Earlier this month, AMD laid months of speculation to rest when it confirmed that it would be releasing two new AM3+ processors based on its Piledriver core. The new chips, the FX-9370 and FX-9590, are both 225W parts. Sampling and retail availability are both uncertain at this juncture; AMD initially implied that the chips would only be available from boutique manufacturers, but that position isn't set in stone.
Still, the announcement got us thinking -- why not take a shot at hitting the FX-9590's 4.7GHz base/5GHz Turbo clock with the FX-8350 (Piledriver core) we reviewed last autumn? That turned out to be something of an adventure, but we got it done thanks to an old cryo-cooler, some bread ties, electrical tape, foam insulation, and a lot of creative verbiage. This article will explain the process and our findings.Cooling a 225W CPU: Phase-change cooling not required
AMD has already stated that the FX-9370 (4.4GHz stock, 4.7GHz Turbo) and FX-9590 will both have a 225W TDP. That's far above the standard 115-130W TDP that AMD and Intel have generally stuck to in recent years, but that limitation has been largely self-imposed. AMD and Intel both pivoted away from higher TDPs when the size and weight of newer heatsinks would have caused problems for mainstream desktop shipments, not because it's impossible to cool a CPU.
We asked premium heatsink manufacturer Noctua if its products could handle CPUs with a 200W+ dissipation. According to company spokesperson Jakob Dellinger, the Noctua NH-D14 and NH-U14S are both capable of dissipating that much heat, though these are top-end air solutions.
Unfortunately, in our case, air cooling alone was not enough to hit the overclock frequencies we were looking for. So we stepped it up a notch.Setting the stage
When it comes to overclocking headroom, every CPU is different. The FX-8350, unfortunately, wasn't a particularly great sample. Even high-end air cooling wasn't enough to hit the clock speeds we wanted. Fortunately, I've got something a bit more esoteric -- a Cryo-Z phase change CPU unit from OCZ, circa 2007.
Phase change cooling works on the same principles that keep a refrigerator cold. A compressor is used to pressurize R507 (the refrigerant) into a liquid, which is then pumped into the condenser. Excess heat is bled off, and the liquefied coolant is pumped into the evaporator, or coolant head. This is the section of the phase change unit that makes direct contact with the CPU. As the coolant expands, it draws heat off the processor. As it absorbs heat it expands, turning from liquid back into gas. Now in a gaseous form, the coolant is drawn back to the compressor, where the cycle begins again. A single-stage phase change unit using R507 will hit a head temperature of around -50 degrees Celsius.
That said, the Cryo-Z isn't a shining example of phase-change capability. It's loud. It's noisy. I'm not sure it ever shipped commercially in any great volume, and to be perfectly honest, mine has a bit of a short -- I had to take the front off to fiddle with the wiring, at which point the temperature readout fell off. It was also designed in an era when CPUs drew less power and is spec'ed for a maximum of 120W. An FX-8350 running at 4.7GHz dissipates about twice that.
The closed-cell foam on top is cut irregularly to give the cooler head more room. There's additional foam underneath this piece. The cooler head imprint is visible above.
My first test got the chip down to -15 Celsius, but keeping the CPU at that temperature under load was a different story. Once the core was fully loaded, system power consumption went from 75W to well over 250W and the Cryo-Z's temperature went with it. The situation was further complicated by the fact that our FX-8350 wasn't a particularly good overclocker. The chip wasn't stable with all eight cores clocked at 5GHz, even when cooled well below freezing.
I cannibalized a second mounting system from an older phase-change unit that had used hard plastic for mounting rather than the semi-flexible plastic on the Cryo-Z. I attached the hard plate underneath the flexible one, secured them together, added additional foam to insulate the socket, and fired the system up again. All eight cores at 5GHz would have been nice, but we ended up backing off that goal and matching the FX-9590's 4.7GHz base/5GHz Turbo.
All these changes got the CPU down to -23 Celsius. More importantly, it increased the time it took for the CPU's temperature to push the cooler head out of subzero territory. Even so, we were rather limited in tests we could run. While the CPU was stable at sub-zero temperatures in any workload we threw at it, once temperatures rose past the freezing point, the CPU would crash in short order. This made comprehensive benchmarking impossible.
Consider this a preview of what AMD's 5GHz CPU is likely to offer -- not a comprehensive review or a final word on the product.
Next page: Testing at 5GHz...5GHz achieved
The final overclocked speeds achieved for the FX-8350 were a 4.7GHz base clock, 5GHz Turbo on four cores -- just like the upcoming FX-9590. I opted for DDR3-2133 (albeit running at fairly high latencies). I also raised the northbridge clock speed to 2600MHz, up from a stock speed of 2200MHz. AMD hasn't said how fast the IMC on the upcoming FX-9590 is, but the FX-8350's memory controller and L3 cache are clocked at 52% of the chip's 4.2GHz Turbo Core speed. At 2600MHz, the IMC and L3 are clocked at 52% of a 5GHz Turbo Core speed.
