Server chips vs. personal use chips

dragonfly522

Distinguished
Sep 16, 2010
123
0
18,680
I presume that most people know the answer to this one or I probably would have found it in the search for the answer. I am wondering why people don't buy server chips for personal use when they are faster or have more cores. Are there features in them which make them incompatible for personal use? I see in many benchmarks that the fastest opterons smoke many of the i7 chips and the fastest pc amd "personal use" chips.
 
Solution


Server CPUs are different from client CPUs (desktop and laptop CPUs) in a few different ways.

- Server CPUs almost always support features like error-correcting (ECC) memory that can increase reliability of a machine that is always left running. Some client CPUs support this technology, some do not. Almost all...
They are not specifically made that way.

In fact, server chips and consumer chips are the exact same. But the server chips are binned higher that will last longer. As well they do tend to have some features that are specifically geared towards server based tasks.

They also are on a different socket that supports multi-socket systems, such as dual/quad CPU configurations and as well support more memory in the system. They also are programmed to work best with a server based OS.

AMD used to make their server chips and consumer chips on the same socket and Intel did make some Xeons for LGA 775. But they were the same in terms of performance they were the same as their consumer counterparts.

In most cases though there is no reason to buy a server CPU due to higher costs.
 
Not really. Server chips are exactly like desktop chips, except most of them support extra features like ECC RAM and Multi-processor configurations. On some notes, it can also use higher binned chips.

Plenty of people buy server chips for personal use, the Xeons for examples. The reason Opterons don't see much usage is because it uses its own socket/chipset.

Most sever boards have a different form factor, and has a completely different focus compared to desktop boards; for example you lose the PCIe slots, pretty looks, ample connectivity and OC features for extreme stability.
 
Tell me if I need to do a little bit of research because of the amount of info. needed to answer this, but what are binned chips? I think my understanding, with the little knowledge that I have about the subject, is that mobos which support server chips don't have the support for features that personal use chips have. I would take it a step further and conclude that mobos which support sever chips don't support chips for personal use.
 
Thanks for the insight on this, but I am confused. In laymans terms, how do multi-sockets affect the performance of processors? Tell me if the answer is too long to explain in a single paragraph and I will research this. My other main objective now is to find out what kind of task(s) a server does that is so labor intensive.

Also, I have been looking into getting a new cpu and I find cpubenchmark.net useful. Do you use their site? It looks like a good site to use because from what I have read on their site they use the average or overall score of all of the performance tests. More people these days are using their computers for multi-core and multi-threaded operations, including multi-tasking, so an overall score in a benchmark seems particularly useful. I have found out by reading a couple of the comment sections after the articles, including from Tom's Hardware, that people think that both AMD and Intel have been reaching a standstill. They think that these companies are aren't about to make processors put out much more MHz, if at all, because their engineers haven't figured out a way to improve upon the cpu's architecture. They also think that making more cores as a way to improve overall performance isn't working and they say that this is because software today isn't designed to make optimal use of multiple cores. According to a speaker for AMD, the Bulldozer chips will stop these accusations for future processors. I have read up on how their new chips are going to be better and I am very impressed. I want one, but I think that AMD is going to price these ones a little bit closer to what Intel's price range is for Intel's top of the line processor chips. Not to the degree that Intel prices them, but it will start to compare to the prices of their server chips.
 
Use the Anadtech bench, CPUbenchmarks is just too fuzzy to compare CPUs on a fair basis, as an average isn't really the way to tell performance.

On servers, pricing/performance isnt everything. AMD actually has the value win and performance win in certian catagories due to superior core scaling in certian programs.

For servers, it takes right marketing/partners. Afterall, most servers are bought in bulk from an OEM, and no matter how good your CPUs are, when an OEM is bound in a contract with the other company, or doesn't find it financially viable to switch to another platform, the sales will still be down.
 
Binning is where Intel and AMD test a CPU to reach to top possible clock speed or lowest voltage needed for a certain clock speed.

It also would take into account if cache works or not.

For example, a Core i7 950 is clcoked at 3.06GHz. The Core i7 920 is at 2.66GHz. But they are the same CPU.
 
They also think that making more cores as a way to improve overall performance isn't working and they say that this is because software today isn't designed to make optimal use of multiple cores.
That is not entirely true. There are very specific mission critical (think massive online banking software)/scientific research software that are designed to scale very well with the hardware.

As far as the CPU speeds goes, yes, that matters in single threaded performance. Say you have a very long program that is serial in nature. If you have 100 cores that run at 500Mhz vs a Dual core CPU that runs at 6Ghz, the dual core CPU will have the advantage (this is assuming there are no performance differences per clock). Also realize, that a Core i7 at the same clock speed can do more instructions than a P4 at the same clock speed.

