AMD’s Polaris — where is the efficiency?

Sapphire Nitro+ Radeon RX580

The Radeon Rx 400 series launched over four years ago and was later refreshed as the Rx 500 series. It was based on AMD’s Polaris gpu architecture and was intended for the midrange and budget segments of the market. It delivered extremely good performance for the price point, but had poor power consumption — often using more power than competing graphics cards, primarily the Nvidia Gtx 1060. I’d like to see how much the efficiency improves with undervolting and underclocking.

During Polaris’s development, however, the card was shown to be much more efficient. People wondered if the power numbers were some kind of fantasy spun by AMD marketing, or if something had changed in the design of the gpu midway through. What happened to the efficient gpu we expected?

Part of the problem could simply be how AMD ships the card from the factory — they ship the cards with higher than optimal clocks. Computer chips have efficiency sweetspots, where the power draw relative to the performance is the best. Gpu makers often push the performance of their cards past this sweet spot at the cost of performance per watt in order to remain competitive. I believe that AMD did so with the polaris generation. I’ve benchmarked a radeon rx 580 at lower clock speeds to see if efficiency improves.

Here’s the specs of the test system I’m using today.

GPU Sapphire Nitro+ RX 580

CPU Ryzen 5 3600

RAM 16 gb Crucial Ballistix 3600 16–18–18–38

Motherboard MSI b450 Gaming Plus Max

PSU Corsair Tx 850

Power draw and fps was measured with the gpu clocked at 1400, 1300, 1250, 1200, 1100, and 1000 mhz, all undervolted to the threshold of stability. I added the tests at 1250 mhz as the results at 1200 were very interesting and I wanted another datapoint. For 1000 mhz, I also underclocked and undervolted the memory as well, as the gpu core voltage could not be lowered below the memory voltage. This probably lowers the power draw and performance of the card further, but I had to do this in order to undervolt the core fully at 1000 mhz.

All games were run at a resolution of 1080p, with all settings on ultra quality. Power draw was measured for the gpu core using AMD’s gpu monitoring tool and for the entire system using a power meter.

Doom (2016): Doom was benchmarked on ultra settings, using the Opengl API. Framerate was recorded with fraps.

As we can see from the fps/watt measurement, the efficiency rapidly increased going down to 1300 mhz, and then slowly climbed from there. The efficiency gain leveled off at about 1200 mhz, but picked up again going to 1000 mhz, likely due to the lowered memory voltage. Performance barely dips by 5 fps going from 1400 mhz to 1200 mhz — a negligible decrease, typically unnoticeable in real world situations. Only once we go below 1100 mhz and reduce the memory clock does performance begin to noticeably decrease.

Borderlands 3: Borderlands 3 was benchmarked with the inbuilt benchmark.

It’s a similar story with borderlands 3, efficiency rises considerably while performance barely dips as the clock speed is lowered. Again 1200 mhz seems like the best trade off while 1000 mhz memory starts to impact performance while being even more efficient.

For Honor: For Honor was benchmarked with the inbuilt benchmark.

For Honor repeats the trend of reducing power draw without a meaningful impact on performance, at least up to 1200 mhz. As usual, 1000 mhz really takes off with the efficiency.

Witcher 3: The performance of Witcher 3 was measured with fraps.

Witcher 3 solidifies the trend. The only note of interest with the Witcher 3 results is that 1100 mhz was slightly less efficient than 1200 mhz, although results are within margin of error.

The biggest reason for AMD’s power draw problem here was setting the reference clock speeds too far from the sweet spot on their cards. Polaris is capable of much greater efficiency. With the clocks dropped slightly the voltage requirements go down by substantial margins, and the performance impact is often minimal. Going from 1400 mhz to 1200mhz dropped power consumption by nearly 25% but only dropped performance by 10%. 1200 mhz and 1100 mhz were tied for the best efficiency. 1000 mhz was the most efficient out of all the frequencies tested, but the results here are skewed due to the necessary memory underclocking.

That performance fails to increase that much at the higher frequencies suggests to me that the card might be memory limited — It might be interesting to see if overclocking the memory can improve performance, which might be a test for another time.

Polaris had every last drop of performance extracted from it due to competitive pressures, making it quite hot and power hungry. The architecture of these chips is capable of much higher efficiency than how they were shipped to consumers. Maybe it would have been better to clock things slightly lower, and give a quieter, less power hungry card to the consumer. But maybe It wouldn’t have. Pc gamers often prefer performance to all other metrics, and the way that these gpus were shipped reflects that.

--

--

--

Programmer, web developer, computer hardware enthusiast. Interested in everything.

Love podcasts or audiobooks? Learn on the go with our new app.

Recommended from Medium

Basics of Git

Documentation as Code — Adventures with DocFx

How to start your tech startup for real estate

This Is What An Engineer Looks Like

Garnet Silver dishy wholesales Earring Red L-1.2in UK gift

Using job queues in Go for resilient systems

Understanding the structure of an object of class Phylo

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Thomas Hayes

Thomas Hayes

Programmer, web developer, computer hardware enthusiast. Interested in everything.

More from Medium

DSAR probing ‘attack’ from Princeton researchers

Giorgio Parisi: Ordering Chao

Joe Biden and Xi Jinping’s video call on 18 March 2022 — What they talked and what were agreed?

Release The Dover Stone