AMD has made a tremendous leap compared to the previous generation, impacting various areas. Instead of focusing on benchmarks, interested individuals can look online (e.g., Phoronix) to witness the impressive performance of these processors.
The main reason I decided to upgrade from the 2990WX to the 3970X (in my opinion, the most significant upgrade) is the elimination of the peculiar NUMA architecture present in its predecessor (4 NUMA nodes / 2 CCX per node / 4 cores per CCX - 2 nodes with dual-channel memory each - 2 nodes with no memory channel at all). This architecture resulted in inconsistent performance in various workloads (not CPU-bound) due to memory latency. Many times, I even observed the 1700 outperforming the 2990WX. A clear example in my case was virtualization. While nested virtualization ran smoothly on the 1700, the same VMs on the 2990WX experienced significant steal time (vCPU waiting for the regular CPU to serve other vCPUs). Even after hours of optimization at the operating system, VM, and BIOS levels (e.g., memory interleaving, memory affinity, numactl, CPU pinning, etc.), I couldn't achieve the desired results. Ultimately, the only way to improve performance was to disable SMT, which was disappointing. This issue isn't limited to virtualization but affects all multicore workloads sensitive to memory latency.
To summarize, the 2990WX is a powerful processor that truly demonstrates its capabilities only in CPU-bound workloads like content creation and video editing. Those running such applications and considering an upgrade to the 3rd Gen Threadripper won't theoretically see more than a 10-15% performance increase (which matches the IPC gain between the 2nd and 3rd generations).
However, with the new Threadripper generation, all these peculiarities are a thing of the past. Besides many new features that improve performance in both single and multicore scenarios, all cores appear as a unified group (UMA) to the operating system, with equal distances between the cores and memory controllers. This architecture ensures consistent performance across all cores for every type of workload, out of the box, without additional configuration. In simple terms, it delivers true power with 32 cores/64 threads without any fine print.
In my case, with the 3970X, I no longer experience any steal time in virtualized workloads, and I've noticed significant improvements in VM deployments, kernel compilation, Spark jobs, and Gromacs, reaching up to 70% faster performance!
Regarding thermal/energy efficiency, AMD works wonders. No, the 280W TDP (notably in all-core boost) is not high; it has the highest performance-to-watt, watt/core, and watt/thread ratios in the market (see AnandTech, Phoronix). Considering it has 32 cores, almost twice as many as the 3950X, with a slightly higher base clock of 200 MHz and an additional 40 PCIe 4.0 lanes, the TDP appears modest. As for temperatures, I observe 35°C idle, 40°C average, and a maximum of 69°C under stress testing with a 560mm cooler, NFA14 3000 fans, and an EK-Velocity water block. It's important to note that these temperatures are accurate and there's no offset as with older Threadrippers (27°C). The mentioned numbers are based on stock settings alone. Anyone attempting overclocking by increasing power will realize that the performance gain is disproportionate to the power consumption and the temperatures that skyrocket. Additionally, AMD does not cover Threadripper overclocking under warranty.
AMD currently claims to have the most