First, SPEC2017 is not a 5 year old benchmark, it's updated regularly
They only up the year version when it's an entire redesign of the whole thing.
SPEC is not purely synthetic either, it uses real applications for benchmarking without the UI/IO. e.g. 526.blender_r really runs blender. 602.gcc_s really runs gcc. 625.x264_s really encodes x264, etc... This isn't your Cinebench or CPU-z benchmark.
SPEC2017 is currently the gold standard for CPU benchmarks in industry, this would be the no-nonsense benchmark. It measures CPU performance (including the memory subsystem) as clearly as it gets, because it doesn't run equivalence on multiple platforms using completely different libraries that are supposed to do the same task, doesn't use off-core accelerators or non-standard instructions, etc... And not all SPEC sub-tasks are memory intensive, btw. Most of SPECfp MT is (biggest exception is blender), but FP calculations do saturate memory bandwidth so this is a real concern for multithreaded tasks, this is partly why HEDT platforms and servers have more memory channels (again this is a test that shows why you need more memory channels as you increase your cores).
SPEC benchmarks are how AWS, Google, Microsoft Azure, etc decide to buy CPUs for their datacentres. How laptop makers decide their chips. This is how every academic computer architecture paper measures their research. And this is how Intel/AMD/ARM/Apple/etc test their CPUs internally (because they know their customers test them this way). If it's good enough for the entire chip industry, it's as no-nonsense as it gets.
SPEC runs a search programme for new workloads that aren't covered by existing benchmarks. It's incredibly difficult to find new ones. You need to make a new benchmark whose result isn't predictable by blending of existing benchmarks. Your typical online CPU reviewer benchmarks don't even come close (you get R2>0.99 for all typical benchmarks across multiple CPUs, including games, cinebench, geekbench, browser, AI, etc). So if you came up with that, they pay you up to $9000:
https://www.spec.org/cpuv8/
There are already a few of these with regards to SPEC2017, other benchmarks cover them such as TPC-C, PerfKitBenchmarker or SpecJBB.
This is why you can look at your workload, split it into different tasks (e.g. compression, decompression, code compile, search algorithms), profile your tasks to see how much time is spent on each part, then cross reference it to specific SPEC sub-scores and it will give you an excellent understanding of how a chip performs doing those tasks.
This is also why Intel themselves, when they release Xeon products, they release SPEC benchmarks:
https://www.intel.co.uk/content/www/uk/en/benchmarks/xeon-scalable-benchmark.html
AMD does the same:
https://www.amd.com/system/files/documents/amd-epyc-7002-scalemp-speccpu-performance-brief.pdf
https://www.amd.com/system/files/2018-03/AMD-SoC-Sets-World-Records-SPEC-CPU-2017.pdf
As no-nonsense as it ever gets.
Once you move on from industry-standard benchmarks to random application benchmarks, tests become more valuable (for purchasing decisions, you see benchmarks for the apps you use) and far less valuable for CPU comparisons, especially as you differ between platforms. These become whole-platform benchmarks, which again are valuable for purchasing decisions, far less for microarchitecture comparisons.
Give me something you can't do with SPEC2017, you and I can each buy an M1 Ultra Mac
). The slides are just beyond embarrassing and misleading. Seems like they're learning a thing or two from Intel's marketing department 