Coincidentally, this also aligns the IMC clock with the HyperTransport clock, though the benefits of this, if any, are negligible. The test bed was configured with 8GB of Mushkin DDR3-2133 RAM, a GeForce GTX 680, and an OCZ Vector SSD.Cinebench 11.5
Piledriver has always struggled in Cinebench 11.5, being weighed down by its weak single-threaded performance. 5GHz helps alleviate the problem -- a score of 1.32 puts AMD midway between a high-end Nehalem and a lower-end Sandy Bridge. The multi-threaded test is a different story -- here, AMD manages to beat out Ivy Bridge and tie things up with Haswell.
At 4.7GHz (on all eight cores), the FX-8350 brings nearly 40GHz of CPU power to the table (37.6GHz to be exact). Four Haswell cores, in contrast, are running at a total of 14GHz, with Hyper-Threading pitching in about another 20% performance. An overclock at this scale gives AMD room to leverage sheer clock rate, and the performance is impressive relative to the company's previous products.
Our x264 benchmarks show an interesting story as well. In a benchmark from TechSpot, AMD is well behind both Haswell and Ivy Bridge in Pass 1, though it competes well in Pass 2. With the newest version of the program (not an AMD-optimized flavor), the showings change. Suddenly, AMD is neck-and-neck with Ivy Bridge in Pass 1, and beating out both Haswell and Ivy Bridge in Pass 2.
There are, however, still significant areas where our overclocked Piledriver still falls well short of its Intel counterparts.
In PC Magazine's 12-filter Photoshop test, the Core i7-3770K clocks in at 183 seconds, the Core i7-4770K is at 169 seconds, and the FX-8350 is well behind at 221 seconds.
In Kribibench, our browser-based, AVX-enabled 3D rendering suite, the overclocked Piledriver is still well back from the Intel processors in all three tests. We suspect AVX optimization may be partly to blame here, though workload clearly plays a role.
Let's go low-level, and look at some reasons why. First up -- Cache bandwidth and latency.
We've listed cache latencies in nanoseconds, rather than clock cycles. At 5GHz, Piledriver's L1 cache latency is substantially better than either Ivy Bridge or Haswell, while its L1 bandwidth hits Ivy Bridge levels. We know that Bulldozer and Piledriver both have L1 cache miss issues with both cores in a module enabled -- accelerating the L1 indirectly helps this. L2 bandwidth improves significantly as well, though access times remain far back from Intel. L3, for Piledriver, is far too slow. That's a known problem, and AMD simply doesn't have a fix for it on the horizon. Still, the 5GHz boost changes the performance picture.
SiSoft Sandra's multimedia tests illustrate some of why the FX-8350 looks very different depending on workload. In integer x16 workloads, the overclocked FX-8350 is significantly faster than the i7-3770K and bids fair to tackle Haswell. Native float performance is also generally comparative.
The AMD chip still struggles in other workloads, however. We suspect the performance differences visible in even this preliminary test come down to code execution efficiency and, to some extent, multi-threading. In fully threaded scenarios, the chip can bring its higher number of cores to bear.
Next page: Does 5GHz make sense?The 5GHz rationale
You could claim that AMD's push for 5GHz is a desperate move to grab headlines and market share with a CPU that draws far more power than the Intel equivalent and doesn't sweep benchmark results, even running at a much higher clock speed. You'd be right. You could also claim that pushing to 5GHz gives AMD a marketing opportunity and narrows the performance gap between Sunnyvale and Santa Clara to the smallest it has been in years -- and you'd also be right. Personally, I'm impressed that Piledriver yields have improved enough for AMD to launch a 4.7GHz chip with a 5GHz Turbo.
Overclocking is always luck of the draw, but using phase-change cooling is one way to improve the odds of a high-frequency processor. Even taking this route, I had trouble stabilizing the processor -- a full 5GHz overclock on all eight cores was out of the question.
One of the notable differences between the A10-6800K that AMD launched a few weeks back and the A10-5800K that debuted last fall is that the 6800K I tested had no trouble hitting 5GHz on all four cores. Launching these new chips is a sign that AMD has successfully implemented those improvements across its eight-core product line as well.
What it takes to get to 5GHz
So who is going to buy a 5GHz FX chip? These are meant for the handful of stalwart AMD enthusiasts who bought into AM3 and firmly believed in Bulldozer.
Will they sell? No idea. But I can see why AMD -- which built itself up with grass roots support and dogged determination fourteen years ago -- would return to those roots now. The consumer desktop market is still important to AMD, even if the company is betting on consoles, tablets, and low-power servers in the longer term. The FX-9370 and FX-9590 acknowledge that.
Now read: ExtremeTech's Haswell Core-i7 4770K review
