CPU Threadripper 2- Where's the bottlenecking?

Gautam

Cable-Tie Ninja
Sep 5, 2016
148
123
I'm aware of the 5 GHz stunt with the liquid chiller. But Intel has had 28 core parts for servers and datacenters not needing such cooling, though still very expensive and power-hungry.

There's going to be a refresh of the i9 line later this year, which will likely have some sort of answer to AMD, whether on 2011 or 3647 or both, and given how much better the mesh architecture scales versus AMD's, they will probably win out. It's as you said, Intel can counter when they want to, but they don't need to resort to two dies.

AMD will always win on price per core, because their architecture is inherently cheaper, but they're not quite as much of a slam dunk in price per performance as it may seem.
 

lhl

SFF Lingo Aficionado
Nov 16, 2015
121
143
I think the big thing we're seeing is how terrible Windows' scheduler does w/ NUMA. TR2 presents as 4 NUMA nodes and Linux does what it should (working NUMA is actually great b/c you can have independent processes w/ less line contention/resource sharing). I think AMD does a little bit of marketing where they encourage "Enthusiasts & Gamers" to go w/ the 2950X vs the 2990X, but neither Intel or AMD really outright say "if you have to ask, you probably don't need HEDT" since the margins are too good. I think the big thing holding back AMD at the high end atm is AVX-512, but it's interesting to see that AMD is not just doing well on perf/$, but also perf/watt. I'm very interested to see how this plays out with Zen2 on 7nm.

On the server side, Intel can still compete, but they're so hyper-segmented - I think they'd rather give up some market share than lose out on their margins (plus they have so many dies to bin so they sort of need the SKUs), but at least w/ AMD being competitive they have to work a bit harder. Competition is a win-win for customers.
 

Arie

Trash Compacter
Jul 4, 2018
37
70
Shared resources for the CPU cores is always their downfall, look at the FX series from AMD. They physically had 4/6/8 cores, but since each set of 2 cores had shared cache in modules, it acted more like 2/3/4 cores respectively with SMT.

This time though, it's a deliberate nerfing of the 2990WX as to not compete with Epyc, not a fundamental limitation of the architecture. Add to that Windows 10s poor handling of the WX architecture and you get the poor results. I expect there will be a Windows 10 update to fix that and bring performance to the same levels as seen in Linux.
 
  • Like
Reactions: Gautam