"It feels strange, but in this generation, I am more excited to use the Ryzen AI 300 series than the Ryzen 9000 series based on the demos I have seen."
I do not agree with this.
While it is true that the Ryzen AI 300 brings significant improvements, especially in the GPU for everybody and in the NPU for those interested in machine learning inference, the Ryzen 9000 series brings the greatest advance in CPU throughput since 2019, when the Ryzen 9 3950X Zen 2 CPU has been introduced.
The Ryzen AI 300 series has the same vector throughput as Zen 3 and Zen 4, only in the desktop and server Zen 5 the vector throughput is doubled.
The evolution of the floating-point throughput in desktop CPUs (i.e. non-server non-workstation CPUs, which are much cheaper per core) has been the following. After a time of rapid throughput increase in the sequence Core 2 => Nehalem => Sandy Bridge => Haswell, Intel had remained without competition and the desktop CPUs have stagnated at 32 FP64 FMA units per socket.
When Zen has begun to apply pressure on Intel, they have increased the number of FP64 FMA units to 48 in Coffee Lake in 2017 and to 64 in Coffee Lake Refresh in 2018 (Zen 1 and 1+ had lower floating-point throughput per core and per socket).
One year later, in 2019, AMD launched Zen 2 with up to 128 FP64 FMA units per socket and Intel has not been able to match AMD since then.
Now, in 2024, Ryzen 9 9950X brings another doubling of the floating-point throughput, with 256 FP64 FMA units per socket.
Despite the hype for machine learning, there still are much more important engineering applications that depend on traditional floating-point computations. Even if the most efficient way for those would be to use FP64 GPUs, the prices for those have grown enormously and they are no longer acceptable for any small business or for any individual.
I still have some old AMD GPUs from almost a decade ago, which were cheaper and faster at FP64 computations than any later CPUs or GPUs.
Until this year there was nothing that could replace them. Now, Ryzen 9 9950X has about the same speed as the old GPUs (close to two FP64 Tflop/s), but it has a lower price and a lower power consumption and it is also much easier to write programs for it, so it is a good upgrade.
No other current CPU or GPU will have a similar FP64 performance per dollar with Ryzen 9 9950X, at least not at reasonable total price. While after its launch AMD Instinct MI300X was offered at $16500, which means a much lower performance per dollar than 9950X, I have seen recently an offer from Dell at a reduced price of $9000. At that price, the performance per dollar would be similar to 9950X. Nevertheless, even if that offer had been valid for those who do not use Dell servers, $9000 is too much for any small business or individual. Such a price is acceptable only for an organization that can ensure that such a GPU will be kept busy almost 24/7 with doing work that would support some profitable activity, allowing to recover the paid price.
So for most people interested in FP64 computations, from now on 9950X will be the only good choice and it will bring a great jump in performance over the computers of similar cost or a much lower cost in comparison with the computers of similar performance, like 3950X did 5 years ago.
In comparison with 9950X with 256 FP64 FMA units per socket, a Ryzen AI 9 HX 370 has only 96 FP64 FMA units per socket, the same with a Ryzen 9 5900X. It is a 50% increase vs. Phoenix/Hawk Point (due to 12 cores vs. 8), but it is not a doubling like in the desktop Zen 5.
