Key Takeaways
1. RDNA 4’s performance and pricing did not meet expectations, and AMD needs to improve offerings in both low-end and high-end markets.
2. The upcoming RDNA 5 series is expected to be built on TSMC N3 nodes, with plans for high-end AT0 and mid-range AT2 dies.
3. The high-end RX 10090 XT model may offer significant performance boosts, potentially being 264% faster than the RTX 4080.
4. The mid-range RX 10070 XT is projected to be 120% faster than the RTX 4080, aiming for a price below $550.
5. All RDNA 5 models may lack Infinity Cache and could incorporate 3D stacking for the L3 cache in future designs.
RDNA 4 has not really wowed people with its performance boosts, and the pricing that AMD was counting on to be a strong point didn’t end up being as affordable as first promised, mainly due to worries about import taxes. The RX 9070 series is still a dependable mid-range option, but AMD needs to step up its game in both low-end and high-end markets. This could happen with the new RDNA 5 chips, as hinted at in a recent leak from Moore’s Law Is Dead.
Speculation on Specs
MLID always starts off by reminding us that the specifications from his reliable insider sources might change before the actual launch (which is also subject to change). So, it’s smart to stay a bit cautious and remember that these are AMD’s current intentions for a launch expected around 2027.
Upcoming Chips
According to MLID’s sources, which have reportedly been verified by several AMD teams working on the next-gen GPUs, the RDNA 5 series will be built on TSMC N3 nodes. This includes a top-tier AT0 die for high-end and HPC models and a mid-range AT2 die for the successor to the RX 9070 cards. There’s even a leaked table showing all the planned versions. MLID has also mentioned that AMD might be thinking about an AT3 die for entry-level models or APU chiplets, but this hasn’t been confirmed by all sources.
Specifications Overview
The AT0 die seems quite large, boasting up to 184 CUs or possibly even 200 CUs, as the table indicates that one work group is turned off per shader engine for the 184 CU variant. The primary two models likely target HPC and data centers, as they are expected to support PCIe 6.0 and include either 128 GB or 96 GB of VRAM, with prices likely above $5,000. Moreover, the table reveals an AI/ML model with 72 GB of VRAM and 184 CUs aimed at this HPC segment.
Gamers might be thrilled to learn about a high-end PCIe 5.0 gaming version with 154 CUs, 40 MB of L3 cache, 36 GB of GDDR7 36 Gbps VRAM, and a 384-bit bus width, which is set to compete with the upcoming RTX 6090. Sources from MLID estimate that this high-end gaming model, tentatively named RX 10090 XT, could be 264% faster than the RTX 4080, offering 1,728 GB/s bandwidth. Surprisingly, the total board power (TBP) is predicted to be around 380 W, and MLID suspects that the 264% performance increase over the RTX 4080 may actually be a bit lower.
Mid-Range Chips
For the mid-range segment, the AT2 die is expected to have as many as 70 CUs, but the table shows the top mid-range version with 64 CUs. This RX 10070XT model is projected to have a TBP of 275 W, a 192-bit bus, 18 GB of GDDR7 36 Gbps VRAM, and 864 GB/s bandwidth. Importantly, this model aims for the sub-$550 price range and is expected to be 120% faster than the RTX 4080.
Other models in this category include the 10070 GRE with 48 CUs, 15 GB of VRAM, and performance similar to the RTX 4080, while another 10060 XT model would have 44 CUs and 12 GB of VRAM, competing primarily with the RTX 5070.
MLID also notes that AMD might introduce additional models like the 10090 XTX, 10080 XT, or 10050, among others. Another noteworthy detail is that all RDNA 5 models appear to lack Infinity Cache, hinting that AMD might be considering 3D stacking for the L3 cache.
Source:
Link
Leave a Reply