Setting Expectations: A Preview of What's to Come in Mobile

Sitting in the audience at the iPhone 5s launch I remember seeing this graph showing iPhone CPU performance increase since the first iPhone. Apple claimed a 41x increase in CPU performance if you compared the Cyclone cores in its A7 SoC to the ARM11 core in the first iPhone. What’s insane is just how short of a time period that comparison spans: 2007 - 2013.

I ran SunSpider on all of the iPhones in our 5s review to validate Apple’s numbers. I came out with roughly a 100x increase in performance, or something closer to half of that if you could run later versions of iOS (with Safari/js perf improvements). SunSpider is a very CPU and browser bound workload, but even if we turn to something a bit closer to real world usage like Browsermark 2.0 I measured a 5x increase in CPU performance over the past 6 years of iPhones.

I frequently refer to the progress we’ve seen in mobile as being hyper-moore’s-law. Until recently, the gains in mobile hadn’t yet approached a point where they were limited by process technology. Instead it’s variables like cost or time to market that govern how much performance was delivered each year. We’re at the beginnings of all of this changing, and it’ll eventually look a lot like what we’ve had in the desktop and mobile CPU space for years now.

When performance results from the new Mac Pro first hit, there seemed to be disappointment in how small some of the gains were. If you compare it to the progress in CPU performance Apple has demonstrated on the other side of the fence, you’re bound to be underwhelmed.

Having personally reviewed every CPU architecture that has gone into the Mac Pro since its launch, I had a rough idea of what to expect from each generation - so I decided to put it all in a chart.

I went back through all of my Conroe, Penryn, Nehalem, Westmere and Ivy Bridge data, looked at IPC improvement in video encoding/3D rendering workloads and used it to come up with the charts below. I made a table of every CPU offered in the Mac Pro, and scaled expected performance according to max single and multicore turbo.

Let’s first start by looking at what you can expect if you always buy the absolute cheapest Mac Pro. That means starting off with the Xeon 5130, moving to the E5462, then the W3520, W3530, W3565 and ending up with the E5-1620 v2 in today’s Mac Pro. I’ve put all of the choices in the table below:

Mac Pro - Cheapest Configuration Upgrade Path
  CPU Chips Cores per Chip Total Cores / Threads Clock Base/1CT/MaxCT Launch Price
Mid 2006 Xeon 5130 2 2 4 / 4 2.0/2.0/2.0 GHz $2199
Early 2008 Xeon E5462 1 4 4 / 4 2.8/2.8/2.8 GHz $2299
Early 2009 Xeon W3520 1 4 4 / 8 2.66/2.93/2.8 GHz $2499
Mid 2010 Xeon W3530 1 4 4 / 8 2.8/3.06/2.93 GHz $2499
Mid 2012 Xeon W3565 1 4 4 / 8 3.2/3.46/3.33 GHz $2499
Late 2013 Xeon E5-1620 v2 1 4 4 / 8 3.7/3.9/3.7GHz $2999

If you always bought the cheapest Mac Pro CPU offering, this is what your performance curve in both single and multithreaded workloads would look like:

The first thing that stands out is both workloads follow roughly the same curve. The entry-level Mac Pro has always been a quad-core option, so you get no increased MT scaling (if you exclude the initial Nehalem bump from enabling Hyper Threading, which all subsequent Mac Pros have supported).

If you’ve always bought the slowest Mac Pro you’ll end up with a Mac Pro today that’s roughly 2.2x the performance of the very first Mac Pro. It’s a substantial increase in performance, but definitely not the sort of gains we’ve seen in mobile. For anyone who has been following the x86 CPU evolution over the past decade, this shouldn’t come as a surprise. There are huge power tradeoffs associated with aggressively scaling single threaded performance. Instead what you see at the core level is a handful of conservatively selected improvements. Intel requires that any new microarchitectural feature introduced has to increase performance by 2% for every 1% increase in power consumption. The result is the end of unabated increase in single threaded performance. The gains you see in the curve above are more or less as good as they get. I should point out that this obviously ignores the ~10% IPC gains offered by Haswell (since we don’t yet have a Haswell-EP). It’s also worth noting that Intel presently delivers the best single threaded performance in the industry. Compared to AMD alone you’re looking at somewhere around a 40% advantage, and ARM doesn’t yet offer anything that competes at these performance levels. It’s bound to be harder to deliver big gains when you’re at this performance level.

Back to the curve at hand, the increase in performance the 2013 Mac Pro offers is arguably one of the best upgrades over the life of the system - assuming you always opted for the entry level quad-core configuration.

What if you always did the opposite though and picked the highest-end CPU configuration? Same deal as before, I’ve documented the upgrade path in the table below:

Mac Pro - Most Expensive Configuration Upgrade Path
  CPU Chips Cores per Chip Total Cores / Threads Clock Base/1CT/MaxCT Launch Price
Mid 2006 Xeon X5365 2 4 8 / 8 3.0/3.0/3.0 GHz $3999
Early 2008 Xeon X5482 2 4 8 / 8 3.2/3.2/3.2 GHz $4399
Early 2009 Xeon X5570 2 4 8 / 16 2.93/3.33/3.06 GHz $5899
Mid 2010 Xeon X5670 2 6 12 / 24 2.93/3.33/3.06 GHz $6199
Mid 2012 Xeon X5675 2 6 12 / 24 3.06/3.46/3.2 GHz $6199
Late 2013 Xeon E5-2697 v2 1 12 12 / 24 2.7/3.5/3.0 GHz $6999

Now things start to get interesting. For starters, single and multithreaded performance scaling is divergent. The high-end CPU option started as two quad-core CPUs but after three generations moved to a total of twelve cores. What this means is that after the early 2009 model you see a pretty significant increase in multithreaded performance for the fastest Mac Pro configuration. Scaling since then has been comparatively moderate as you’re looking at IPC and frequency improvements mostly with no change in core count.