There is also a good chance that Intel could offer an Extreme Edition Sandy Bridge CPU clocked at above 4Ghz, assuming the 4.9Ghz OC on SB on air are true.
 


Server CPUs are different from client CPUs (desktop and laptop CPUs) in a few different ways.

- Server CPUs almost always support features like error-correcting (ECC) memory that can increase reliability of a machine that is always left running. Some client CPUs support this technology, some do not. Almost all AMD CPUs made in the past half-dozen years or so support ECC memory, Intel FSB-equipped CPUs could support it if the chipset did, and no LGA1156/1366 Core i-series CPUs have ECC support.

- Server CPUs can sometimes support registered memory, which allows for more than two memory modules per channel and more than two ranks of memory ICs (more than 8 or 9 ICs per side) per DIMM.

- Server CPUs can sometimes support multi-socket operation so you can have more CPU cores on your motherboard. No client CPUs since the Pentium III support being run in multiple-socket mode, even if the CPUs will fit in the motherboard.

- Server CPUs sometimes have more cores than client CPUs. AMD's Opteron 6100s have 8 and 12 cores while their client CPUs top out at 6 cores; Intel's server CPUs go up to 8 cores while their client CPUs also top out at 8 cores.

- Server CPUs sometimes have different sockets than client CPUs, but not always. Right now AMD uses completely different sockets for servers (C32 and G34) than they do for client (AM3, S1) and Intel shares LGA1156 and LGA1366 with both client and server. They still do have LGA1567 solely for the largest Xeon MP setups though.



Multiple sockets allow for multiple processors. Adding a second CPU adds another set of CPU cores and another set of memory slots, so you get more cores, more memory bandwidth, and more memory capacity from adding a second CPU. That second CPU doesn't help any unless you need more memory, more memory bandwidth, or more CPU cores to do your work. Many client applications such as games do not gain much from going from 3-core to 4-core and 6-core CPUs, and the 2-memory-channel LGA1156 Core i7s perform similarly to their 3-memory-channel LGA1366 Core i7 counterparts. As such, there's typically not a lot of benefit in using multiple processors for client programs. Video encoding is the big exception as a good encoder will use as many cores as you can throw at it.

Servers run things like virtual machines and Web serving programs, which are highly multithreaded and scale well on servers with lots of CPU cores.

Also, I have been looking into getting a new cpu and I find cpubenchmark.net useful. Do you use their site? It looks like a good site to use because from what I have read on their site they use the average or overall score of all of the performance tests.

It's PassMark's CPU benchmark website. I take it with a BIG grain of salt as four of the eight tests the benchmark runs are completely synthetic benchmarks with no real-world-application applicability (integer math, prime number test, FP math, SSE/3DNow! test). Case in point- the Pentium 4s were champs of a lot of synthetic benchmarks yet lost to just about everything else out there on real programs. I'd also wonder at the exact code they use, their compiler and their compile-time optimizations. The benchmark seems to well with both clock speed and cores, but it seems to absolutely kneecap AMD's CPUs in comparison to Intel's. Lo and behold, in their "About Us" page they have an "Intel Software Partner" logo there. Maybe it would be okay to get a VERY rough idea of how two different CPUs from the same maker stack up against each other, but I'd certainly NOT use it to compare AMD vs. Intel CPUs.

More people these days are using their computers for multi-core and multi-threaded operations, including multi-tasking, so an overall score in a benchmark seems particularly useful.

The best benchmarks are the ones that reflect what YOU do with the computer and use real life applications. If you do a lot of gaming, look for tests that run real games in their benchmarks rather than some synthetic "gaming" score from a benchmark like 3DMark. It is awfully tempting to want to see one number that says "this CPU is faster than that one," but that number really doesn't exist any more. It used to when everybody copied Intel's designs and different makers' CPUs were very similar except for clock speed and price, but now they are very different and much more difficult to compare.

Also realize that a lot of "performance" is responsiveness and not speed. A computer with plenty of RAM and fast disk drives will feel a lot snappier than one with a stonking huge CPU but a mediocre amount of RAM and a slower disk drive. Also never overlook the effect of your software on how snappy your computer feels. A well-tuned OS free of cruft feels fast, while the same OS with a bunch of crap running in the background feels dog-slow.

