AMD Radeon RX 6800, 6800 XT, and 6900 XT aim at Nvidia's best

Welcome back to the high end, AMD.

For the first time since the days of the Radeon R9 290X—seven long years ago—AMD’s Radeon group is prepared to unleash enthusiast-level graphics cards that can do battle with Nvidia’s best and brightest. During the company’s second “Where Gaming Begins” event on Wednesday, AMD revealed a trio of Radeon GPUs based on a new RDNA 2 architecture, all armed with “revolutionary Infinity Cache” technology, sky-high clock speeds, and interesting synergies with Ryzen. They take direct aim at Nvidia’s RTX 30-series offerings, culminating in a $999 Radeon RX 6900 XT that seeks to seize the gaming crown from the monstrous $1500 GeForce RTX 3090 when it launches December 8—for $500 less than Nvidia’s beastly GPU costs.

The $579 Radeon RX 6800, meanwhile, will go toe-to-toe with the $500 RTX 3070 releasing tomorrow, and the $649 Radeon RX 6800 XT intends to challenge the $700 RTX 3080. Both launch November 18, about two weeks after AMD’s Ryzen 5000 CPUs hit the streets.

While Nvidia’s recent gaming GPU releases went light with raw memory capacity, opting for an upgrade to faster GDDR6X VRAM instead, AMD equipped every high-end Radeon RX 6000 card with a substantial 16GB of standard GDDR6, plenty to handle 4K gaming well into the future. And while Nvidia’s RTX 3080 and 3090 consume much more power than previous GeForce incarnations, AMD claims it’s achieved its lofty goal of a 50 percent performance-per-watt increase compared to the original RDNA architecture, which makes the Radeon RX 6000-series XT graphics cards less power-hungry than their Nvidia rivals.

Yes, you read that right: Radeon GPUs will need less power than competing GeForce cards. My, how the tables have turned.

We’ll start by diving into the raw specs and performance claims for the Radeon RX 6800, 6800 XT, and 6900 XT, because we know that’s what you’re here for, but stick around afterward when we dig into key changes in the RDNA 2 architecture and new Radeon ecosystem features, like Smart Access Memory that ties into Ryzen 5000 CPUs and AMD’s new Infinity Cache. Those provide crucial additional context for some of the slides you’re about to see.

Radeon RX 6800 vs. GeForce RTX 3070

Let’s start at the more affordable end.

The Radeon RX 6800 packs 60 compute units and a 1,815MHz game clock—90MHz higher than the rival GeForce RTX 3070’s rated Boost speed. More importantly, AMD’s card triumphs in memory capacity, packing a full 16GB of GDDR6 RAM versus the 3070’s 8GB. Nvidia pitched the 3070 as the ultimate 1440p graphics card, but it can do well in 4K games, too. That 8GB memory buffer probably won’t hold up well for 4K gaming as the years go on though, as some games already exceed it. The Radeon RX 6800 won’t suffer the same problem. Nvidia’s card enjoys a slightly lower total board power rating.

Now things are getting interesting. Vendor benchmarks should always be taken with a grain of salt, and you should always wait for independent reviews of pricey hardware, but AMD’s data tends to be reliable—and it shows the Radeon RX 6800 beating the GeForce RTX 3070 across the board here, at both 4K and 1440p resolution. Wow. Radeon chief Scott Herkelman told me that AMD’s lab used identical systems for all tests, changing only the graphics card, and that the company used the best-performing API for a given graphics card in all tests. So if Nvidia’s GPU performed better in a game using DirectX 11, and AMD’s exceled with DX12, these results reflect apex performance.

It’s not quite as cut-and-dry as it looks, however. See that tiny “+Smart Access Memory” next to the Radeon RX 6800’s name at the bottom? This graphs shows AMD performance with that feature active, which can provide an additional performance boost when you drop a Radeon 6000 GPU into a system with a Ryzen 5000 CPU installed. We’ll get into it more later, but these aren’t stock results for the RX 6800.

Radeon RX 6800 XT vs. GeForce RTX 3080

Stepping up the stack, the Radeon RX 6800 XT increases the compute unit count to 72, game clock speeds to a hair over 2GHz, and the total board power to 300 watts. It also offers a roomy 16GB of memory. By comparison, Nvidia’s GeForce RTX 3080 has a Boost clock nearly 300MHz slower (though it tends to run faster in reality) and just 10GB of memory, though it uses the faster GDDR6X variety for more bandwidth. The Radeon RX 6800 XT’s 300W power rating is 20W less than the RTX 3080’s.

In these results, AMD isn’t putting its thumb on the scale with Smart Access Memory. In its stock state, the Radeon RX 6800 XT meets or flat-out beats the GeForce RTX 3080 in 7 of the 10 benchmarks here at 4K resolution, where Nvidia’s card performs its best. Drop the resolution to 1440p, where Nvidia’s Ampere can’t take full advantage of its abundant shaders, and AMD opens its lead even more, triumphing in every game except Division 2 and Resident Evil 3. Hot damn.

