New leak suggests AMD is planning high-end RDNA 5 gaming GPUs to compete with upcoming Nvidia RTX 6090

RDNA 4 did not impress too much with performance gains, and the pricing that was supposed to be AMD’s ace up the sleeve did not turn out to be as low as initially advertised with all the uncertainties around the import taxes. The RX 9070 series remains a solid mid-range product, but AMD really needs to compete on the low-end and high-end as well. This should be possible with the upcoming RDNA 5 chips, as revealed in a fresh leak from Moore’s Law Is Dead.

As always, MLID opens with the disclaimer that the specs provided by his trusty insider sources are subject to change between now and launch times (including launch times). As such, we should remain moderately skeptical and keep in mind that these are Team Red’s current plans for a launch happening sometime in 2027.

According to MLID’s sources that have apparently been confirmed with multiple AMD teams working on upcoming GPUs, the RDNA 5 family of gaming GPUs wlil be fabricated on the TSMC N3 nodes and includes a flagship AT0 die for high-end and HPC models, as well as a mid-range AT2 die for the successor of the current RX 9070 cards, and there is a leaked table with all the planned variants. MLID mentions that AMD might be planning an AT3 die for entry-level models or APU chiplets, but this has not been corroborated with all sources.

High-end RDNA 5

The AT0 appears to be a massive die, with up to 184 CUs or even 200 CUs, since the table notes that 1 work group is disabled per shader engine for the 184 CU model. The top two models are targeting HPC and data centers most likely, as they also feature PCIe 6.0 support and pack 128 GB or 96 GB VRAM, and they will most likely cost upwards of $5,000. Additionally, the table shows an AI/ML model with 72 GB VRAM and 184 CUs that should fit this HPC category.

However, gamers would be excited to find out that there is also a high-end PCie 5.0 gaming model with 154 CUs, 40 MB L3 cache, 36 GB GDDR7 36 Gbps VRAM and 384-bit bus width that is slated to compete with the upcoming RTX 6090. MLID’s sources estimate that this high-end gaming model with RX 10090 XT placeholder name is 264% faster than the RTX 4080, and features 1,728 GB/s bandwidth. Surprisingly, the TBP is estimated at 380 W, and MLID believes the 264% performance gains over the RTX 4080 might end up slightly lower.

Mid-range RDNA 5

On the mid-range side, the AT2 die is supposed to have up to 70 CUs, but the table lists the top mid-range model with 64 CUs. This so-called RX 10070XT model would have 275 W TBP, 192-bit bus, 18 GB GDDR7 36 Gbps VRAM and 864 GB/s bandwidth. Most importantly, nonetheless, it would target the sub-$550 price bracket and it’s expected to be 120% faster than an RTX 4080.

Other models in this die category include the 10070 GRE with 48 CUs, 15 GB VRAM and around the same performance of an RTX 4080, while a 10060 XT model would feature 44 CUs, 12 GB VRAM and would compete with an RTX 5070 at most.

MLID notes that AMD may launch other models as well, like the 10090 XTX, 10080 XT or 10050 etc. Another interesting point is that all the RDNA 5 models seem to be missing Infinity Cache, which suggests that AMD could use 3D stacking for L3 cache.