That's pretty much all there is to compare against Intel since they're just not producing proper new server CPUs. They'll be releasing something now, aren't they? And that will be competing with Milan-X and whatever is based off Zen4 Zen 3 now. I can't remember all the names TBH, so sorry for not being super specific.
Well, I am going to start my answer with your own words…
the thread is not just about AL and ML, but the process node.
What I explained in my previous post is that a 64-core running at 2GHz will always win in efficiency against a 40-core running at 3.2GHz even if both cpus had the same architecture, the same overall performance and were manufactured on the same process node. So EPYC possibly winning on some efficiency metrics against Xeon (a Xeon on a two generations old microarchitecture mind you) is not a feat of the process node. It is a feat of the design choice of going with more cores and running them at lower clock speeds.
And no you are not comparing their best. By comparing Icelake to Milan we are not comparing the best Intel and AMD managed to extract from the latest nodes they have had access to. Although Zen 3 derivatives are really the best AMD managed to extract from TSMC’s 7nm, Icelake is certainly NOT the best that Intel managed to do on their 10nm node. For starters since the 10nm used for Icelake, Intel has had two new iterations of that manufacturing node. The first was the 10nm superfin that was used with Tigerlake and the second is the 10nm enhanced superfin (later renamed to Intel 7) that is currently used with Alderlake. Both of these improvements increased performance per watt compared to the 10nm used for Icelake in a way equivalent to moving to a new full node. Furthermore, and equally important is that since Icelake we had two new microarchitectures (Willow Cove and Golden Cove). Besides the increase in IPC a microarchitecture can improve efficiency as well. AMD managed just that with Zen 3 against Zen 2. Remember both Zen 2 and Zen 3 were made on TSMC’s 7nm yet Zen 3 is more efficient due to the new architecture. What is more is that Intel has yet another architecture on their Intel 7 node that is expected to have even better efficiency. It is foolish really to think that anyone can optimise and tune a given node better than vertically integrated Intel especially given the impressive recent record of Intel at optimising and extracting the most out of their nodes (both in 14nm and now at 10nm).
I already provided evidence that AMD is ahead via AnandTech's server tests. Do you need me to Google more for you? I could also get more Laptop data so I can double down on the point, but nah.
No you did not provide such evidence. Just because you linked an article from a reputable website doesn’t make the content of what you linked relevant to the topic in discussion.
The matter of the fact is: Intel even using a more efficient design: monolithic and bigLITTLE is barely on par with AMD's chiplet approach which is inherently less efficient on the desktop and server front.
You are misguided here on multiple fronts.
1. The hybrid design really only comes to play in overall power consumption comparisons across e.g. a daily workload including heavy, mixed and light workloads as well as idling. And in such comparisons Intel is doing wonders. Comparisons when pummelling the cpu with a single heavy workload such as Blender or Prime 95 are irrelevant.
2. Intel’s main reason for using a hybrid design is not really efficiency (although this is one of the reasons it is not the
main one, not in desktops at least). Instead the main reason is area efficiency for multithreaded (MT) performance. You see the only reason to have more than 8 cores is scalable MT performance and it turns out that a cluster of 4 Ecores (which occupies the same area as 1Pcore) offers twice the MT performance compared to 1Pcore for the same power envelope. In other words, with 8P+8E cores Intel is achieving the same MT performance as a 12P+0E core cpu but only using the same die area as a 10P+0Ecore cpu.
3. Although in workloads like Cinebench and Blender the efficiency cores add enough performance for Intel to match or beat the performance of AMD’s 12 and 16 cores without requiring much extra die space or needing much extra power, it is still important to note that Alderlake cpus only have 8 big/performance cores and most performance comes from these 8 big cores. So we really again have the situation of comparing a cpu with fewer cores but higher per core performance (due to running at higher clockspeed and having higher IPC that comes from more per-core computational resources) going up against twice the cores at lower per core performance (due to lower clockspeed and lower IPC). The former is never going to win in power consumption metrics.
4. Power consumption alone is not a proper metric for efficiency. Energy is what is. If a cpu pulls twice the power but completes the job in half the time versus another cpu, then both cpus have used the exact same energy and have the same energy efficiency (and the former also saves you half the time). Power consumption is only relevant in cooling.
Facebook
Twitter
Instagram