By Joel Hruska
AMD’s RDNA2 GPUs land this morning, in the form of the Radeon RX 6800 ($579) and Radeon RX 6800 XT ($649). These two graphics cards are based on AMD’s RDNA2 GPU architecture, also known as “Big Navi.” Big Navi is an extension and improvement of the RDNA architecture that AMD debuted last year. RDNA2 is the GPU technology inside the Xbox Series S, Series X, and Sony’s new PlayStation 5, all of which launched this month.
On the PC side of things, AMD’s RDNA2 is facing off against Ampere. Nvidia’s latest RTX 3000 series debuted a few months ago, and the only serious complaint about the Nvidia GPUs has been their limited availability. AMD is fighting uphill in this comparison, but the company has promised it can deliver equal or better performance than the Ampere GPUs it’ll face off against.
We’ll be checking that hypothesis, at least at the high end of the market.ADVERTISING
Let’s Talk RDNA2
RDNA2 is a 7nm GPU design built on the same process node as RDNA, AMD’s first-generation Navi architecture. AMD has built multiple generations of chips on the same process node, but RDNA2 was billed as a significant architectural leap over and above RDNA, not just a widening of the original architecture.
According to AMD, RDNA2 is more efficient than RDNA, hits higher clocks, and incorporates unique first-ever features for a GPU, like a 128MB onboard cache. We’re still exploring what kind of performance it enables. One thing AMD told us: Running at low resolution isn’t expected to deliver a huge performance gain because the point of the IF is to reduce pressure on the 256-bit memory bus, and running in 720p or below doesn’t put enough meaningful pressure on it to meaningfully stretch the cache. Anyone with ideas on how to copy the entirety of Quake II or Quake III’s working data set into a 128MB GPU-mounted cache is invited to get in touch with me. Bonus points if we can guarantee loading the executable into a Zen 3’s 32MB L3.
Architecturally, the 6800 XT we’ll be reviewing today is a 4608:288:128 core with 72 ray accelerator units (one per CU). This is a bit below the Biggest Billy Goat Navi, the RX 6900 XT at 5120:320:128 and 80 ray accelerator units. We’ve previously discussed how RDNA improved performance over GCN when that architecture debuted last year — here’s a slide breaking down how AMD boosted RDNA2 performance over and above RDNA:
Design Frequency Increase means AMD leveraged resources from the Ryzen side of things when it came to designing for higher clock frequencies. It deployed transistors and other logic optimized for high performance in certain places, and it found ways to reduce latency and allow for higher clocks across the chip.
Improvements to power and gating improve overall consumption and, by association, thermals. The goal is to propagate clock signals across the GPU as efficiently as possible and to allow parts of the GPU to power back down when they don’t need to be turned on. Improving how efficiently power moves across the chip is critical to overall GPU efficiency.
Finally, there’s straight performance per clock. AMD is suggesting that 21 percent of their 54 percent uplift over RDNA came from general efficiency improvements within the GPU itself.
The Infinity Cache in particular has an interesting role to play here because data found within the cache will make it back to the GPU far faster than anything retrieved from GDDR6. During our engineering round table, AMD acknowledged that it didn’t have any new memory compression deployed within RDNA2, but that it had improved the implementation and applicability of its existing technology.
The new Infinity Cache has an impact on the overall cache architecture, as you’d expect. The 4MB L2 in Big Navi is the same size as in RDNA, and while we can’t really talk about saving die size or power with an extra 128MB of cache as opposed to an 8MB L2, AMD would have had to add additional L2 cache if it hadn’t come up with this unusual solution.
According to AMD, its compute units have been optimized to deliver 1.3x more throughput in the same power envelope.
AMD’s “Infinity Cache’ is a reference to how it connects to the GPU over the Infinity Fabric, not an attempt to claim AMD’s latest graphics card violates space-time principles. According to AMD, the cache distribution in the RDNA GPU wasn’t ideal:
AMD’s solution — well, one of AMD’s solutions — was to slap 128MB of cache down beside Big Navi, wire it up with 16x64B channels, and clock it at 1.94GHz. The cache provides ~2TB/s of bandwidth or nearly 4x the peak of GDDR6. Infinity Cache, according to AMD, is the secret sauce that allows it to compete so effectively with Nvidia and its RTX family.
There seems to be a period law of computing that allows us to slap cache down in new and exciting places every few years to good results. We’ve done L1, L2, L3, flirted with L4, built GPUs with L1, L2, and L3 caches of their own, and now AMD has gone and slapped 128MB of it directly on the graphics card. That’s how you know you’ve made it in the big leagues, in the PC industry — somebody finally comes along and slaps a huge amount of cache on you, just to see if anything interesting happens when they do.
Today, AMD is launching the Radeon 6800 and Radeon 6800 XT. Both GPUs are an attempt to claim an aggressive position against Nvidia, with the RX 6800 debuting at $579 versus a theoretical $499 for the RTX 3070, and the RX 6800 XT at $649 against the Nvidia RTX 3080 at $700. All of these prices are theoretical because of ongoing problems with bots, scalpers, and COVID-19.
