[ad_1]
An in-depth analysis has been performed on AMD’s newest RDNA 3 GPUs and has delivered some surprising results. The report covers the Radeon RX 7900 XTX flagship GPU and its new midrange RX 7600 graphics card, analyzing how their different architectures can affect performance. It’s an interesting comparison, as AMD has chosen to use a chiplet design for its big Navi 31 GPU while opting for a monolithic design for Navi 33 used in the Radeon 7600. Additionally, Navi 31 has more of everything than Navi 33, allowing us to see the benefits and tradeoffs AMD has made by bifurcating its product line.
To set the table a bit, the Navi 31 is AMD’s flagship die utilizing a chiplet design manufactured on TSMC’s 5nm process. The whole package is 529mm squared, with the compute die measuring 300mm squared. There is 96MB of Infinity/L3 Cache spread across six Memory Cache Dies (MCD), and the interconnect between them is capable of 5.3TB/s of bandwidth. It also has 24GB of GDDR6 memory running on a 384-bit bus for 960GB/s of memory bandwidth. In contrast, the comparatively small Navi 31 is made on an older 6nm process, also by TSMC, and is just 204mm squared with a monolithic design. It has 32MB of Infinity Cache and uses 8GB of GDDR6 memory across a much narrower 128-bit memory bus for 288GB/s of memory bandwidth.
The two GPUs are quite close up to L2, then we see quite a gap developing as the 7600 offers improved latency.
Credit: Chips and Cheese
Despite these radical differences, the new analysis by Chips and Cheese shows the smaller Navi 31 die offers improved cache and memory latency compared with its big brother. The site says while it’s normal for a smaller die to provide these kinds of benefits over a larger die, it’s especially pronounced with RDNA 3. The most notable example is that their tests found the Radeon RX 7900 XTX took 58% longer to retrieve data from its Infinity Cache than the Radeon RX 7600. This is a notable increase in latency from RDNA 2 when comparing large and small dies, as there was only a 16.5% latency penalty between the 6900 and 6600 XT GPUs. They posit that the chiplet configuration may be the culprit for this increased latency.
A similar gap between the two GPUs exists for VRAM access, though it’s not as pronounced as with the Infinity Cache. Instead, it ends up being 15% slower for the bigger die, according to Tom’s Hardware. This is likely due to the formerly discovered scenario where it just takes the 7900 XTX longer to access its Infinity Cache than the smaller GPU. Memory latency might be less of an issue with the 7900 XTX, though as it has so much more cache, it might need to access memory less often than the 7600.
This analysis raises interesting questions about what could have been if AMD had gone with a monolithic die for the 7900 series GPUs. Its chiplet design has many benefits, including lower costs and better yields, but clearly, it is giving up performance in some areas, at least in these tests. Despite all that, the Radeon 7600 has probably been an underwhelming launch for AMD, as the GPU was met with a tempered reception by reviewers for having gone a bit too far into the cost-saving realm. The card fared so poorly with reviewers prior to launch that AMD dropped the price by $30 just before its unveiling, but many felt it should have gone even further to make it a better value.
[ad_2]
Source link