Radeon RX 6900 XT vs. GeForce RTX 3090

Now we get to the battle of the (almost) Titans. The Radeon RX 6900 XT packs identical specifications to the Radeon RX 6800 XT except one, boosting the critical compute unit count from 72 to 80. Yes, it still runs over 2GHz, and yes, it still sticks to 300W of total board power—50W less than the power-sucking GeForce RTX 3090.

Nvidia’s marketing for the 3090 revolved around 8K gaming and the card’s massive, creator-friendly 24GB of GDDR6X memory. The company even calls the step-down RTX 3080 its gaming flagship—not the 3090. AMD is keeping the focus on normal gamers with the Radeon RX 6900 XT. This card comes with 16GB of standard GDDR6 memory, just like the other Radeon GPUs revealed today. It’s not as spacious and not as fast as the RTX 3090’s memory, but 16GB of GDDR6 should hold up well for 4K gaming over the long term nonetheless. Opting for a less extreme VRAM setup means AMD can keep the cost of the Radeon RX 6900 XT much lower compared to its rival.

So many people thought this was impossible, but here we are—at least in AMD’s supplied benchmarks. The Radeon RX 6900 XT beats or ties the GeForce RTX 3090 more often than not at 4K, and substantially so in Forza, with all games blowing well past the golden 60-frames-per-second mark even with their highest visual settings enabled.

Note that AMD is tipping the scales again here, and more so than before. Not only is Smart Access Memory active on the Radeon system, but so is Rage, AMD’s new automatic overclocking tool in Radeon Software. A victory is a victory, but you can’t help but wonder how much extra performance those features are adding to the top. A slide you’ll see later shows the duo improving performance by anywhere from 2 to 13 percent depending on the game, with an average uplift of 6.4 percent across eight games.

It’s also worth noting, however, just how much cheaper the Radeon RX 6900 XT is compared to the luxurious GeForce RTX 3090. $500 ain’t chump change.

That’s it as far as the Radeon RX 6000-series graphics cards themselves go. For a deeper look at the RDNA 2 architecture and new features inside AMD’s high-end GPUs—including Rage, Smart Access Memory, and Infinity Cache—keep reading.

RDNA 2 and Infinity Cache

AMD knew that to catch up to Nvidia, its GPUs needed to become much more efficient than they used to be. RDNA 1 achieved a 50-percent power improvement over previous Radeon GPUs, but a large part of the efficiency gains came from moving to the advanced 7nm manufacturing process. The RDNA 2 architecture stays on 7nm, so AMD’s engineers needed to boost its efficiency by another 50 percent in other ways.

It started with the compute units—the beating heart of the GPU. AMD’s graphics chief engineer, Laura Smith, said the company went through the design with a fine-tooth comb to root out inefficiencies: rejiggering data paths, rebalancing pipelines, and (in a page from the Ryzen design handbook) enabling “pervasive fine-grain clock gating,” which is an interesting twist I’m looking forward to hearing more about, as GPUs behave very differently than CPUs. The new compute units are 30 percent more efficient than before, which Smith says is crucial for scaling RDNA up to these enthusiast-class GPUs.

AMD’s engineers also managed to squeeze much higher clock speeds out of RDNA 2 on the performance side of things. Where the Radeon RX 5700 XT sported a 1,755MHz game clock, AMD says the Radeon RX 6800XT and 6900XT should sustain 2,015MHz speeds during gaming, and can even go to 2,250MHz in other tasks. (Smith actually says the Radeon RX 6900 XT’s power-per-watt increase versus RDNA 1 was even higher, at 65 percent.)

But the most intriguing—and mysterious—power gains come from the new 128MB “Infinity Cache” built into the Radeon RX 6000-series. Smith says the Infinity Cache is a high-density design that has its roots in Ryzen’s L3 cache decisions, but rearchitected for Radeon’s gaming needs. Details of what Infinity Cache actually does remain murky, but Smith says it basically allows Radeon RX 6000 graphics cards to deliver twice the memory bandwidth at lower power cost. Including Infinity Cache lets AMD stick to a smaller 256-bit bus for the Radeon RX 6000-series, which provides substantial power savings over moving to the larger bus sizes you normally see in high-end gaming cards. We can’t wait to hear—and test—more.

Ryzen synergy, Rage, and DirectX 12 Ultimate

Hardware can’t run without software, and part of AMD’s presentation was dedicated to showcasing some nifty new capabilities found in Radeon RX 6000 graphics cards. Most intriguing? AMD Smart Access Memory.

Like AMD’s Smart Shift technology in laptops (and the RDNA 2-based PlayStation 5), AMD Smart Access Memory is designed to give you better performance if you’re all-in on AMD’s hardware. If you have both a Ryzen 5000 CPU and a Radeon 6000 GPU in your system, AMD Smart Access Memory gives your CPU access to your graphics card’s full memory buffer, rather than the usual 256MB chunks. That direct communication cuts down on buffering and potential latency. Even though game makers haven’t had a chance to optimize for the closely guarded tech yet, Radeon chief Scott Herkelman told PCWorld that simply flipping this feature and Rage overclocking on simultaneously in AMD’s labs resulted in performance increases across a series of games, with Forza Horizon 4 running 13 percent faster.

