Do Dual CPU Sockets Matter in 2018?

Linus Tech Tips ·Linus Tech Tips ·2019-05-06 · 1,284 words · ~6 min read
Floatplane YouTube

Transcript

JSON SRT VTT 128
0:00 Zeon W,
0:05 Zeon scalable.
0:08 You know, they're both called Xeon,
0:12 but these things are really different.
0:16 One of them is basically a Core i9 with
0:20 ECC memory support and the other one is
0:24 a server CPU that I fanled all over
0:29 because I love super high-end expensive
0:32 tech toys. Now, in the past, you needed multiple
0:38 CPUs in one multi-socketed motherboard
0:43 in order to handle intensive multi-threaded workloads. But is that
0:48 still the case today? Do you still need
0:52 two of these given that a single Xeon
0:55 Platinum 8180 is 28 cores and 56 threads
1:02 on a single chip? Well, I don't know
1:06 what is the purpose today of a dual
1:09 socket machine like this one and how much have single high core count CPUs
1:16 eroded the market that they used to enjoy? Let's find out, shall we?
1:31 All right, there's a lot more room down here and we are going to need it for
1:35 this honking, not to mention heavy test bench. On this
1:40 test bench, you will find the ASUS C621E
1:44 Sage. This is a dual socket motherboard
1:48 rocking two LGA 3647 sockets for Intel's
1:54 Xeon scalable lineup of CPUs. And setups
1:59 like this have actually been around as far back as the 486 in 1989
2:06 with the resulting secondhand hardware giving enthusiasts the ability to get
2:11 multiple physical cores in their homes
2:14 over the years with the peak being somewhere in the mid 2000s or so. But
2:21 that was then and this is now. Now you
2:25 can get multiple processing cores in a
2:28 single chip. So
2:32 to see how far things have come, what we're going to do is pit this machine
2:38 against the fastest single CPU that
2:41 we've tested to date. We're going to try to keep the number of variables to a
2:46 minimum in order to gauge the impact that these extra CPU cores will have on
2:51 our setup. Though, it should be noted that there aren't many options when it
2:56 comes to aftermarket LGA3647
2:59 coolers because most of the folks
3:02 selling these kinds of systems would figure out their own solution. So, that
3:07 means that our dual socket workstation will run a little bit toasty, but we
3:11 didn't observe any thermal throttling, so it shouldn't affect our performance.
3:16 Let's start off then with good oldfashioned
3:19 Cinebench. I mean, we've we've seen this run before, but it's always fun to see
3:25 it finish that quickly.
3:31 So, in a surprise to no one, the dual
3:34 socket machine is faster. But
3:38 considering its 56
3:41 processing cores, not all of our workloads scale in the way that we might
3:48 expect. 7zip, for example, shows a
3:51 smaller thanex expected gain over our Core i9 Extreme Edition, and YC Cruncher
3:57 even finds itself losing ground. ASUS
4:00 Realbench demonstrates this, though, with that said, the encoding benchmark
4:05 ees out a lead over our Core i9 7980XE.
4:08 and then Blender. Well, here we actually get a victory for our dual CPU system
4:14 again, showing this platform's potential
4:17 for expanding render farms.
4:21 But what is really going on here? Well,
4:27 something you guys have to realize is that there is more to a dual socket
4:33 configuration than just more cores. Do
4:37 you remember when AMD managed a 3%
4:40 improvement in IPC with second gen Ryzen
4:44 just by improving cache latency? So on
4:47 this motherboard, we've got two separate CPUs with two separate sets of cache and
4:56 memory. See, these six banks go to this
4:59 one and these six banks are wired into this one. And that means a lot of
5:04 latency for compute tasks that require
5:07 the same data sets. This latency is a
5:11 necessary evil in the design of multiprocessor systems because of the
5:15 need for nonuniform memory access or
5:18 NUMA for short that allows these two
5:22 processors to efficiently share resources or as efficiently as they can.
5:27 So the short version of this is that it works by transparently allocating
5:31 devices and memory to each CPU which means they can more easily avoid
5:36 interrupting each other while accessing those resources. This in turn reduces
5:41 the amount of waiting around that they have to do for those resources to become
5:45 available. So that's what we're seeing during our testing like in Yunchruncher
5:51 for example where both CPUs are working
5:54 on the same data but it's not really the
5:57 intended use case for this kind of thing. What if we could use different
6:02 data sets? Then we should be able to find this kind of setups true calling.
6:08 And how better to do that than to effectively turn this system into two
6:14 independent computing machines using virtualization. So let's fire up on
6:20 which uses Red Hat KVM as a hypervisor
6:23 to see what kind of results we get splitting these resources into multiple
6:29 independent machines. Immediately, we
6:33 see worse results from our VMs than our
6:36 original 56 core testing.
6:40 But look closer at how much lower it is.
6:45 It's not a whole lot. In every test,
6:49 it's basically the same story here. And we are still way out ahead of the Core
6:55 i9 Extreme Edition, particularly when it comes to Blender. Now if we consider the
7:01 fact that we are getting simultaneous
7:04 work done that gives us a good look at
7:07 what an optimized workload might look like. I mean or heck uh virtualization
7:13 itself is a legitimate task too. I mean
7:16 this thing could be so many gamers in
7:20 one PC. But I digress. I mean, nobody is going
7:24 to buy something like this for their personal rig anytime soon, given the
7:29 $10,000 per CPU price tag, which puts it
7:35 squarely in the territory of big the
7:39 size of the check doesn't matter business. What they care about is
7:43 density. The more processing power a
7:47 single computer can manage, the more processing power that can be physically
7:52 fit into a building. And this is perhaps most important for data centers and
7:58 render farms in particular. The less those guys have to spend on setting up
8:03 the electricity and cooling management
8:07 for a data center versus the amount of
8:10 performance they can get, the better.
8:14 So, will multiple sockets make a comeback in
8:18 the proumer space? Outside of, you know, oil and gas
8:23 exploration where there are still workloads that can benefit from this
8:28 kind of thing, the chances look pretty slim if you ask me. But I don't
8:35 necessarily think that it's Intel's intent to sell these chips in the
8:39 proumer space. And for that matter, even
8:42 in the enterprise space, I don't think they move a ton of them. For me, I look
8:48 at a product like this as more of like a a future crafting exercise where it is
8:54 available today, but it's more of a representation of what actually might be
8:59 attainable a generation or two from now,
9:02 just like the 22 core processors that we were playing around with a couple of
9:07 years ago. Nowadays, those are much more
9:10 affordable and businesses are using them to power your cloud computing services.
9:17 So, thanks for watching, guys. If this video sucked, you know what to do. But
9:20 if it was awesome, get subscribed, hit that like button, or check out the link
9:24 to where to buy the stuff we featured in the video description. Also, linked down
9:27 there is our merch store, which has cool shirts like this one, and our community
9:31 forum, which you should totally join. Now I'm off to finally put this to use
9:37 for the reason that I obtained it.
9:41 Many video editors, one CPU.