COMPARING INTEL AND AMD at the high end can be difficult. A benchmark might show that an Intel Core i7-975XE obliterates an AMD Phenom top offering, but ignore the fact that the Phenom CPU might cost only one-third of the Intel CPU. Even when you add in all the other system components, the AMD system might cost 30 per cent less than the Intel system for 25 per cent lower performance.
The arrival of less expensive socket LGA1156 Intel Core i5 and i7 chips changed that a bit. They are cheaper, but slightly slower than the top level Intel 975XE CPU. The system cost is also about on a par with the AMD top CPU bins. Yet, trying to put together an apples-to-apples comparison is anything but easy. After all, even though they both use dual-channel DDR3 memory, the CPU core speeds and even cache sizes are still different. So, here I tried my best to set up two configurations with as many common components as possible.
The first one, using the Intel Core i7-870, is based on Gigabyte P55 UD6 mainboard, while the other one, for the AMD Phenom II 965BE, uses the Gigabyte MA790FXT UD5 mainboard. Both are the top end models from this vendor for the Intel P55 and AMD 790 chipsets, respectively, so they both have maxed out power and voltage conversion circuitry with the best capacitors, 2-ounce copper layers, optimised memory traces and such.
While the UD6 mainboard uses Intel's P55 chipset, where the PCIe bus hangs off the CPU directly for lower latency GPU access, the UD5 mainboard relies on the AMD 790FX chipset and its 'classic' PCIe attachment off the North Bridge for GPU access. The benefit here is more total PCIe lanes available, of course - full 2 x 16 for the AMD 7909FX chipset versus 1 x 16 or 2 x 8 in the Intel P55 chipset.
Also, both mainboards use twin 2GB modules of optimised memory. Both Kingston DDR3-2000 CL8 and G.skill DDR3-2000 CL9 were used for the tests. The CPU cooler was shared too, the brand new Thermaltake SpinQ VT, a good looking yet fairly large and mostly silent high-end unit which managed to keep both CPUs at less than 45C when idle in the Singapore heat at 4GHz and above.
The graphics card was the current performance favourite, the reference AMD ATI Radeon HD5870 1GB with the Catalyst 9.10 drivers for Windows 7. The Thermaltake 1000W PSU powered the whole contraption in both cases. So, plenty of common stuff there.
What happens when we overclock and tune the machines? The Intel Core i7-870 does 4 GHz fairly easily here with the Turbo Boost still on to give an extra clock step or two, but the AMD Phenom II 965BE is no slouch either, with 4 GHz achieved on air cooling here easily. However, the Intel Core i7-870 needs its HyperThreading multithread feature turned off before it can reach the 4GHz plus Turbo Boost level safely.
The voltages needed to achieve that are a little different too, even though both processors are made in 45 nm process technology. The Intel Core i7-870 needs 1.42 volts to run stable through all the benchmarks, while the Phenom II 965BE needs 1.45 volts for the same, but again, this difference is nothing major.
Since the Intel Core i7-870 has a locked multiplier, I had to set the base clock to 182MHz, a hefty over one=third jump up from the base 133MHz, with the same locked 22x multiplier. With that, I had 4GHz base frequency, and two extra Turbo Boost bins for either 4.18GHz or 4.36GHz net Turbo Boost frequency across all cores when running. In some benchmarks like CineBench, it ran at 4.36GHz, while in others like Sandra, it didn't go above 4.18GHz.
At the same time, I attempted to push forward the memory settings using both G.skill Ripjaws brand new DDR3-2000 CL9 modules (see them inside the mainboard on the above picture) as well as the classic reference, the fast Kingston HyperX DDR3-2000 CL8 series with Elpida chips. How did they fare on each CPU?
On the AMD, the G.skill DIMMs did well up to DDR3-1600 CL8 without any problems at around 1.6V. However, getting anything better above 7-8-7 at that frequency proved not possible even with a 1.65V setting. Nevertheless, knowing these are aimed to be inexpensive DIMMs, it's fair. On the Intel, they did run at DDR3-1800 CL8-8-8 but at 1.66V only.
