An update for AMD's ROCm software spotted by @Kepler_L2 has purportedly revealed general specifications for AMD's yet-to-be-announced RDNA 3-based discrete graphics processors codenamed Navi 32 and Navi 33. The rumored core counts of the GPUs may give some clues regarding the performance of AMD's upcoming Radeon RX 7000-series offerings for high-end and mid-range market segments.
AMD's Navi 32 will feature 60 compute units (3840 stream processors), whereas the Navi 31 chip will pack 32 compute units (2048 stream processors) based on ROCm code exposed over at GitHub. To put the number into context, AMD's range-topping Navi 31 powering the company's Radeon RX 7900 XT and RX 7900 XTX graphics cards has 96 CUs, or 6144 SPs, in its full configuration.
Unfortunately, the number of compute units/stream processors does not exactly disclose the performance of actual graphics cards based on AMD's Navi 32 and Navi 33 GPUs. We must remember that there are factors like yields, clocks, memory bandwidth, power consumption and price points the company wants to address.
However, they certainly give us some food for thought. For example, with 62.5% of the core count of Navi 31, the Navi 32 could offer 60% to 65% of the Radeon RX 7900 XTX's performance at around the same clocks if it is not constrained by memory. Meanwhile, with a third of the core count of the top-of-the-line RDNA 3 chip, the Navi 33 could end up offering 30% to 35% of Radeon RX 7900 XTX performance at the same clocks and without memory bandwidth constraints.
Of course, since we have no idea about the performance of AMD's Radeon RX 7900 XTX from independent sources (such as our own Jarred Walton), we won't know how the cheaper Radeon RX 7000-series offerings will stack up against AMD's existing Radeon RX 6000-series products. However, assuming that the efficiency of RDNA 3's stream processors is higher than that of RDNA 2, it is safe to say that even the midrange Navi 33 will give its predecessor a run for its money.
Meanwhile, remember that the ROCm update reveals the maximum number of CUs and SPs that Navi 32 and Navi 33 feature. The actual GPUs may have fewer compute units and stream processors.