AMD Ryzen 3D V-cache CPUs surprisingly show big benefits on Windows 11 in a new test

Yesterday we reported on the purported specifications of Intel's next-gen Nova Lake-S desktop CPUs that are rumored to pack massive amounts of cache, even more than what we have on AMD's Ryzen X3D processors.

We already know that AMD manages to stay ahead of Intel CPUs in gaming thanks to that massive level 3 (L3) last-level cache (LLC), and so we expect Intel to overtake current-gen AMD 9000X3D parts.

However, while AMD has always highlighted the benefits 3D V-Cache brings in terms of gaming, the company has never shown that in particular AI workloads. Now a new test has revealed the potential for huge cache-based processors as AMD's 9000X3D parts show huge performance gains over their non-X3D counterparts.

The performance was evaluated by a Korean site called GiggleHD, and it used the X3D Rag Benchmark to test the CPUs. Both AMD and Intel parts were included in its assessment and the testing was done on Windows 11 25H2. Hence performance on Linux may vary.

For those wondering, the x3d-rag-benchmark measures how cache can improve Retrieval-Augmented Generation (RAG) performance by leveraging large L3 cache when it's available. In RAG pipelines, queries are embedded on the GPU, then passed to the CPU for vector search before the GPU generates a final response. Since vector search involves random memory access patterns, the cache size becomes very crucial, similar to how DRAM cache improves random performance on SSDs.

In essence, the large L3 cache in X3D processors reduces the reliance on RAM lookups, which are generally much slower than CPU cache, thus increasing hit rates and lowering latency. This translates into faster query throughput, reduced time-to-first-token (TTFT), and better handling of concurrent searches. Hence, much like in gaming workloads, the cache advantage provides measurable gains for RAG workloads too.

First we have Batch Search measured in queries per second (QPS). The octa-core 3D parts manage to outdo the 16-core 9950X3D, indicating that cache and boost clocks probably matter more here. This is shown by the Intel Core Ultra parts too, as the higher-clocked 285K gets ahead of the 270K Plus.

In index building, measured in vectors per second, however, we see the X3D CPUs falling substantially behind both AMD non-3D and Intel CPUs. So the large L3 is still clearly impacting the test results, just in a negative way for the Ryzen CPU. This is a surprising result for sure, even on the Intel side, as 270K Plus easily beats 285K.

Finally we have throughput in requests per second and TTFT in seconds (hence for TTFT, the shorter bar implies better performance). The octa-core X3D parts again do really well here. It seems the single CCD design is definitely at a major advantage and the dual CCD 9950X3D is likely having cross-CCD latency issues since only one of its CCDs has 3D cache.

The test also suggests that the new 9950X3D2 will do much better than the 9950X3D, as both CCDs on the former pack 3D V-cache. Intel's upcoming Nova Lake-S dual chiplet parts may show even better performance.

Source and images: GiggleHD