The Kingston again ran at very low latency, with the AMD Phenom II configuration easily achieving it's lowest latency DRAM setting, the CL 6-6-6-16 at DDR3-1600 dual channel, at just 1.62V. In the case of Intel, it was DDR3-1820 at CL 7-7-7-18 at 1.64V, not bad either. Since we couldn't exactly match the clocks and latencies, this compromise - a 12 per cent faster clock but 15 per cent higher latency - is a close approximation for a fair comparison.
Another point of observation was the GPU. The AMD (ATI) Radeon HD5870 is the fastest single GPU graphics card around right now, and it puts considerable demands on the system CPU, the PCIe bandwidth and latency, as well as on the system memory during operation. It was interesting to see whether the lower PCIe latency in the Intel Core i7-870 versus the more AMD specific optimisations in the AMD 790FX chipset could help either win there.
With all these equal, I ran the following benchmarks: Sandra 2009, CineBench, and 3DMark Vantage in Performance mode.
3Dmark on Intel
3Dmark on AMD
CineBench on Intel and AMD together
As you can see, the i7 still has the advantage in most cases. Here are the overclock setup benchmarks with a couple more results, with the accompanying screenshots:
As well as the memory benchmark
Well, the Intel Core i7-870 has somewhat faster memory system, as you can see - DDR3-1800 at ease versus DDR3-1600 on the AMD Phenom II 965BE, but that's not a showstopper difference at the end. It's more the memory controller efficiency here - for the same DRAM settings, the Intel Core i7-870 seems to outpace the AMD Phenom II 965BE quite a bit, with the difference for the nearly the same dual-channel - DDR3-1600 CL6-6-6-16 on AMD versus DDR3-1800 CL7-7-7-20 on Intel - settings being some 40 per cent in favour of Intel.
Also, while I disabled HyperThreading, which tends to skew the benchmark results a bit, and therefore enabled full core versus core and thread versus thread comparison, I left Turbo Boost on as it is a free feature that, in most cases, a user would keep. If you disable Turbo Boost, you can run the CPU at a somewhat lower voltage - 1.36V in this case - but even with Turbo Boost on, the required voltage is still lower than the voltage required on the Phenom.
System architecture wise, both platforms are well developed - optimised dual-channel DDR3 memory, decent PCIe performance (although Intel P55 needs more PCIe lanes for everything, whether dual graphics at 16X each, or extra PCIe for USB3 and SATA3 controllers), and upgradeability.
However, as mentioned before, AMD really needs an improvement to its basic core performance. Even if I disabled the Turbo Boost completely, the Intel CPU would only be some 3 per cent to 7 per cent slower in the benchmarks shown here. That still wouldn't change the picture significantly, and the picture is that, clock for clock, core for core, with the surrounding systems as similar as possible, and - importantly - price levels as similar as possible, Intel's platform is still ahead.
Of course, with the right price coupled with some power usage reductions, the AMD Phenom II is still more than good enough for most desktop applications, but price alone should not be the only argument to use.
So, the AMD Phenom II is on its own a great CPU nevertheless and, even if AMD ends up late with the Bulldozer next generation, there is still room for AMD to optimise performance rather than just fight on the price front. One approach could be for AMD to move the six-core Istanbul die into the Phenom II socket and let you have six cores.
In my mind, that's not the right choice as the clock speed would invariably suffer and frankly, there is not that much desktop software that can use six cores. However, adding more L3 cache while lowering its latency and tuning up the cores to get another 10 per cent or more of per-clock performance could be a good interim solution. Remember, Intel's six core Westmere-based Gulftown high-end CPU for workstations and desktops is less then six months away, and is most likely held back simply because Intel has no competition in that space. A year from now, the first Intel Sandy Bridge cores with further speedups like twice the floating-point unit throughput are expected, so the least we can hope for is that AMD's Bulldozer comes out a little earlier. µ