2

I built a desktop computer probably about 8 years ago when the first series of i7 processors (i7 920) were released. I pulled this out of my dusty cupboard yesterday where it has sat unused for about 4 years.

I am interested in possibly building a new PC soon, and so I have started researching various components.

I noticed that my current system supports both hyper-threading and triple channel memory. Most of the current intel CPU's seem to fall into one of 2 categories.

  • They are either about the same price as I bought my i7 920 for (I think about 250 GDP) and have dual channel memory and often do not have hyper-threading.

  • Or they are significantly more expensive and have both hyper-threading and quad channel memory.

I wonder why it is that CPU's are being manufactured without hyper-threading considering that technology has been around for quite some time. (My understanding is this is an intel only patented technology, however.) But this is not my main question.

I want to know how do current intel CPU models perform compared to my i7 920.

My expectation is that even though I have triple channel memory, my memory clock is either 1066 Mhz or 1333 Mhz, and current DDR3 clocks exceed 2 Ghz. DDR4 modules I have seen start at about 2 Ghz. Therefore I expect modern dual channel configurations probably produce about the same memory bandwidth performance if not more.

Current intel CPUs clock about 3.5 Ghz stock, whereas my CPU clocks at 2.66 Ghz stock. Naively higher clock is faster, but does the performance vary with and without hyper-threading?

My question comes down to should I invest money upgrading my system? (Since this is perhaps subjective, what I really want to know is measurements of the performance difference.) And most importantly, are there any benchmark tests which compare older i7's to current i7's?

Subsequently Added Info

I will be using my system for data processing. Typically this involves a small number of CPU operations on large (~ 6 terrabytes) quantities of data. As a consequence, performance is usually limited by memory bandwidth and disk bandwidth.

Disk bandwidth is typically a problem which you can either deal with in an effective way by changing how your code works (eg; parallelize it so less time is spend loading data from disk), or you can't deal with at all.

Memory bandwidth is subject to a large number of things. In my experience, 3 channel memory will ensure exactly 3 times more bandwidth than an identical system in single channel mode. I would assume intel have reverted to average consumer CPU's being dual channel because most users, even gamers, probably do not run up against memory bottlenecks in the everyday use of their computers. Having "more megahertz" in your RAM is often only useful if your CPU/motherboard combination can run at those higher speeds. For example, the motherboard I am using with the i7 920 has memory speeds of 1066 / 1333 Mhz, and up to 1866 Mhz as "overclock". But I could never get anywhere near that. 1333 was sometimes a challenge, and the memory I was using was rated 1600 Mhz...

Appendix

I found some info here which compares intel i7 920 to 6700K and 4790K.

http://cpuboss.com/cpus/Intel-Core-i7-920-vs-Intel-Core-i7-6700K

But the info on this website makes no sense. They conclude that they recommend the 6700K even though it performs worse in single process and mutli-process tests. I personally do not think this is a reliable source.

3

Are there any benchmark tests which compare older i7's to current i7's?

  1. Cinebench (benchmarks) (6700k performs 72% better)
  2. 3DMark (benchmarks) (comparison1) (6700k performs 33% better)
  3. Passmark (comparison2) (6700k performs 122% better)
  4. Mixed Benchmarks by AnandTech (comparison) (6700k performs 85% better average)

I wonder why it is that CPU's are being manufactured without hyper-threading considering that technology has been around for quite some time.

  1. Marketing
  2. It costs them money to add it
  3. Demand isn't high because ridiculous amounts of threads aren't particularly useful in most situations.

Most of the work I do is memory bandwidth constrained.

Memory bandwidth depends on your motherboard and your CPU.

i7 920

  • RAM Bandwidth - 25.6 GB/s
  • RAM Types - DDR3 800Mhz/1066Mhz
  • RAM Channels - 3
  • RAM Max - 24 GB

i7 6700k

  • RAM Bandwidth - 34.1 GB/s
  • RAM Types - DDR3L 1333Mz/1600Mhz & DDR4 1866Mhz/2133Mhz3
  • RAM Channels - 24
  • RAM Max 64 GB

Footnotes

  1. You are mostly looking at Physics score
  2. Not necessarily reliable
  3. Does not support your current RAM
  4. This is a downgrade

By the way, CPUBoss is about the worst site for comparisons there is; same goes for GPUBoss. They simply give biased opinions on processors, i.e. any new Intel processors are considered superior to old ones, and almost every Intel processor is onsidered superior to AMD.

| improve this answer | |
  • Does the "RAM bandwidth" account for the number of channels; ie; in single channel mode the i7 6700k RAM bandwidth will be about 15 GB/s instead of about 30 GB/s – user3728501 Jul 6 '16 at 20:53
  • @user3728501 RAM bandwidth is the maximum the processor can handle, I'm assuming without regards to channels. Quoting Intel's website, "Max Memory bandwidth is the maximum rate at which data can be read from or stored into a semiconductor memory by the processor (in GB/s)" – Rubydesic Jul 6 '16 at 20:59
0

Passmark is a better benchmark than that cpuboss, which basically just scrapes data from other sources.

