Test Platform (Sponsored by Corsair)

For AMD, we primarily tested on these platforms:

• Gigabyte X470 Gaming 7, EFI revision F4B
• ASUS Crosshair VII, EFI revision 0505
• ASUS Crosshair VI, EFI revision 6008

Note that our initial Gigabyte X470 BIOS revision did not support disablement or tuning of XFR2. We ended up scrapping initial (internal) test data and replacing it once we received more consumer-ready BIOS revisions.

Wherever “X370” is mentioned, that’s the Crosshair VI. When X470 is mentioned, in gaming and production benchmarks, that’s the Gigabyte X470 Gaming 7. For overclocking, memory scaling, and game streaming tests, that’d be the ASUS Crosshair VII motherboard.

For Intel, we’re primarily using the ASUS Maximus X motherboard. Synthetic and Blender results, along with power results, are using the Gigabyte Ultra Gaming. We tested with the latest BIOS and with new Windows versions.

Ryzen 2 Minimum Stable Voltage to Hold a Frequency

Our review begins by discussing the most immediate improvement in the new Ryzen CPUs, which is the lowest stable voltage at a given frequency. The frequency curve for Ryzen is somewhat exponential, in that a 4.2GHz clock might take 1.38V to 1.4V to sustain, but a 4.3GHz clock takes beyond acceptable safe voltages in our testing.

To this end, we found that, at a given frequency of 4.0GHz, our R7 2700X held stable at 1.175V input at LLC level 4, which equated to 1.162V VCore at SVI2 TFN. The result was stability in Blender and Prime95 with torturous FFTs, while measuring at about 129W power consumption in Blender. For this same test, our 1700 at 4.0GHz required a 1.425V input at LLC level 5, yielding a 1.425VCore, a 201W power draw – so 70W higher – and pushed thermals to 79 degrees Tdie. That’s up from 57.8 degrees Tdie at the same ambient.

Prime95 produced similar results, but you can find those in the article below.

Lower voltage for a given frequency also means lower power consumption for the same frequency. To some extent, this is binning – but most of that large delta is from improvement of the product’s clock efficiency at the “old” high clocks of 4.0GHz. To get to 4.2GHz and beyond, granted, does take over 1.42V on our chip. It’s a significant, nearly exponential curve to increase frequency by a couple hundred megahertz. We found 4.3 to be impossible to sustain on 3 of our CPUs that we’ve tested.

AMD Ryzen 2 (R7 2700X) Memory Overclocking

Memory overclocking hasn’t changed much with Ryzen+ and the X470 platforms. Initial reports suggested that 4000MHz would be achievable and somewhat standard on X470, including some off-record statements we received many months ago, but that’s not the case. Memory overclocking has hardly changed.

We found that both of our platforms were capable of handling 3600MHz XMP for our GSkill Trident Z kit on cold boot. Versus Ryzen’s launch, this is an incredible improvement; versus a month ago, it’s really not any different. Overclocking to 3666MHz, not that this is much of an OC, was also possible – but it required training. We could cold boot 3600, but we had to train to 3666MHz. Anything beyond that was unachievable on our board, CPU, and memory. Note that our memory was the kit used to take a top 5 world record recently, so we know it hits at least 4000MHz.

As for memory scalability, we did some testing on that as well:

This test was conducted with completely controlled memory timings, meaning we controlled every single timing presented in UEFI. That includes subtimings and tertiary timings.

When you control timings 100% and change only the memory frequency, plus whatever unpresented timings are being modified, the result for TimeSpy is what you see on the screen now. TimeSpy is highly memory sensitive, and we end up with 3666MHz plotting 9465 points, an uplift of 0.6% over 3600MHz, which is 0.4% over 3466MHz. You get the idea. There is absolutely scaling here, but it’s incremental. At 3600MHz versus 3200MHz, we observed a 3.5% difference. Given the price difference in some regions, you may as well stick with 3200MHz.

The difference between 3200MHz and 2400MHz is tremendous, though, marked at 12.5%. That’s with TimeSpy, which is memory sensitive.

Here’s where it gets interesting: With the Crosshair VII Hero, tuning the frequency down will result in an automatic tightening of subtimings, which results in increased performance overall. If you were to manually enforce 16-18-18-36 timings and allow all other timings to be auto, which most users do, and then set the memory frequency, you’d end up with this chart.

The result is much tighter. We suddenly have a range of a couple hundred points, rather than 1400 points. Instead of 8070 to 9465, we’re at 9100 to 9300. That’s because the timings automatically tighten to a point of improving beyond the frequency improvement.

Basically, if you are really serious about performance, you’ve got a lot of work to do on auto-tuning timings. ASUS does an exceptional job at this automatically on their Crosshair VII Hero, and this is where you see the value of good motherboards emerge. This also illustrates how testing can be completely invalidated if you don’t pay close attention to those lower-level, buried timings, but more importantly, it shows that buying 3200MHz kits with tighter timings can perform every bit as well as higher frequencies and looser timings. That said, it’s best to go faster and with tighter timings, and not everything responds the same as TimeSpy did.

Here’s an example. In Firestrike, which doesn’t use the same API as TimeSpy, we observed nearly 0 scaling between memory frequencies, even when controlled for timings.

Cinebench also demonstrated nearly zero scaling difference; at least, not until 2400MHz.

Continue to Page 2 for AMD R7 2700X vs. i7-8700K game streaming.