Fascinating stuff. As with Infinity Cache, we’re looking forward to seeing deeper details, but several AMD-supplied benchmarks for the Radeon RX 6000 GPUs show results from numerous games with Smart Access Memory active.

AMD’s new cards will also support Microsoft’s upcoming DirectStorage API, which will let your GPU talk directly to your NVMe SSD for vastly improved loading times in games. Getting the various parts of your PC to talk directly rather than flowing through system RAM suddenly seems like the hot new trend for 2021. (Nvidia’s RTX 30-series also supports DirectStorage, using “RTX IO” branding.)

AMD is also spiffing up the one-click overclocking capabilities found in its Radeon Software. Going forward, you’ll find a new “Rage” preset tuning option that applies an automatic overclock tuned to your specific card—a helpful tool indeed for people who just want their games to play as quickly as possible with no headaches. Herkelman told me there are some slight differences compared to the previous automatic overclocking feature offered by Radeon software, but he didn’t get into specifics.

Fun fact: The “Rage” name is a shoutout to a popular line of graphics cards released by ATI in the early days of 3D acceleration, before AMD purchased the company. ATI’s “Fury” moniker made a comeback in graphics card form a few years back, and now Rage will be back on every Radeon-powered PC. I dig it.

Speaking of Radeon Software features, AMD also briefly teased how its existing “Radeon Boost” and “Radeon Anti-Lag” features combine to improve latency in games that demand quick responses, especially when paired with a FreeSync monitor. GeForce cards just gained the Nvidia Reflex suite of latency boosting features, so it’s clear that game responsiveness is becoming a new battleground for graphics card makers.

Finally, like the GeForce RTX 30-series, AMD’s Radeon 6000 graphics cards will be fully DirectX 12 Ultimate-compliant. Microsoft calls DX12 “a force multiplier for the entire gaming ecosystem” by unifying an array of new features—mostly ones introduced in Nvidia’s RTX 20-series, but largely ignored by developers—across all PC and next-gen Xbox hardware. “When Xbox Series X releases, there will already be many millions of DX12 Ultimate PC graphics cards in the world with the same feature set, catalyzing a rapid adoption of new features, and when Xbox Series X brings a wave of new console gamers, PC will likewise benefit from this vast surge of new DX12 Ultimate capable hardware,” Microsoft said when announcing the API.

Being DirectX 12 Ultimate-compliant means the Radeon RX 6000 graphics cards will support variable rate shading, mesh shading, sampler feedback, and yes, real-time ray tracing—though AMD didn’t go into details about its hotly anticipated ray tracing implementation.

AMD wasn’t entirely mum, though. Nvidia has a clear lead on ray tracing, and the company’s proprietary Deep Learning Super Sampling (DLSS) plays a big role. The AI upscaling feature leverages the dedicated tensor cores in GeForce GPUs to speed up frame rates, counterbalancing ray tracing’s performance impact—but remember, it’s an Nvidia-proprietary technology. AMD teased a solution of its own as part of the company’s open FidelityFX tools during today’s presentation. Tucked away in the corner of the slide, there was a small box that simply said “Super Resolution.”

Herkelman wouldn’t go into details, but he confirmed to PCWorld that this will be AMD’s open response to Nvidia’s DLSS, designed to work on any graphics hardware—even APUs—and across various platforms, because AMD’s RDNA 2 GPUs also power the next-gen PlayStation 5 and Xbox Series X consoles.

That broad adoption of AMD technologies could be an ace in the hole for Radeon’s ray tracing future. The new consoles will be getting off the ground at roughly the same time as the Radeon RX 6000 series, all bearing RDNA 2 GPUs. With the Radeon RX 6000 series, Xbox Series X, and even Nvidia’s RTX 30-series all driving DirectX 12 Ultimate forward in the same time frame as well, AMD is betting that DX12U’s features will become ubiquitous. Everyone is working on the same page now.

amd consoles — RDNA 2 isn’t just for the Radeon RX 6000 graphics cards.

As the graphics chip inside the consoles, AMD’s yet-to-be-revealed ray tracing implementation could become the standard target for game developers despite Nvidia’s two-year lead and impressive proprietary hardware. If AMD’s ray tracing becomes the standard, its open-source “Super Resolution” technology inside of FidelityFX could become the standard as well. DLSS 2.0 is black magic, but it’s locked to Nvidia GPUs on PCs, and requires custom code.

Of course, as with Infinity Cache and AMD Smart Access Memory, it depends on the details. AMD isn’t revealing anything more about Super Resolution today.

Bottom line, though? After seven long years, AMD’s RDNA 2-equipped Radeon RX 6000 graphics cards are finally ready to bring the fight to Nvidia at the very edge of enthusiast-class performance. We’ll need to see if AMD’s lofty claims hold up to scrutiny by independent testers, but we can’t wait to get our hands on these graphics cards. Competition is a wonderful thing—and long overdue at the high end. Welcome back, AMD.