As you can see in that lineup, paying close attention to the Single Thread rating, even AMD CPUs are substantially more powerful than the i7-920 today, and the i7-6700K absolutely annihilates both AMD's common offerings and the older i7-920.

Now, to address some of your other concerns, which may end up being much more important to you than just comparing the absolute performance numbers:

Hyperthreading is not always useful. What it does is allow a single CPU core to process two program "threads" simultaneously at a small hit to its overall performance. In situations where your workload can take advantage of parallel processing, adding threads can nearly double your performance. But in situations where your workload doesn't make use of more threads than cores on your CPU, hyperthreading will offer you NO speed advantages, and may in some edge cases result in LOWER performance. This is why Hyperthreading has seen limited adoption - it is no substitute for more cores and even then, many applications are not/cannot be parallelized.

Your workload leans heavily on RAM and disk performance - your CPU will probably not be the "bottleneck" in your system (the part everything has to wait on in order for work to get done). My guess, given that you have triple channel RAM, is that your bottleneck will be with storage - and this kind of puts you in a good position, because the upgrades I'm about to go over would be things you'd end up needing whether you upgraded your current system or moved to a whole new architecture.

In modern computers there are a wide variety of methods for speeding up disk read and write. I will go over the two options which make sense for you, using a desktop PC, assuming all data is local to that PC (not stored remotely on a network):

  1. Tiered storage - you can use a combination of RAMDisks, SSDs, and HDDs to cache your storage, enabling chunks of your total dataset to be processed very quickly. This is a relatively simple solution to set up, and is limited only by the filesize you need to be working on at any given time, because RAM and SSDs are expensive storage solutions compared to HDDs and it can be prohibitively expensive to build a cache capable of storing terabytes of data all at once - it would be better to have, say, a 1gb RAMDisk cache paired with a 32gb SSD cache, managed by some cheap software like PrimoCache, moving files on and off your HDD array as needed.

  2. RAID - you can really speed up your HDD array by putting them behind some good, quality hardware RAID. Dropping an old SAS to SATA II MegaRAID controller into your PCI-E 16x slot could provide you with a way of creating a RAID0 or (ideally) RAID10 array that moves data with a small built-in RAMcache for dedicated writeback performance storage behavior. These older RAID controller cards can often be had on ebay for under $100, which can make it a more cost-effective option than building a working tiered storage system. It is harder to set up, though, requiring you to have a good understanding of how RAID levels work and how to set up your system to play nicely with the RAID array you end up creating.

Either of these options can be deployed into your old PC or a new replacement, but you'll probably want one or the other no matter which way you go.

| improve this answer | |
  • I have to politely disagree with the usage of Passmark. Although it is much, much better than CPUBoss which is a biased piece of crap, not even a benchmark, Passmark isn't very accurate. I'd recommend CineBench or 3DMark/PCMark – Rubydesic Jul 6 '16 at 20:18
  • @RubyJunk It is certainly difficult to know what a score of 10000 passmarks compared to 5000 passmarks really translates to in terms of real-life performance. This could be a very memory intensive test, or alternatively it could be very cache/register level intensive, and involve no memory read/write operations at all. Finally, although RAID is a good idea it's not what is limiting performance here. Most of the work I do is memory bandwidth constrained. – user3728501 Jul 6 '16 at 20:26
  • With all due respect, can you give us more details about 1. the exact nature of your work - specific program/platform names help us a lot sometimes. 2. how you are determining that memory bandwidth is the limiting factor? The passmark test was just the first thing I could think up which would give you a general idea about CPU performance (which is all you're ever going to get without specific program and workload information). It is accurate within itself, as in it is consistent. Passmark is also well-documented (as are Cinebench and 3DMark/PCMark, I'm sure), so you can read up on it. – Adam Wykes Jul 6 '16 at 21:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.