The larger VRAM buffers on the 6800 and 6800 XT are an easy point of superiority over the Nvidia competition, but we’ll have to wait and see if those pay dividends in non-gaming workloads. For now, nothing we’ve tested is going to push those kinds of frame buffers.
Smart Access Memory
One feature we’ll be testing today is Smart Access Memory, also known as the first time AMD has claimed any kind of ability to tune performance between CPU and GPU due to owning the entire platform. This feature allows for memory transfers between CPU and GPU in blocks larger than 256MB, which can improve performance in some games.
Enabling SAM isn’t difficult — it required flipping two UEFI switches — and the performance boost from doing so is modest but real. It’s not a reason to buy an AMD card in and of itself, and Nvidia is claiming it can deliver an equivalent capability in an update, but it’s a good option to have.
Test Setup, GPU Configuration Issues
A large chunk of the time I’d set aside for working on the 6800 and 6800 XT review was focused on troubleshooting the GPUs instead. Initially, my testbed would not POST. Three days, 20 hours of work, one weekend, four motherboards, three UEFI updates, four CPUs, two CPU manufacturers, three SSDs, five sets of RAM, and four video cards later, I finally figured out the problem: The 6800 and 6800 XT are incompatible with the Dell UP2414Q.
I’ve used the UP2414Q as a testbed monitor ever since I wrote about it here. While it’s small, it offers support for three simultaneous display inputs (useful for a multi-system monitor) and its color reproduction is first-rate. A lot of my comparative DS9 work has been done on this panel. Unfortunately, it’s not compatible with the 6800 or 6800 XT and will not activate the monitor at POST.
When I brought this to AMD, the company assured me it’s virtually unheard of, affecting just my Dell panel and one other Acer model. Obviously, this isn’t an absolute guarantee — everyone is working off their own best knowledge — and I’m willing to believe that an early 4K panel could have some random bugs of its own. There are also some known issues with the LG CX OLED family, though AMD is working with LG to resolve them and expects the manufacturer to issue its own firmware update in the future.
We tested the 6800 XT on an Acer XB280HK and had no further problems.
There are several benchmarks where AMD’s 1080p and 1440p results are on top of each other. This is a known issue and the company is working on a solution. The implication here is that 1080p results aren’t as fast as they could be.
This review compares the RTX 3080 against the Radeon RX 6800 XT. Comparisons between the RTX 3070 and the RX 6800 will come in the near future. I wanted to test specifically on an AMD platform to evaluate SAM and the value proposition of an all-AMD platform — and due to the various issues I encountered, I wound up with less time to benchmark alternate cards than intended.
The RTX 2080 and Radeon VII data included here has been transcluded from our RTX 3070 review and the listed scores derive from those configurations. This is unlikely to have much practical impact — the RTX 2080 and Radeon VII are both last-generation GPUs, and I’m only providing them here for reference in the first place. The reason for not including the RTX 3070 the same way is that I’m specifically comparing that GPU against the 6800 in an all-AMD solution, and I didn’t want to clutter the situation any further than it already was.
Look for a follow-up on that comparison — with more lower-end GPUs — in the near future.
The RTX 3080 and Radeon RX 6800 XT were both tested on an AMD Ryzen 5900X running in an MSI X570 Godlike motherboard flashed with a beta UEFI provided by AMD. RAM was 32GB of Crucial DDR4-3600 memory in four DIMMs with a Corsair 2TB MP600 PCIe 4.0 SSD for storage.
Ashes of the Singularity: Escalation
Ashes of the Singularity: Escalation began life as a big AMD win but has flipped back and forth between Nvidia and AMD more recently. Interestingly, it’s not a win for SAM — overall performance falls back somewhat with Smart Access Memory enabled — but it does showcase the 6800 XT, slugging it out with the RTX 3080. AMD wins all three resolutions.
Assassin’s Creed: Origins
AC:O has always been an Nvidia-friendly title, as evidenced by the RTX 2080’s decisive tromping of the Radeon VII. In 1080p, this remains the case. Above 1080p, the RTX 3080 and 6800 XT are tied. SAM provides a modest improvement here.
BL3 is a new addition to our testing suite. Here, AMD again wins past Nvidia in all three benchmarks, taking the test entirely. SAM provides very small improvements. Not much new here as far as adding to our performance understanding, but AMD wasn’t kidding when it said it could take on Nvidia on its own turf.
Deus Ex: Mankind Divided
In Mankind Divided, the 6800 XT finally manages to shove its head above the 100 fps mark in 1080p (running 4x MSAA in this application incurs a wicked penalty as far as frame rates). There’s no impact from SAM, really, but we do see AMD holding the test.
Far Cry 5
Far Cry 5 doesn’t put much load on the GPU at 1080p, as evidenced by how scores bunch up. This is with the game’s HD textures loaded as well, so the game is tuned to maximum overall detail. Larger gaps appear at higher resolutions. Interestingly, SAM actually costs AMD a small performance hit in this game.
Final Fantasy XV
Nvidia has always led in Final Fantasy XV, and it’s not surprising to see that pattern hold up here. AMD can at least claim to have slashed the gap — at 4K, the RTX 2080 is 1.2x faster than the Radeon VII, while the RTX 3080 is only 1.1x faster than the RX 6800 XT. SAM is not very helpful to AMD in this benchmark either, but the performance rankings come out in the wash.
