Up until this month, most people only knew AMD’s latest GPUs as Navi or RDNA 3, which are the codename and architecture upon which they are based upon. AMD announced the new Radeon RX 7900 XTX and 7900 XT in Las Vegas in front of dozens of press and streamed the presentation to the public as well. While there was a lot learned about the new GPUs in subsequent sessions, that will have to wait for another date. What AMD did talk about, however, was the world’s first chiplet-based GPU design and pricing and availability. Performance was roughly discussed, but I think many of those claims will have to be assessed in December when these GPUs hit the market.
AMD’s new Radeon RX 7900 XTX and 7900 XT are fairly similar GPUs, with both of them based on the RDNA 3 architecture. The Radeon RX 7900 XTX, AMD’s flagship GPU, claims a raw performance of 61 TFLOPS with a 5.3 TB/s chiplet interconnect, 24GB of GDDR6, and 58 billion transistors. Keep in mind that chiplet interconnect is necessary because AMD has split out the GPU into logic and cache. The main GPU die, also known as the graphics compute die (GCD) is where most of the actual GPU logic resides on a 5nm process node, while the memory cache dies (MCD) uses a slightly older 6nm process node, allowing AMD to optimize both for performance and cost while simultaneously manufacturing smaller dies which can help to improve yields.
Each GCD has a 300 mm2 die size, while each MCD is only 37mm. Better yields can mean better performance and cost, which may explain AMD’s pricing for the new GPUs. While the 7900 XTX has 24GB of memory, the 7900 XT features 20GB, which is still more than ample for most applications. AMD has also decoupled the front-end clock speed from the shader clock speed which it claims enables for a 25% power savings and a 15% frequency uplift in the front-end. The RX 7900 XTX has 96 RDNA CUs clocked at 2.3 GHz and has a 384-bit memory controller driving 24GB of GDDR6. The RTX 7900 has 84 of the same RDNA 3 CUs with a 2 GHz ‘game clock’ with a 320-bit memory controller driving 20GB of GDDR6.
In addition to raw compute capabilities, AMD is also leading the way with support for DisplayPort 2.1, which will enable a display link bandwidth of up to 54 Gbps and will enable 12-bit color. What does this translate to? Well, it translates to supporting 4K at 480 Hz or 8K at up to 165 Hz, both applications that do not exist today due to the existing display drivers and panel capabilities, let alone GPU horsepower. But AMD did, along with Samsung, tease a new 8K version of Samsung’s MicroLED Odyssey G9 curved 49” gaming monitor, which has traditionally shipped in 240Hz refresh rate in all of its previous iterations, and I have a feeling it may do so again, but at 8K. In addition to AMD’s new display engine, the new media engine delivers hardware support for AV1 encoding and decoding up to 8K60. Speaking of displays, AMD also says that it expects high-resolution DisplayPort 2.1 displays to hit the market from all the major gaming display manufacturers starting in early 2023, including the 8K Odyssey G9.
AMD claims up to a 54% improvement generation on generation for performance per watt, which coincidentally happened with RDNA 2 over the original RDNA. While AMD did not compare itself to Nvidia’s products at all, it did compare itself to the Radeon RX 6950 XT at 4K, which is where many publications have chosen to extrapolate the 7900 XTX’s performance against Nvidia’s RTX 4090 and 4080. AMD claims that the 7900 XTX is up to 70% faster than its 6950XT which it released as a refresh to compete with the RTX 3090 and 3090 Ti. Many people claim that this puts the RX 7900 XTX somewhere in the ballpark between an RTX 4080 and an RTX 4090 in terms of absolute performance, which is still quite a win for AMD when you consider power, which I’ll be talking about next.
The power and cooling
Ever since Nvidia announced the RTX 4090 and people started reporting melting power connectors, there has been a lot of talk about power. The RTX 4090 has a stated TDP of 450W, but has been observed by some reviewers to consume north of 600W under heavy loads. The new ATX 3.0 12VHPWR connector, which enables these new higher power levels, is unique to Nvidia in gaming today and while some power supplies use it, in many cases results in people using dongles. Some GPU manufacturers have already started to recall some of these connectors, and it hasn’t been a good look for Nvidia. So, naturally, AMD seized on this opportunity to talk about how its GPUs ‘only’ have total board power of 355W and 300W for the 7900 XTX and 7900 XT, respectively.
Not only that, AMD also talks about how its reference design cards for both GPUs ‘only’ require two 8-pin connectors and ‘only’ take up 2.5 slots on the motherboard. AMD also claims that while the 7900 XTX is considerably more powerful than the 6950 XT, it is only 11mm longer and 25W more board power, which Nvidia can’t claim with the 3080 and 4080. AMD has also leaned into the fact that its new GPUs don’t require power supply upgrades or getting rid of smaller form factor cases, or any modifications if you already have a high-performance GPU. This may help AMD win a big segment of users who aren’t interested in upgrading anything other than the GPU. However, this argument doesn’t work that well for new builds unless we’re talking about SFF PCs which likely won’t fit these new AMD GPUs either.
The AMD advantage
AMD also took the wraps off one of its secret weapons that is only possible when you have deep access to both the CPU and GPU, SmartAccess Video. AMD’s only competition in this space comes from Intel with its Deep Link capability, which, while similar, isn’t necessarily the same. AMD claims that by distributing the encode and decoded workloads between the Ryzen CPU and Radeon GPU, it can achieve an up to 30% uplift in workflows like 4K multi-stream transcoding.
AMD also announced the AMD Advantage program for desktops, which was previously a notebook-only program, combining the best of what AMD has to offer on CPU with the best that it has to offer on GPU in a single, deeply co-engineered product. I am writing this article on AMD and Alienware’s crowning glory of this laptop program, the Alienware M17 R5 AMD. It is by far the most powerful laptop I have ever used and one of the quietest and most enjoyable gaming laptops I have had the chance to use. The success of laptops like this are why I believe OEMs and AMD have gone and decided to create a desktop program as well. This program will pair AMD’s Ryzen 7950X with the Radeon RX 7900 XTX in desktops, paired with AMD’s latest Smart Technologies and Adrenalin Edition software. These systems will also require liquid-cooled CPUs, a minimum of 2TB of NVMe, 32GB of EXPO DDR5 memory, tool-less entry and 80 Plus Gold certified power supplies. At AMD’s event, I got a chance to see a few of these systems running AMD demos and games live, including Falcon Northwest, Origin PC, Xidax and Maingear.
Conclusion – pricing and availability
AMD’s go-to-market and how it competes with Nvidia ultimately boils down to the price-to-performance equation that many gamers will make in their heads (or reviewers will do for them). Nvidia’s stated MSRP for the RTX 4090 is $1600 while the RTX 4080 is $1200, both of these are within expectation for most gamers familiar with Nvidia’s recent pricing; however it is still historically very high. That is why AMD’s pricing for the RX 7900 XTX at $999 is considered much more palatable to consumers, as is the $899 price for the RX 7900 XT. That said, anyone willing to spend $1,000 on a GPU is also equally as likely to spend $1600, so I don’t really think it makes much of a difference to those buyers. The important pricing will be closer to the mid-range of the lineups where most gamers can afford to buy GPUs and how both companies will deliver in performance per dollar for gamers. Based on what we’re seeing today, it is very likely that it will be one of the most competitive climates for Nvidia that the company has ever seen. I also believe that AMD’s RDNA 3 family is part of the reason why Nvidia ‘unlaunched’ the RTX 4080 12GB model and will likely re-introduce it as an RTX 4070/Ti when the 7900 XT drops along with the 7900 XTX on December 13th.