AMD Radeon VII Review: Performance Benchmarks With 7nm Vega

Article Index

AMD Radeon VII: Overclocking, Power, And Noise

We also spent some time messing around in the Radeon VII’s WattMan driver control panel, to tweak and tune a few things and see what kind of additional performance we could squeeze out of the card.

Like previous-gen Vega-based Radeons, when the GPU is boosting, frequencies and voltages scale upwards (power and temperature permitting) based on the GPU's workload at the time. With the WattMan tool built into AMD's drivers though, users have the ability to manually alter frequencies, voltages, fan speeds, and the power target to fine tune performance and power consumption. There are some additional wrinkles to consider with the Radeon VII, however.

health data

In previous-generation designs, though the GPUs had multiple sensors built-in, a single sensor has been used to determine the GPU temperature and data gathered from that sensor was used to control the card's cooling fans and throttle the GPU when necessary. With the Radeon VII though, AMD has incorporated a network of 64 thermal sensors at strategic locations across the GPU die -- double the number of sensors found in Vega 64. Data from this array of sensors is used to determine what AMD is calling the "Junction Temperature" on Radeon VII, and it's the Junction Temperature data that is used to tune the Radeon VII's power and thermal profile. AMD claims the increased resolution and accuracy from the Radeon VII's additional thermal sensors allows it to increase overall performance, because thermal throttling based on the Junction Temperature is more reliable and effective.

Unfortunately, it appears all of the new bells and whistles didn’t seem to play nice with the Radeon VII and the drivers we had available for testing. The WattMan tool built-into the Radeon VII’s drivers offer manual tuning, along with automatic under-volting and automatic GPU and Memory overclocking. Manually tweaking any of the GPU related settings, however, resulted in lower performance for us. We were advised by AMD that we were likely to achieve the best GPU overclocking results if we under-volted the GPU, because the cards quickly run into a power limit. However, even following their recommendations to the letter, performance always decreased. Even simply turning the power target up and manually increasing fan speeds to keep temps lower dragged performance down – it frankly didn’t make much sense.

health data oc

That said, auto-overclocking the memory seemed to work properly. When auto overclocking, we saw memory speeds peak at about 1,225MHz, which is a fairly big jump. With that data in mind, we tried to manually tweak memory speeds further, only to find that manually altering the memory frequency via the slider brought the clock down into the 800MHz range.

So, overclocking is effectively a lost cause at this point with the Radeon VII. We suspect as AMD spends more time with the cards, fine-tunes the Radeon VII’s power plan and works the kinks out its drivers, overclocking may become a thing, but for now the Radeon VII is not very tweaker-friendly.

oc1

oc2
We did, however, re-run a couple of tests with the auto memory overclocking option enabled and saw some mild up-ticks in performance. While its memory was overclocked, the Radeon VII was able to match the RTX 2080’s performance in Tomb Raider at 4K and its performance increased slightly in 3DMark Time Spy as well, though not enough to overtake the RTX 2070.
 

Total System Power Consumption
Tested at the Outlet
Before bringing this article to a close, we'll cover a couple of final data points for power consumption and acoustics. Throughout all of our benchmarking and testing, we monitored noise output and tracked how much power our test system was consuming using a power meter. Our goal was to give you an idea of how much power each configuration used while idle and also while under a heavy gaming workload. Please keep in mind that we were testing total system power consumption at the outlet here, not the power being drawn by the graphics cards alone. It's a relative measurement that does give you a decent view of how much additional, or less power draw a graphics card is placing on a system.

power

The Radeon VII's power consumption characteristics are significantly improved over the Radeon RX Vega 64. Although Radeon VII significantly outpaces the Radeon RX Vega across the board, the card consumed 46 fewer watts under load. Idle power was somewhat higher, but was nothing to be concerned over. That said, the Radeon VII did use more power than the GeForce RTX 2080, though it was in the same ballpark. The move to 7nm significantly enhances the Radeon VII's performance per watt versus its predecessor.

We should also note that using the auto-undervolt feature in Wattman shaved about 22 - 30 watts off the Radeon VII's peak power consumption, without sacrificing any performance (though we only ran a handful of tests).

Despite its more power-friendly nature and beefy triple-fan cooler, however, we found the Radeon VII to be rather loud under load. When idle, the card is essentially silent and can't be heard over the slight hum of a typical CPU cooler and PSU. Once the card heats up though, the Radeon VII's fans spin up considerably and the card generates a fair bit of noise. It's not obnoxious, but the Radeon VII can easily be heard outside of a closed chassis and it's clearly louder than competing cards from NVIDIA.

We suspect AMD is being overly aggressive with the Radeon VII's fan profile, because the GPU doesn't get excessively hot. GPU temperatures under load were typically under 70ºC. The Junction Temperature, however, would hover in the high 90s, but that's to be expected, considering that number is derived from multiple sensors situated in the hottest parts of the GPU. It's possible the Radeon VII GPU's 7nm manufacturing process is less forgiving at higher temperatures and AMD must keep temperatures lower, but we're just speculating at this point.

Related content

Comments

Show comments blog comments powered by Disqus