests are multi-threaded because, let’s face it, the vast majority of CPU-intensive real world computing has moved to multiple threads. If the program you use doesn’t, it’s time to upgrade.

So, we’ll start by rendering some images with Cinebench R10. In R10 Piledriver has a respectable performance, showing strong gains over the FX-8150, beating the 2600K and coming close to -but not beating- the 3570K. When you overclock it, things shift squarely in the FX-8350’s favor.

Cinecench R11.5 is a more modern flavor of the rendering benchmark and the FX-8350 is here with a win over its older brother and its Intel competition. Overclocking just brightens the picture, coming ahead of the 3770K.

PoV Ray and Piledriver get along very well, with the stock FX-8350 actually beating the 3770K and then coming quite close to the 3960X when overclocked.

The x264 Benchmark has two passes, one is a reading pass (Pass 1), where the CPU basically looks over the file. That one doesn’t look so good for the FX-8350. However, when doing the actual work of encoding, which happens in Pass 2, Piledriver flips the script, with numbers more in line with the PoV Ray results above – beating the 3770K at stock and coming near the 3960X when overclocked. The 3570K is left in the dust.

7 Zip loves some threads too. At stock, the FX-8350 trounces the 3570K and comes very close to the 3770K. Overclocking brightens the picture considerably.

When looking at real world, multi-threaded applications like rendering, compression and encoding, the FX-8350 does very well for itself. It is a good showing over the FX-8150 and looks good when comparing to Intel’s offerings, especially the 3570K.

FX-8150 vs. FX-8350 – Out of the Box Comparison

Now we’ll see how the stock FX-8150 compares to the stock FX-8350. AMD says up to 15% performance increase.

You already know the issue with SuperPi and my dearly departed RAM, so you can basically disregard the disparity there. WPrime is more accurate and shows roughly a 5% decrease in times.

These two benchmarks test only the floating point cores and AMD obviously wasn’t able to increase floating point performance quite as much a 15%. AIDA showed an average FPU performance increase of 9.5% and WPrime gives a performance increase of 4.9%. Any increase is good, but don’t expect a large floating point difference between Bulldozer and Piledriver.

Moving on to the remainder of our benchmarking suite (warning, this is a large graph, just over 1100px wide), we see as-expected performance gains, up to 18% in our testing. The average of all performance gains ends up being 11.9% across the board.

Overall, we see a solid performance increase over the previous generation.

Pushing the Envelope

Of course, freezing Bulldozer was one of the best things about it. Top frequency records were set all over the place and just like previous AMD chips that came before, you could run a full pot and never worry about any sort of cold bug or cold boot bug. I’m very happy to report that Piledriver continues that tradition, with hours of full pot benching to be had.

So how did the FX-8150 do after several hours under the cold? Not  bad, not bad at all. WPrime 1024M was able to complete at over 6.4 GHz and WPrime 32M at over 6.5 GHz. Both of these are close to what I got on an FX-8150, but not quite as far, trailing by ~200 MHz each, but at ~.1V less.

Of course, single core benchmarks were able to run faster than that, with PiFast coming in just north of 6.8 GHz and SuperPi 1M just breaking the 7.0 GHz mark. Both of these are well short of my FX-8150 clocks, but both of them were also at almost 0.2V less. This is a good start for Piledriver; a very good start.

Last, but not least, we come to the CPUz validation – where Bulldozer shined. This Piledriver chip managed 7.3 GHz fully CPUz validated, which is an impressive number considering the BIOS limitations. You see, Bulldozer could clock so high because there was a BIOS that allowed you to disable cores within modules. You could then enable one core in your strongest module and one in another module. Then you would leave the weaker core clocked at a “stable” frequency and then shoot the moon with the other one. Unfortunately with the Piledriver BIOS for the MVF, you have to use full modules, which doesn’t lean itself to as high clocks. Hopefully there will be a BIOS released with that feature!

Regardless, 7.3 GHz validated with a full module is impressive in and of itself.

It’s also important to note that I needed this chip to survive so we could bring you more testing as promised, so I didn’t push voltages as high as I did with the Bulldozer chip (over 2.0V, killing it in the quest for a higher CPUz validation…you can see all that chip’s results at HWBot). I’m sure with more voltage, this chip could do even better. In any case, kudos to AMD for keeping Piledriver just as fun to overclock as its older brother!

Final Thoughts & Conclusion

I didn’t give Bulldozer as bad a review as a lot of people did. Unlike many, I tried to look at Bulldozer as just any other CPU, not the return of the FX of days past, when AMD ruled the roost and Intel was playing catch up. I gave Bulldozer an Overclockers Approved, in large part because it was fun to overclock -especially under extreme conditions- and because it was priced right. However, when it came out, retailers/etailers boosted the price over what AMD said it would be. I felt duped, not just because of the pricing but because AMD didn’t step in and make the retailers price the CPU where it should be.

As long as it is priced in line with what AMD says the price will be this time, that all changes with Piledriver. As AMD points out, this is a processor evolution from Bulldozer. It isn’t a die shrink or a micro-architecture redesign, it’s basically optimization and higher clockspeeds over the CPUs released a year ago. They reinvented the wheel with Bulldozer. They chiseled the edges and made it roll smoothly with Piledriver.

So, what does AMD promise here? Up to 15% performance increase over Bulldozer. That they do deliver. They also deliver lower temperatures and higher overclocks in our testing.

What it does not do is perform with high efficiency. This is a very power hungry CPU, there is no way around that. Without major changes to the architecture or a die shrink, Piledriver couldn’t do anything about that. If the name of your game is efficiency, Intel has that in the bag. We tend to not really care too much about efficiency here, but it does bear mentioning since there is such a large gap in power pull.

When it comes to performance, their target in the current CPU market is the i5 3570K. As you saw above, in single-threaded applications Intel still reigns, but when you get into multi-threaded applications the FX-8350 really does quite well. It also does it for less money, coming in at $195, which is a very economical price for performance. Heck, with heavily threaded metrics, it often hangs with or beats the much more expensive i7 3770K.

Add to the very aggressive pricing strategy the fact that the FX-8350 is still a blast to overclock – both with ambient cooling and sub-zero – you have a formula for a CPU that deserves to hold its head high among its competition. Piledriver is what Bulldozer should have been. It performs well against its targeted competition and is perfectly priced to compete. As long as they keep that price (or very close, say $199.99 to $209.99), AMD has a good competitor on its hands with the FX-8350.

– Jeremy Vaughan (hokiealumnus)