AMD RX 7900 XTX vs. RX 7900 XT: AMD’s monster GPUs, compared

Announced for the first time on November 3, 2022, the 7900 XT and 7900 XTX are up for grabs starting December 13, priced at $900 and $1,000, respectively. The pricing structure here is similar to the RX 6000 series, which had the $999 RX 6900 XT, then added a $900 GPU.

This could help bridge the gap between the higher-end RX 7000 GPUs with the top-end models; by contrast, the RX 6800 XT launched for $649, well below the 6900 XT’s price tag, with no GPU in between the two models.

Here’s a quick look at how these two models differ on the spec sheet.

Essentially, the 7900 XTX has the full Navi 31 GPU with all the bells and whistles, while the 7900 XT is slightly cut down. The XT model has 84 of the 96 Compute Units (or CUs) that Navi 31 has to offer, a clock speed of 2GHz as opposed to the 2.3GHz on the XTX, and 20GB of memory instead of 24GB GDDR6.

The Total Board Power (TBP) for the 7900 XT is also a little lower, of course, as is the clock speed. Overall, both GPUs are quite similar — the only real difference lies in the performance.

The 7900 XT and 7900 XTX are both powered by the new RDNA 3 architecture, which is radically different from RDNA 2, the architecture inside of RX 6000-series GPUs. The two GPUs share the same architecture, but it’s worth exploring when considering AMD’s latest and greatest.

Most importantly, chiplets, the technology that propelled Ryzen into first place in 2019, have finally come to GPUs. That’s big news, but the chiplet design for RX 7000 is in many ways very different from what we’ve seen with Ryzen and other AMD CPUs. Instead of having multiple dies for cores and one extra die for I/O and other functions, RX 7000 GPUs have one die with all the cores and several dies that each have one GDDR6 memory controller and 16MB of Infinity Cache. The chip with all the GPU cores is now called the Graphics Compute Die (GCD) and the chips with memory controllers and cache are Memory Cache Dies (MCD).

This means an RX 7000 GPU (specifically the 7900 XTX) can have up to 96MB of Infinity Cache and a bus width of 384-bit, which together allow the 7900 XTX to hit a peak bandwidth of 5.3 TB/s. Meanwhile, the 7900 XT appears to have one less MCD and thus has just 80MB of cache and a 320-bit bus, as well as 4GB less VRAM. Lower-end RX 7000 GPUs will likely use a different GCD and even fewer MCDs, but we’ll have to wait until those are announced next year for all the details.

RDNA 3 also comes with a new node. Last-generation RX 6000 GPUs were fabbed on TSMC’s 7nm node; now, RX 7000 GCDs are fabbed on 5nm, which has up to 1.8 times the transistor density of 7nm, as well as a 15% clock speed boost for the same power consumption or 30% less power at the same clock speed. Combined with architectural improvements, 5nm reportedly allows RDNA 3 to be 54% more efficient than RDNA 2.

Meanwhile, RX 7000’s MCDs are fabbed on TSMC’s 6nm node, which is basically 7nm, but with a focus on efficiency, low power, and low cost. Cache and I/O devices like memory controllers are increasingly more difficult to shrink with every new generation, which is why AMD is sticking with 6nm for these MCDs; 5nm is just too expensive while providing little improvement for this kind of stuff.