AMD Radeon RX 6800 XT and RX 6800 Review: Nipping at Ampere's Heels

[harlan4096]

AMD's best GPUs to date are strong in rasterization but fall behind in ray tracing.

The AMD Radeon RX 6800 XT and Radeon RX 6800 have arrived, joining the ranks of the best graphics cards and making some headway into the top positions in our GPU benchmarks hierarchy. Nvidia has had a virtual stranglehold on the GPU market for cards priced $500 or more, going back to at least the GTX 700-series in 2013. That's left AMD to mostly compete in the high-end, mid-range, and budget GPU markets. "No longer!" says Team Red. 

Big Navi, aka Navi 21, aka RDNA2, has arrived, bringing some impressive performance gains. AMD also finally joins the ray tracing fray, both with its PC desktop graphics cards and the next-gen PlayStation 5 and Xbox Series X consoles. How do AMD's latest GPUs stack up to the competition, and could this be AMD's GPU equivalent of the Ryzen debut of 2017? That's what we're here to find out.

We've previously discussed many aspects of today's launch, including details of the RDNA2 architecture, the GPU specifications, features, and more. Now, it's time to take all the theoretical aspects and lay some rubber on the track. If you want to know more about the finer details of RDNA2, we'll cover that as well. If you're just here for the benchmarks, skip down a few screens because, hell yeah, do we have some benchmarks. We've got our standard testbed using an 'ancient' Core i9-9900K CPU, but we wanted something a bit more for the fastest graphics cards on the planet. We've added more benchmarks on both Core i9-10900K and Ryzen 9 5900X. With the arrival of Zen 3, running AMD GPUs with AMD CPUs finally means no compromises.

Update: We've added additional results to the CPU scaling charts. This review was originally published on November 18, 2020, but we'll continue to update related details as needed.

AMD Radeon RX 6800 Series: Specifications and Architecture

Let's start with a quick look at the specifications, which have been mostly known for at least a month. We've also included the previous generation RX 5700 XT as a reference point. When AMD fans started talking about "Big Navi" as far back as last year, this is pretty much what they hoped to see. AMD has just about doubled down on every important aspect of its architecture, plus adding in a huge amount of L3 cache and Ray Accelerators to handle ray tracing ray/triangle intersection calculations. Clock speeds are also higher, and — spoiler alert! — the 6800 series cards actually exceed the Game Clock and can even go past the Boost Clock in some cases. Memory capacity has doubled, ROPs have doubled, TFLOPS has more than doubled, and the die size is also more than double.

Support for ray tracing is probably the most visible new feature, but RDNA2 also supports Variable Rate Shading (VRS), mesh shaders, and everything else that's part of the DirectX 12 Ultimate spec. There are other tweaks to the architecture, like support for 8K AV1 decode and 8K HEVC encode. But a lot of the underlying changes don't show up as an easily digestible number.

For example, AMD says it reworked much of the architecture to focus on a high speed design. That's where the greater than 2GHz clocks come from, but those aren't just fantasy numbers. Playing around with overclocking a bit — and the software to do this is still missing, so we had to stick with AMD's built-in overclocking tools — we actually hit clocks of over 2.5GHz. Yeah. I saw the supposed leaks before the launch claiming 2.4GHz and 2.5GHz and thought, "There's no way." I was wrong.

AMD's cache hierarchy is arguably one of the biggest changes. Besides a shared 1MB L1 cache for each cluster of 20 dual-CUs, there's a 4MB L2 cache and a whopping 128MB L3 cache that AMD calls the Infinity Cache. It also ties into the Infinity Fabric, but fundamentally, it helps optimize memory access latency and improve the effective bandwidth. Thanks to the 128MB cache, the framebuffer mostly ends up being cached, which drastically cuts down memory access. AMD says the effective bandwidth of the GDDR6 memory ends up being 119 percent higher than what the raw bandwidth would suggest.

The large cache also helps to reduce power consumption, which all ties into AMD's targeted 50 percent performance per Watt improvements. This doesn't mean power requirements stayed the same — RX 6800 has a slightly higher TBP (Total Board Power) than the RX 5700 XT, and the 6800 XT and upcoming 6900 XT are back at 300W (like the Vega 64). However, AMD still comes in at a lower power level than Nvidia's competing GPUs, which is a bit of a change of pace from previous generation architectures.

It's not entirely clear how AMD's Ray Accelerators stack up against Nvidia's RT cores. Much like Nvidia, AMD is putting one Ray Accelerator into each CU. (It seems we're missing an acronym. Should we call the ray accelerators RA? The sun god, casting down rays! Sorry, been up all night, getting a bit loopy here...)

The thing is, Nvidia is on its second-gen RT cores that are supposed to be around 1.7X as fast as its first-gen RT cores. AMD's Ray Accelerators are supposedly 10 times as fast as doing the RT calculations via shader hardware, which is similar to what Nvidia said with its Turing RT cores. In practice, it looks as though Nvidia will maintain a lead in ray tracing performance.

That doesn't even get into the whole DLSS and Tensor core discussion. AMD's RDNA2 chips can do FP16 via shaders, but they're still a far cry from the computational throughput of Tensor cores. That may or may not matter, as perhaps the FP16 throughput is enough for real-time inference to do something akin to DLSS. AMD has talked about FidelityFX Super Resolution, which it's working on with Microsoft, but it's not available yet, and of course, no games are shipping with it yet either. Meanwhile, DLSS is in a couple of dozen games now, and it's also in Unreal Engine, which means uptake of DLSS could explode over the coming year.

Anyway, that's enough of the architectural talk for now. Let's meet the actual cards.

Meet the Radeon RX 6800 XT and RX 6800 Reference Cards

We've already posted an unboxing of the RX 6800 cards, which you can see in the above video. The design is pretty traditional, building on previous cards like the Radeon VII. There's no blower this round, which is probably for the best if you're worried about noise levels. Otherwise, you get a similar industrial design and aesthetic with both the reference 6800 and 6800 XT. The only real change is that the 6800 XT has a fatter heatsink and weighs 115g more, which helps it cope with the higher TBP.

Both cards are triple fan designs, using custom 77mm fans that have an integrated rim. We saw the same style of fan on many of the RTX 30-series GPUs, and it looks like the engineers have discovered a better way to direct airflow. Both cards have a Radeon logo that lights up in red, but it looks like the 6800 XT might have an RGB logo — it's not exposed in software yet, but maybe that will come.

Otherwise, you get dual 8-pin PEG power connections, which might seem a bit overkill on the 6800 — it's a 250W card, after all, why should it need the potential for up to 375W of power? But we'll get into the power stuff later. If you're into collecting hardware boxes, the 6800 XT box is also larger and a bit nicer, but there's no real benefit otherwise.

The one potential concern with AMD's reference design is the video ports. There are two DisplayPort outputs, a single HDMI 2.1 connector, and a USB Type-C port. It's possible to use four displays with the cards, but the most popular gaming displays still use DisplayPort, and very few options exist for the Type-C connector. There also aren't any HDMI 2.1 monitors that I'm aware of, unless you want to use a TV for your monitor. But those will eventually come. Anyway, if you want a different port selection, keep an eye on the third party cards, as I'm sure they'll cover other configurations.

And now, on to the benchmarks.
...

Continue Reading