Just yesterday we took a look at the AMD reference design Radeon RX 7900 XTX and RX 7900 XT. Today the review embargo for the custom-design partner cards expires and we have two reviews for you: ASUS RX 7900 XTX TUF and XFX Radeon RX 7900 XTX Speedster Merc 310 OC. There's more reviews in the pipeline, including of XT versions, so stay tuned.

AMD's new RDNA 3 architecture brings with it the world's first GPU based on chiplet technology. Why is that a big deal you ask? Making large chips is expensive, more expensive than making several small chips. AMD realized that years ago and built their Ryzen CPUs using the chiplet approach, which is the foundation for the company's tremendous comeback in the CPU space. Team Red is betting that the same can happen in the GPU world, and today we're testing their first products built using that philosophy. Using chiplets gives another major advantage—you can combine multiple production processes. For the case of the Navi 31 GPU that powers the Radeon RX 7900 Series, the central compute die is fabricated on TSMC's leading 5 nanometer node, because efficiency greatly matters for its design. On the other hand, the memory-cache dies don't put out as much heat, and contain analog technology, which doesn't scale as well with process size. That's why AMD decided to build those with 6 nanometer tech, making them cheaper to produce.

The RX 7900 XTX is AMD's flagship for this generation—it comes with the full Navi 31 GPU: 6144 cores, 96 compute units, 24 GB GDDR6 and six MCDs with 96 MB of L3 cache. RDNA 3 also introduces an upgraded display engine, which has support for DisplayPort 2.1, for higher refresh rates on upcoming 4K and 8K displays, and you also get support for hardware-accelerated AV1 encoding—the video format of the future.

For this round of reviews I've switched my testing to a Core i9-13900K, replacing our aging Ryzen 7 5800X that served us well for many years. Originally I wanted to switch over the holidays, but constant feedback on how "outdated, slow, and terrible" the 5800X is (it is not), made me switch early—just last week. So in addition to working on the RDNA 3 reviews, I also had to retest all my cards on the new setup, which is why there's fewer comparison cards in this review than what you're used to. I'll keep retesting, and update this review with data for the remaining GPUs. All the important high-end cards are included right now, so the outcome won't change in any way, it will just give you a broader overview.

Averaged over our whole 25-game test suite at 4K resolution, with RT off, the factory-overclocked XFX Radeon RX 7900 XTX has a 3% performance lead over the AMD reference design, which doubles the performance uplift to RTX 4080 from 3% to 6%. While a few percent here and there isn't "much," it's still a welcome improvement. The gap to RTX 4090 shrinks to 19% now. Compared to the RTX 3090 Ti, the XFX card is 22% faster and the increase over the RTX 3090 is 37%. Compared to last generation's RX 6900 XT, the XFX Merc 310 is a whopping 52% faster. The differences between individual games are huge, in some titles the XTX is 25% faster than the 4080, in others it's 10% slower. I've added a new chart at the end of the "Relative Performance" page, to break that down for you.

It's also possible that the press driver isn't fully optimized for all our games yet. RDNA 3 introduces new dual-issue compute units, which require special code optimization, so that they can achieve the +100% performance uplift. In briefings AMD has made it clear they have been optimizing the driver for the new units, and I'm sure a lot of work has already been done in the compiler, but I'm just as certain that there's some cases where hand-optimization can yield further benefits. During testing I also encountered crashes in AC:Valhalla and Elden Ring, no doubt these will be fixed soon.

With those performance characteristics, RX 7900 XTX is a formidable choice for gaming at 4K, with maximum details and 1440p at high-refresh-rate. You can crank up everything and you'll still run at over 60 FPS. Things are different when you enable ray tracing though, here the RX 7900 XTX is considerably weaker than what NVIDIA offers. On average (new chart in the RT section), the RTX 4080 is around 15% faster than the RX 7900 XTX with ray tracing enabled, which isn't monumental, but definitely more than what I would have expected. I think everyone agrees that ray tracing is the future, and just disagrees on how quickly that future is happening. If you're part in the "I want this now" camp, then you should probably consider the RTX 4080, or RTX 4090. On the other hand, if you feel like ray tracing is just minor additional eye candy, that comes with a huge performance hit, then you can happily grab the RX 7900 XTX. That's not to say that AMD's new cards are useless with ray tracing, but if you consider the differences in price and RT performance, then the value-proposition of both cards is virtually identical, with NVIDIA RTX 4080 giving you the higher overall performance.

