-1

Why does the clock speed of CPUs decrease the more sockets you add to the motherboard? For example, for a single-socket desktop, I can buy a Core i7-7700K that runs at 4.2 GHz before turbo boost. But if I buy an HPE DL360 gen. 9 server with two sockets, the fastest quad-core CPU I can get with it runs at 3.5 GHz. And if I get a quad-socket DL580, the quad-core CPUs run at 3.2 GHz. Likewise, the new 8-core EPYC CPU from AMD runs 2.1 GHz, whereas the desktop Ryzen eight-core CPU runs at 3.6 GHz.

user1177071
  • 51
  • 4
  • You are performing a meaningless comparison. Comparing clock speeds across architectures serves no purpose. – David Schwartz Jun 22 '17 at 06:14
  • @David Schwartz: I think you meant to say "microarchitecture." They all use the x86 architecture. I understand they're different. I just want to know why they're clocked differently. – user1177071 Jun 22 '17 at 06:42
  • Also, tests performed by Puget Systems comparing Core i7 and Xeon CPUs in 2015 with the same clock speed and core count showed a performance variance of no more than 2%. So, I don't think it's a meaningless comparison. – user1177071 Jun 22 '17 at 06:58
  • You'll find a strong correlation between land vehicle performance and number of seats because almost no bicycles have more than two seats and almost no cars have fewer than four. But it's still a silly comparison. The lack of variance comes from many factors including the strong correlation between clock speed and fabrication node and between fabrication node and transistor count. – David Schwartz Jun 22 '17 at 07:38
  • So, I have a 14 nm Broadwell E5-2637v4 offered in the DL360 server that runs at 3.5 GHz. In the DL580, I have a 14 nm Broadwell E7-8893v4 that runs at 3.2 GHz. So, the process node is the same (14 nm). The microarchitecture is the same, as well. (Technically they're all using the Haswell microarchitecture, but it's named Broadwell as it's a die shrink of Haswell.) So, your explanation doesn't make any sense. I think you just misunderstood my initial question. I was not asking for a performance comparison between them. – user1177071 Jun 22 '17 at 08:04
  • My explanation makes perfect sense -- those two CPUs are very, very similar in that particular case. And there's *still* no point in comparing their clock speeds since you could just compare their performance. What I'm trying to say is that you're asking the wrong question -- it has nothing to do with CPU clock speeds except coincidentally (in the cases where that correlates with performance, but not in those were it doesn't). – David Schwartz Jun 22 '17 at 08:57
  • It's like asking about cars with larger gas tanks when you mean cars with more horsepower. Yes, there's very frequently a correlation. But in the cases where there aren't (say cars meant for long road trips) picking gas tank size gets it wrong. Everywhere it matters, CPU clock speed gets it wrong. And where it doesn't matter, you could just as well use performance. – David Schwartz Jun 22 '17 at 08:59
  • Just a guess, it might have to do with the TDP of the CPUs and that clock-speed and amount of cores has to be balanced to meet the TDP design. Not sure if that is still the case, but in a former Intel CPU and when using AVX registers, the base clock was downclocked as the AVX registers seemed to consume more power than was possible at that clock speed. But using AVX at a lower clock speed is still faster than using e.g. SSE4 at a higher clock speed. – Thomas Jun 22 '17 at 09:13
  • Downvoting this legitimate question is obscene. – ajeh Jun 22 '17 at 17:28
  • @Thomas: That's a good point. Looking at the specs for the two CPUs, I see the 3.5 GHz E5 has a TDP of 135 W. The 3.2 GHz E7 has a TDP of 140 W. So, the E7 generates more heat at a lower clock. So, perhaps it has some additional circuitry that the E5 does not. – user1177071 Jun 23 '17 at 03:14

2 Answers2

1

Fair question, and a much confused topic for many. It is basically the same hardware so why the giant cost and speed difference....

Servers and Desktops, while very similar, have VERY different warranty's and that causes the discrepancy you are seeing. Servers are more "industrial" designed to run at or near full load for hours on end. So many manufactures de-tune and overbuild to make sure their products will last. (I use alot of consumer stuff for servers work too ;-) )

So more chips/sockets not only more work for the CPUs (the talk to each other, I'll not dive into this) but also more work for the main boards, and more work/load on the power systems. So they slow things down to guarantee reliability (something you need when a business's operations and income is based on some software that needs a database server that doesn't fail, ever).

Its like comparing a 400HP tractor trailer motor to a 400HP sports car motor. The tractor motor can put down full power for much longer periods of time with out issue, and requires less maintenance. The sports car will be lucky to last 100k miles where as the tractor is expected to make it to 400k.

sirmonkey
  • 76
  • 1
  • 1
  • 6
  • Another thing to consider is that in the enterprise environment we usually observe equipment being replaced after 5 year lease expires, while the home office/ enthusiasts may keep their machines around for much longer. There goes "overengineering for reliability" theory. Also, in the datacenters the servers are under watchful eye of support staff, where at home they are often neglected and left running with dried up thermal interface for years. The overengineering for reliability theory just does not cut it. – ajeh Jun 22 '17 at 17:30
  • Most home office/ enthusiasts machines have a warranties a year or shorter. So the OEM only cares to keep the device operational for that long. I work in datacenters (equnix, RW, coresite), many servers aren't as closely watch as you think. Often machines run for extend periods of time with errors on their console because no one walks by to see the screen or logs into the remote management to check. If I don't visit a customer cage for months (no service calls), often I'll find something in error. I've found downed PSU's on core routers that have been out for months. – sirmonkey Jun 22 '17 at 18:33
  • Thanks for the info, I'll remember to avoid equnix, RW and coresite. – ajeh Jun 22 '17 at 20:41
  • I have yet to see any study that shows that servers are more reliable than desktops, with the exception of ones related to specific high-reliability technologies like RAID or ECC (which desktops can also use). And pretty much every place I've seen the claim that servers are more reliable than desktops, they also claim they provide higher performance. I don't think this answer matches with reality. – David Schwartz Jun 22 '17 at 22:29
  • Hi Sirmonkey. Thanks for the reply. So, it sounds like the chips are generating extra heat due to the communication between them. Also, there is a higher load on the power circuits in the server due to the extra processors. So, then it would make sense to underclock them to reduce the heat and also the power draw. – user1177071 Jun 23 '17 at 03:32
  • @David Schwartz. You make a good point. We use HPE servers in our data center, so I can only speak for that brand. Everything is redundant in them (fans, power supplies, network ports, etc.). And the servers are all in clusters, so entire machines can go down without any problems. This leads to the illusion that the servers have more durable parts in them. But the parts break all the time. Most surprising to me has been the number of failed motherboards in the servers. (There's no redundancy for that part!) I haven't seen a CPU go bad in a server or even a desktop, yet, though. – user1177071 Jun 23 '17 at 03:49
0

More cores at a lower clock is usually more power optimal, for reasons better explained by an electrical engineer.

With fewer cores, a higher clock is possible for the same power budget. Which is still in demand. In particular, for single threaded applications, and for per CPU licensing schemes, where more cores is not a benefit.

At the extreme, there are fundamental latency and fabrication process limits that make double digit GHz impractical for these general purpose CPUs.

Don't limit your survey to x86 though. A 12 core POWER8 processor can sustain in excess of 4 GHz, and 4 socket nodes can be had. It has a rather hefty power draw.

John Mahowald
  • 32,050
  • 2
  • 19
  • 34