So let's go through your opinions line-by-line then;
So looking at CPUs with good single core performance. One example
would be AMD EPYC 73F3 which is (according to passmark) on par with
Ryzen 5950X, but 10 times more expensive and it's rated over x2 TDP.
This is a terrible comparison - try putting two CPUs in a 5950X-based 'server', or more than 128GB of memory, or having more than 64MB of L3 cache, or more than 16+4 PCIe lanes.
It's not clear if you know what a server is I'm afraid, not in a production/professional environment anyway. Airliners, cruise-ships, oil-tankers, buses, trains etc. are designed to be reliable, resilient and deal with a wide variety of usage requirements over multiple years at a predictable cost - jet-fighters, speedboats, your car etc. are faster yes but designed for single, focussed use cases where reliability and costs are less of an issue.
In the intel realm there's really nothing on par when you look at
single core performance as most of the offerings are marginally faster
from Xeon E5-2660 v3 which is 8 years old. The upside for intel is
that their server CPUs are not NUMA, so memory handling is better, and
support seems more polished so not even sure how these synthetic
benchmarks relate to real life situations where memory needs to be
moved between cores in non-NUMA aware applications (is there really
anything NUMA-aware out there besides oracle server?)
Firstly nobody in a server environment cares about single-core performance, maybe a tiny handful do but >99% of people do not. It appears that you think that clock-speed is the only measure of single-core performance anyway, forgetting the impact of memory/QPI/UPI/IPC performance increases - not only would something as low-end as a 4210T (10c/20t, 2.3-base/3.4-turbo) absolute stomp a 2660v3 into the ground (and they list at only 555 USD by the way) but there are SKUs like the 8732C (28c/56t, 3.2-base/3.5-turbo) that would make the older chip seem silly - plus all their 40c/80t SKUs too.
And of course Intel does NUMA too, you probably mean on-socket-NUMA, which is a different thing, and the 92xx Xeons even did that - look at the 9282 (56c/112t) for instance, which has the same memory concerns as Zen CPUs.
Your assertion that "Oracle Server" is a rare case of being NUMA aware is wrong and pointless anyway as what matters is that the base OS or Hypervisor is NUMA aware, and anything even vaguely recent has had this for years. Any modern Linux/Windows will happily keep processes and their memory 'near' to each other within a NUMA domain unless very highly contended indeed, the same has been true for ESXi/KVM/Xen for even longer, most server applications just don't need to consider NUMA at all as it's all taken care of for them.
Xeon workstation CPUs seems to be on par with AMD's customer grade
CPUs, but still not as powerfull as intel CPUs available for ordinary
customers. The only upside for xeon workstation is that they support
ECC, which seems to be the only reason why this space still exists as
you could get much better customer CPU probably for half the price,
but there's no ECC option and RAM is limited.
Newer Threadripper and Ryzen CPUs and chipsets support ECC, so your point here is moot.
Also i noticed there are some issues with EPYC but Ryzen is working
perfectly fine. So that may be a loaded question but what are people
using nowdays as maybe im missing something but getting a server which
is on par with some super-cheap ryzen setup in terms of power usage,
reliability and performance seems impossible if you don't want to
invest some enormous amount of money and the best you can do is to end
up with single core performance of consumer grade 2YO Ryzen anyway.
Which problems, can you be specific?
Again this your lack of production experience. This site is very specifically for professional sysadmins/system-designers, we make that very clear when you join. And our number one priority when it comes to servers is to maintain the data we have and to maintain service for the dozens/hundreds/thousands of applications and users served by our infrastructure. Yes we care about power-usage, heat-management and overall server performance but these are distant secondary concerns to reliability, resilience, monitoring-capability, pre-failure-warning and capacity in terms of cores/threads/memory/PCIe-lanes - and frankly anything BUT production-level CPUs fail on multiple, if not all, of these criteria.
If you get bored google 'RRAS', see if that helps you understand.
Has innovation in server space stalled?
No not at all, but it's inherently never going to be at the same bleeding edge as consumer parts, simply because we need reliability - why would we risk running a server with a CPU with a few slightly faster cores but lose all of those features we need that I list above.
My analogy above is key - the vast amount of people fly with others on airliners as it's cheaper and more reliable than flying everywhere in a fighter-jet, the same for shipping containers - you could put one on a faster speedboat but the numbers don't add up, the same is true for lots of other ways of getting things done - handle more load, more reliably and more cheaply with larger well-engineered solutions rather than unique custom ones.
Metaphorically you've walked into the pilot's lounge at an airport and ripped into all the Boeings and Airbuses because they can't do a barrel-roll as easily as a Cessna. We're not idiots, just about everyone who comes here regularly has a decade or more (32 years in my case) of successfully doing this job on very large infrastructures, we're professionals who know how to research our work and learn from others in the same field (literally the point of this site).
Do you honestly think this post of yours is groundbreaking, genius-level work that hundreds of thousands of people, including every server and CPU manufacturer, has overlooked - or might it be that maybe you just need to learn more?