The XFX custom design graphics card look fantastic, thanks to an understated smooth look. Under the hood, the cooling capabilities has definitely been upgraded. In our apples-to-apples heatsink comparison test at normalized noise and power output, the XFX Merc 310 cooler can achieve a temperature reduction of 4°C. However, when running at the default settings, which are factory overclock and small voltage increase of 80 mV, this advantage exaporates due to the higher heat output and the card is basically same temperature and same fan noise as the AMD reference card, with higher performance though. Temperatures are super low with 59°C, but noise levels are a tad bit high with 38 dBA. I feel like allowing a little bit higher temperature would have definitely allowed a significant reduction in fan noise levels. Of course it's difficult to justify such an approach if it means running warmer than the AMD ref card. That's why a lot of companies have a dual BIOS on their card for "quiet" mode. XFX is using the dual BIOS differently. The second BIOS is a "Maximum Power" BIOS that increases the power limit from 327 W to 339 W, on a 3x 8-pin card that could pull 525 W in power. A 12 W (or 3.6%) increase is really laughable. What makes things worse is that the fan settings are extremely aggressive, the card now runs at a loud 45 dBA. We've also tested the ASUS TUF today, which has a much bigger cooler, but ends up whisper quiet (28 dBA) with its quiet BIOS active—a huge difference. Just like the AMD cards, the XFX Merc 310 OC includes the idle-fan-stop capability that shuts off the fans when not gaming.

Power efficiency of the new Radeons is fantastic, clearly much better than the previous generation of RDNA2 and NVIDIA Ampere cards. As mentioned before, XFX has overclocked and overvolted their card, to unlock higher performance. This means that power consumption is increased, by roughly 50 W, which is perfectly fine and makes sense, given that this is a 3x 8-pin design. This move does lower energy efficiency by roughly 10% though, which is unfortunate, but still acceptable I'd say. Just like the AMD board, as the card heats up, the frequencies will drop by a lot. In our thermal load test, the card starts out running at 2754 MHz, and stays in that state for around 20 seconds, good to get a boost on short running benchmarks, but then clocks go down to 2552 MHz and stay there until the card cools down again at the end of your gaming session. This 7.4% drop is clearly significant and costs AMD against the RTX 4080, which loses only 1% in the same test.

We measured a shocking power consumption result for multi-monitor and media playback. Here, just the graphics card alone consumes 101 W and 89 W, respectively. This is way too high, RTX 4080 uses only 20-23 W in the same scenario, even the last generation RDNA2 cards were less than half that with 40 W. This can only be some sort of driver bug, because it basically disqualifies the new Radeons for multi-monitor use. Remember, this is idle sitting at the desktop, not gaming. Wasting that much power is simply a big no-no, especially in these times. AMD has had a long history of drawing a lot of power in these power states, so I'm not 100% convinced this really is so easy to fix. I also find it hard to imagine that nobody at AMD tests multi-monitor power draw, so in some meeting somewhere, someone decided "we will release it like that."

Overclocking worked so much better on the XFX Merc 310 OC than on the AMD reference board, thanks to the 3x 8-pin power inputs. While the AMD very quickly ran into its power limit, even at the +15% power setting, the XFX card has more headroom, which allows higher overclocks. Overclocking is still complicated, but I have to admit it was fun seeing all the pieces come together to unlock an overclock that's better than what we usually see in our reviews.

AMD's RX 7900 XTX reference card comes at $1000, XFX is asking another +$100 for their Merc OC. The biggest selling point of the card is that it comes with a triple 8-pin power input, which ensures the card isn't as power starved as AMD's reference card. With manual overclocking you can clearly reach higher performance levels than what's possible with a 2x 8-pin model, but that comes at the cost of lower efficiency. I also feel like +$100 is quite a lot, because it brings the card so much closer to the $1200 RTX 4080. Even with XFX's factory overclock, the RTX 4080 is still faster in ray tracing, and more energy-efficient too.