GPU Boost 5.0

While NVIDIA doesn't specifically call Ampere GPU Boost "5.0", there are some minor changes compared to Turing.

NVIDIA's GPU Boost was introduced with the Kepler architecture (GTX 680), Maxwell introduced GPU Boost 2.0, Pascal upped the version to 3.0, and Turing ran at 4.0. NVIDIA GPU Boost works by taking into account several parameters reported by the card to dynamically increase the GPU frequency above base clock. The clock increase has been especially huge with Turing: RTX 2080 Ti ran 1350 MHz base clock, was rated for 1635 MHz Boost (+21%), and ended up at 1824 MHz actual measured average frequency in our tests (+35%).

The GeForce RTX 3080 Founders Edition ticks at a base clock of 1440 MHz with rated boost set to 1710 MHz (+19%), achieving an average gaming frequency of 1931 MHz (+34%), which is quite similar to Turing. I'm still surprised that rated boost is so much lower than what is a realistic average frequency. This causes confusions among our readers from time to time; I guess NVIDIA is just playing it safe, taking no risk of overpromising.

For the chart below, I recorded the GPU frequency at a given temperature and plotted it in a scatter plot. The points are semi-transparent, so you have some indication of which temperature-frequency combinations are most active.



As you can see, there are five groupings. While temperature goes up, the GPU frequency will be reduced by 15 MHz for every five degrees Celsius. This 15 MHz interval is fixed, and the drop happens at round 70, 75, and 80 °C—it's not a fine-grained, smooth curve. Once you go above the 83°C thermal limit, the card thermal throttles, which has it reduce clock speeds more aggressively. During normal testing, I couldn't reach this state—yes, unlike the 2080 Ti FE, there is no thermal throttling on the RTX 3080 FE. For the plot above, I had to play with the fan speed, and even manually jammed the fans.

Using your favorite overclocking software, you may increase the 83°C thermal limit up to 88°C.

Overclocking

Overclocking on "Ampere" seems to be identical to "Turing". The cards will sit in their power limit most of the time, also during overclocking. This makes it more difficult to dial in specific clock speeds because when in the power limit, the GPU frequency will be lower than with a light load. If you tune your maximum OC to be stable with heavy loads only, your card will crash with light loads, as Boost will run the card at higher, unstable clocks.

The Founders Edition has a manual power limit adjustment range of up to 370 W, from a default of 320 W. This can help garner a little bit of extra performance, but it's not a big difference, maybe 1% in real-life performance.

The fan-adjustment range is 30% to 100%, same as previous NVIDIA Founders Editions. Adjusting the voltage limit is possible too, within a small margin that is only active when the card is not power-limited. This yields no significant improvements, which means it is usually not worth the time. Undervolting is not possible.


GDDR6X memory comes with an error-detection mechanism that is active on the GeForce RTX 3080. On previous generations, the "ECC memory" feature was only available for much more expensive, professional cards. AMD has enabled it on several Radeon cards, too. While GDDR6X error-detection isn't identical to ECC, it is similar; the memory controller is able to detect most transmission errors, and Error Detection and Replay (EDR) functionality will keep retrying that memory transfer until it succeeds.

For memory overclocking, this means you can no longer rely on crashes or artifacts as indicators of the maximum memory clock. Rather, you have to observe performance, which will reach a plateau when memory is 100% stable and no transactions have to be replayed. Replaying memory errors reduces overall memory performance, which will be visible in reduced FPS scores. For the following OC testing, I took this into account, of course.



With manual overclocking, maximum overclock of our sample is 1320 MHz on the memory (11% overclock), which is pretty impressive for a brand-new memory technology—I'm sure it won't be long before we see 20 Gbps or 21 Gbps GDDR6X officially. The maximum stable GPU overclock was 1855 MHz Boost, which increases average GPU clock from 1931 MHz to 1947 MHz (1% overclock).

Overclocked Performance

Using these clock frequencies, we ran a quick test of Unigine Heaven to evaluate the gains from overclocking.


Actual 3D performance gained from overclocking is 3.9%.