We have two Hitman II graphs: Miami and Mumbai. These benchmarks are basically a party for Smart Access Memory, which boosts the 6800 XT into a leadership position against the RTX 3080. It will be very interesting to see if Nvidia and Intel can copy AMD’s new trick for setting PCIe BAR size because if they can, Nvidia may regain this performance. Note that our figures for Hitman II reflect the “GPU” average logged in the benchmark report and not the “Overall” benchmark result reported briefly when the test is run.
We tested Metro Exodus in two modes: Extreme (no RT) and with ray tracing enabled. This is another area I’ll be visiting more extensively when we talk about the 6800 versus 3070 in the near future.
We test Exodus in Extreme detail, which sets 200 percent supersampling. Our effective test resolutions are therefore 4K, 5K, and 8K. One of these tests is run with ultra anti-aliasing enabled, the other is not. The Radeon VII doesn’t support RT in this title, so we don’t have results for that game run.
In standard Metro Exodus, the 6800 XT crushes it against the Radeon VII, with performance up 1.65x. The RTX 3080 still has an edge in 4K here, but it’s matched at lower resolutions.
Ray tracing changes our performance metrics here. Nvidia leads every metric except, oddly, our 4K test result on the 6800 XT (this may have simply been a blip). I only had time to run one ray tracing benchmark for this coverage, so I’m not going to draw any firm conclusions about how Nvidia and AMD stack up on this point — but it looks as though the situation between the two companies in ray tracing may track what happened with tessellation in the early 2010s, when AMD and Nvidia both had solutions in-market, but Nvidia’s was a bit beefier.
AMD’s ray tracing performance put it solidly above Turing by 1.5x – 2x, but behind Ampere on the whole. Interestingly, however, Ampere’s lead collapses at 4K, implying that something beyond sheer RT horsepower is causing a problem.
Shadow of the Tomb Raider
In SoTR, Smart Access Memory proves its worth again. AMD would win the text at 1080p and 1440p no matter what, but the 4K victory thanks to SAM is a nice cherry on the cake.
The RTX 3080 ekes out narrow wins in Strange Brigade, without much uplift from SAM for the Radeon family. Normally we’ve used this test for Vulkan, but we’ve had trouble with Vulkan for this review, so we fell back to DX12 for this test. Not much to see here one way or the other.
Note: Total War II: Warhammer is tested in DX11 on NV cards and DX12 on AMD cards. I normally don’t split APIs like this, but NV + DX11 is the clear best choice for that card, while DX12 is the clear best-pick for AMD.
Warhammer II shows some interesting patterns. The Radeon VII and RDNA2 hit the exact same 1080p score while the RTX 3080 leaps ahead, implying there’s something fundamental blocking performance. The RTX family has no such problem. The 6800 XT doesn’t get much boost from SAM here and Nvidia sweeps the test.
Without SAM, the Radeon RX 6800 XT has an average of 128.84 fps at 1080p, 108.25 fps at 1440p, and 67.09 fps at 4K. The RX 3080 averages at 134.86, 108.26, and 67.77 fps respectively, which gives it a small lead over AMD at 1080p and a tie position at 1440p and 4K.
SAM changes this situation somewhat. While the boosts are small — roughly 7 percent, 4 percent, and 3 percent, respectively — they do make enough of a difference to pop the 6800 XT over and above the RTX 3080 across our suite of tests. If Nvidia can replicate the advantage AMD has introduced, things will shift back towards Team Green. Otherwise, AMD can fairly claim to have matched its goals.
The big question for AMD is whether the overall driver and software support situation is equivalent to what Nvidia can offer. Historically, AMD drivers have had more problems than their Nvidia equivalents. Problems like the black screen of death last year caused issues for a lot of gamers, including yours truly.
The RX 6800 XT offers competitive performance with Nvidia’s RTX 3080, along with far more VRAM (very useful in certain circumstances). The tradeoff looks to be in ray tracing, where the card is better than Turing but falls behind Ampere.
If this conclusion sounds a little preliminary, that’s because it is. AMD has done an excellent job delivering a hardware followup to RDNA2. The few power consumption measurements I got in suggest a lead over the RTX 3080 as well, which isn’t easy to deliver when you’re working with an extra 6-8GB of RAM to power. Features like Infinity Cache have interesting long-term potential.
I’ll have more to say about where I feel like the card fits once I’ve finished with the 6800 and spent some more time with various Radeon software features. For now, I’d say AMD has re-established overall gaming competitiveness with Nvidia while adding its own ray tracing solution that offers meaningfully improved performance over Turing even if it isn’t quite up to par with what Ampere is bringing to the table. My problems bringing my testbed online colored my initial impressions of the card, and I want to poke at everything a bit more. Right now, I’d say Big Navi offers slightly higher rasterization performance on an all-AMD system in enough cases to make interesting, with some tradeoffs in ray tracing where workloads favor Nvidia.
More to come on this topic, including 6800 results versus the RTX 3070.