Ever since AMD bought ATI, gamers have asked if there was an intrinsic benefit to running an AMD GPU alongside an AMD CPU. Apart from some of the HSA features baked into previous-generation AMD APUs and a brief period of dual graphics support, the answer was always “No.” From 2011-2017, AMD simply wasn’t competitive enough in gaming for the company to invest in that kind of luxury concept.
AMD’s RX 6000 GPUs will be the first cards that can specifically take advantage of platform-level features inside the 500-series chipset. We’re going to talk more about that specific feature and several others later on, but it’s one of the most interesting things AMD discussed today, and I wanted to get it on the board.
Before we go deeper on new features, let’s talk about the new cards. Click on images to enlarge them; all images below are from AMD’s launch event.
Meet the RX 6000 Series
AMD is launching three new GPUs: Radeon RX 6800, Radeon RX 6800 XT, and Radeon RX 6900 XT. Here are the relevant specs on each:
The Radeon 6800 is a 3,840-core GPU with an 1815MHz base clock and a 2.105GHz boost clock. It features 128MB of Infinity Cache (more on that shortly), 16GB of GDDR6, and 250W of total board power. Like the other two GPUs today, it uses a 256-bit memory bus (more on that shortly). Total board power is 250W, including VRAM.
AMD has positioned the 6800 well above the RTX 3070’s $499 launch price, so the GPU will need to demonstrate this kind of lead in our own testing to carry the price point. Features like 16GB of VRAM may help with that, though we’ll have to see if the extra RAM is useful at any practical detail levels the GPUs can reach. (It may be useful for AI GPU upscaling, where VRAM is worth its weight in gold.)
Note that this RX 6800 was tested using Smart Access Memory. This is AMD’s new technology that leverages the 500-series motherboard platform to give the CPU full access to GPU RAM, rather than limiting the window to 256MB. This supposedly improves performance somewhat, even on unoptimized titles. AMD is leveraging it to compete as well as it is against the RTX 2080 Ti. Just something to keep in mind.
Next up, the Radeon RX 6800 XT:
The RX 6800 XT offers 72 compute units (4,608 cores), a 2015MHz game clock, 2250MHz boost clock, 128MB Infinity Cache, 16GB of GDDR6, and 300W of total board power consumption. Performance-wise, it’s expected to compete against the RTX 3080. When you check these numbers, note that AMD is not using Smart Access Memory to show these results:
As for 4K, you can see those figures below:
Eyeballing the graph, the ratios are mostly the same, but Nvidia gains ground on AMD in Doom Eternal, Resident Evil 3, and Wolfenstein: Young Blood for sure. I’m less certain of the others, due to the off-angle comparison, but it’s something we’ll check on when we get cards. AMD also took some pains to point out that this GPU draws just 300W to Nvidia’s 320W. Price? $649.
Finally, there’s the Radeon 6900 XT:
As rumored, the 6900 XT is 80 CUs (5,120 cores), with the same 2015MHz base clock, 2.25GHz boost clock as the 6800 XT. It also packs the same 16GB of VRAM, the same 26.3B transistors (all three chips are obviously using the same chip design), and a price tag of $999. AMD is claiming an absolute uplift in performance per watt of 1.65x, over and above the 1.5x target it set for RDNA2. This implies AMD is binning the cards aggressively, as it did with Radeon Nano.
Note that in this set of comparison figures, AMD is explicitly activating both Smart Memory Access and a one-touch overclocking feature called Rage Mode. With Rage Mode enabled and on its preferred platform, the RX 6900 XT can pace the RTX 3090, even outperforming it in spots. If we didn’t have these features in place, the performance gaps would presumably be larger. The flip side to that, of course, is that the RTX 3090 has an MSRP of $1,500, where the Radeon 6900 XT has an MSRP of $1,000.
AMD’s Special Features
If you’re familiar with high-end GPU design, you’re probably wondering why AMD is building its highest-end chips with just 256-bit memory buses. The answer is a new feature AMD built into RDNA2 dubbed Infinity Cache. We don’t have much detail yet on how the large cache structure is allocated, but the company did show some information on how it compares with using a wider memory bus:
Evidently, it’s more efficient to deploy a large cache backed by a smaller VRAM bus than to simply deploy more GDDR6.
The company credits techniques like Infinity Cache, along with fine-grain clock gating, pipeline rebalancing, and redesigned data paths for boosting the overall performance of RDNA and delivering a total 1.54x uplift in performance-per-watt. Sustained clocks have supposedly improved ~1.3x over and above standard RDNA.
Smart Access Memory is a feature that only works with 500-series chipsets, but allows the Ryzen CPU to access all 16GB of GPU VRAM, rather than being limited to the standard 256MB aperture size. This reportedly allows for more efficient data allocation in VRAM and improves overall performance.
Rage Mode, referenced above, is a one-click overclocking option that will need to be a great deal better than any previous one-click overclocking option I’ve ever tested in order to pay dividends. Between Rage Mode and Smart Access Memory, AMD believes it can boost the baseline performance of the 6800 XT by a fair bit.
Game speed improvements range from 2 percent to 13 percent, with an average performance uplift of around 6.4 percent.
AMD is also continuing to expand its library of FidelityFX features:
Those are the major announcements from the event. Obviously, AMD has thrown down something of a gauntlet here. The RX 6900 XT and RX 6800 XT are both priced below their Nvidia counterparts, while the RX 6800 comes in somewhat above the RTX 3070. AMD clearly believes it’s got a strong position with this part.
The stage is set for two major showdowns in the next few weeks in both the CPU and GPU markets. This is going to be downright interesting. As always, take all manufacturer benchmarks with a grain of salt, though AMD’s performance claims do broadly line up with where we expected the company to fall versus Ampere. The big question, of course, is whether these cards will actually ship to consumers in significant numbers, or if they’ll all end up on eBay.
Now Read:
- AMD Said to be in Advanced Talks to Buy Xilinx
- AMD Wants to Prevent Bots and Scalpers From Wrecking Ryzen 5000, Radeon Launches
- AMD’s Mark Papermaster Dishes the Goods on Zen 3
No comments:
Post a Comment