I have found out by reading a couple of the comment sections after the articles, including from Tom's Hardware, that people think that both AMD and Intel have been reaching a standstill. They think that these companies are aren't about to make processors put out much more MHz, if at all, because their engineers haven't figured out a way to improve upon the cpu's architecture. They also think that making more cores as a way to improve overall performance isn't working and they say that this is because software today isn't designed to make optimal use of multiple cores. According to a speaker for AMD, the Bulldozer chips will stop these accusations for future processors. I have read up on how their new chips are going to be better and I am very impressed. I want one, but I think that AMD is going to price these ones a little bit closer to what Intel's price range is for Intel's top of the line processor chips. Not to the degree that Intel prices them, but it will start to compare to the prices of their server chips.

CPU engineers reached a point where CPU clock speed could not increase very quickly due to thermal issues in the early 2000s. Thermal output increases at a 1:1 ratio with increasing clock speed and with the square of the voltage increase. Running at a higher clock speed requires more voltage, so you can see how a small bump in clock speed can result in a big bump in heat output. That's fine in the Pentium III days when chips threw off 25 watts of heat, but not so much when you had P4s throwing off 100 watts already. Thus to increase performance, they wanted to find a way to make the CPU do more with every clock cycle, since they couldn't just keep getting more clock cycles per second like they used to. That's why we have things like multiple-core CPUs, it's all a way to have a CPU do more in a given period of time.

The thing that has really been lagging is programmers. Writing good multithreaded code is difficult, but it can yield huge performance gains in most programs if done well. Many programmers whine about CPU clock speeds remaining more or less constant since it means more work for them, rather than writing simpler single-threaded code and relying on the CPU makers to keep jacking up clock speed to make their programs run faster. The programmers really need to wake up and realize the days of more GHz is over unless there is some paradigm shift in CPU manufacture that solves the thermal problem. That paradigm shift will most likely be a big change in what CPUs are made of, such as going from silicon to diamond or using carbon nanotubes or some other not-here-yet technology.
 
Solution
^ Excellent reply MU. Well done. +100.

Intel shares LGA1156 and LGA1366 with both client and server. They still do have LGA1567 solely for the largest Xeon MP setups though.
Ok, that's just stupid. I thought all Xeons were LGA1366, but apparently not. WTH? Why would Intel have 2 different sockets for servers too. It makes sense to have a different socket for MP set ups, but really?

Btw, what's the difference between LGA1366 vs LGA1156 Xeons? Support for 2P,etc?
 


LGA1156 Xeons (model number 34xx)
- 2 channels of unbuffered ECC/non-ECC DDR3-1333
- Single-socket operation only
- Onboard PCIe controller
- Communicates with PCH chipset over DMI
- Quad-core, 8-thread 45 nm units with 8 MB L3

LGA1366 Xeons
- 3 channels of unbuffered ECC/non-ECC DDR3-1066 or 1333 (35xx or 36xx series)
- 3 channels of registered ECC, unbuffered ECC or unbuffered non-ECC at DDR3-800 (E/L550x), DDR3-1066 (E/L55xx/E/L56xx), DDR3-1333 (X/W55xx, X56xx)
- Single-socket only for the 35xx and 36xx, two-socket operation enabled for 55xx and 56xx
- Has 45 and 32 nm CPUs with two, four, and six cores. Some have HT, some do not. L3 caches vary from 4-12 MB.
- QPI link to the chipset, chipset controls PCIe.

That's the technical differences I can think of. In reality, it really seems like the LGA1156 Xeons are just like the LGA1156 i7s- basically redundant because of nearly-identical LGA1366 units.
 
^I would gander that the 55xx and 56xx LGA1366 Xeons are 32nm and have at least dual QPI links which would allow for multiple socket setups.

In fact if I remember correctly, all the server based CPUs from Intel above LGA1366 have 2x QPI or more (4x QPI for Bekton).
 


The 35xx and 55xx Xeons are all 45 nm, the 36xx and 56xx units are all 32 nm. All of them have two QPI links, but Intel shuts off the second QPI link in the 35xx and 36xx chips to disable 2-socket operation because they want the customer to buy the considerably more expensive 55xx and 56xx units instead.

In fact if I remember correctly, all the server based CPUs from Intel above LGA1366 have 2x QPI or more (4x QPI for Bekton).

They all have four links. The Xeon 6500s are the same silicon as the 7500s, but Intel shuts off the third and fourth QPI links in the 6500s to limit it to two-socket operation, because they are less expensive than the 7500s. AMD used to do that too with their Opterons as the 2xxx series was simply the 8xxx series with the third HyperTransport link shut off to disable more than two-socket operation. AMD doesn't do that any more as they have switched to completely different sockets and CPU packaging for their 2-processor and 4-processor units. 2-processor is Socket C32 (LGA1207) and has two memory channels and one CPU die per socket, whereas 4-processor is Socket G34 (LGA1944) with four memory channels and two CPU dies per socket.