The single threaded performance improvement, by comparison, is fairly mild. If you bought the most expensive Mac Pro configuration back in 2006 you had a 3GHz part. In the past 7 years peak single core turbo has only improved by 30% to 3.9GHz. Granted there are other efficiency gains that help push the overall improvement north of 50%, but that’s assuming you haven’t purchased anything since 2006. If you bought into the Mac Pro somewhere in the middle and opted for a high-end configuration, you definitely won’t see an earth shattering increase in single threaded CPU performance. Note that we’re only looking at one vector of overall performance here. We aren’t taking into account things like storage and GPU performance improvements (yet).

For the third configuration I wanted to pick something in the middle. The issue is that there is no middle config for entirety of the Mac Pro’s history. In some cases shooting for the middle meant you’d end up with 4 cores, while other times it meant 6, 8 or 12. We settled on trying to shoot for a $4000 configuration each time and never go above it. It turns out that if you always had a $4000 budget for a Mac Pro and tried to optimize for CPU performance you’d end up with a somewhat bizarre upgrade path. The path we took is listed in the table below:

Mac Pro - Mid-Range Configuration Upgrade Path
  CPU Chips Cores per Chip Total Cores / Threads Clock Base/1CT/MaxCT Launch Price
Mid 2006 Xeon 5160 2 2 4 / 4 3.0/3.0/3.0 GHz $3299
Early 2008 Xeon E5472 2 4 8 / 8 3.0/3.0/3.0 GHz $3599
Early 2009 Xeon W3580 1 4 4 / 8 3.33/3.6/3.46 GHz $3699
Mid 2010 Xeon W3680 1 6 6 / 12 3.33/3.6/3.46 GHz $3699
Mid 2012 Xeon E5645 2 6 12 / 24 2.4/2.67/2.4 GHz $3799
Late 2013 Xeon E5-1650 v2 1 6 6 / 12 3.5/3.9/3.6 GHz $3999

Around $4000 the Mac Pro went from a quad-core system to eight-cores, back down to four cores, then up to six, then twelve and finally settling back at six cores this generation. What this means is a cycling between improving single and multithreaded performance over the course of the past 7 years:

Here’s where the comparison gets really interesting. If you spent $3799 on a Mac Pro last year, in order to see a multithreaded performance uplift on the CPU side you’d need to spend more this year. Single threaded performance on the other hand sees a big uptick compared to last year. The 2012 $4K config is the outlier however, if you have a budget fixed at $4000 then a 2013 Mac Pro will be quicker in all aspects compared to any previous generation Mac Pro at the same price point.

The bigger takeaway from this is the following: the very same limited gains in CPU performance will eventually come to ultra mobile devices as well. It’s only a matter of time before those CPU curves flatten out. What that does to the smartphone/tablet market is a discussion for another day.

Introduction, the Hardware, Pricing & Config Plotting the Mac Pro’s GPU Performance Over Time
Comments Locked

267 Comments

View All Comments

  • FunBunny2 - Tuesday, December 31, 2013 - link

    Has everybody forgotten? This is just a Cube with one round corner. I suppose Tim will claim that's been patented too.
  • newrigel - Wednesday, March 1, 2017 - link

    Right.... with a unified core in it he he.... Mac's rule
  • Y0ssar1an22 - Tuesday, December 31, 2013 - link

    Off the Mac Pro topic but how come the 2013 13" rMBP scores significantly lower than the 2012 and various MBAs in the Cinebench 11.5? I'm personally interested as I have one on order :-) It scores better in later tests (so presumably not a typo?) Cinebench caught my eye as the first cross-benchmark in the review.

    Thanks for this review, and looking forward to the rMBPs in depth!

  • iwod - Tuesday, December 31, 2013 - link

    1. What are the likely chances of a Mac that does Desktop Class Gfx card with 2 x8 PCI-E and uses Desktop Haswell instead. Unless i miss anything surely this is a simple change in production line.
    2. SSD speed is slow, for a Peak rate of 2GB/s, it seems Apple firmware or Samsung Controller not capable of feeling up the peak bandwidth? So which is likely the cause?
    3. GFx ECC Ram. How much of a problem is it? For Professional market? And why Apple decide to ditch this since the price difference are minor for the price of Mac Pro.
  • dwade123 - Tuesday, December 31, 2013 - link

    Who the **** put a trashcan here!?
  • e375ued - Wednesday, January 1, 2014 - link

    Is there some convenient reason Anand let the Mac Pro off easy by using Prime95 instead of Intel Burn Test or linpack?
  • Ryan Smith - Wednesday, January 1, 2014 - link

    It was my suggestion to try maxing out the Mac Pro, just to see if it would throttle (and if so, by how much). I picked Prime95 because it's good enough; not that there's anything wrong with IBT or Linpack, but all 3 of those are close enough that it shouldn't matter (and P95 is easy to use).
  • jrs77 - Wednesday, January 1, 2014 - link

    Good test that shows that the thermal core design works like a charm, even when applying very heavy and rather unrealistic loads to the system.

    Most people will run these new Mac Pros with only having a scene rendered or a video-filter applied etc and in this case the system is basically dead-silent and street-noise totally drowns the noise of the fan anyways.

    Just a tad too expensive for me tho.
  • Kevin G - Wednesday, January 1, 2014 - link

    The ‘mid range’ config is a far better value on the 2012 model since it is a 12 core model. The $200 savings can be put toward a better GPU.

    With regard to Cinebench, does it use AVX under OS X? I suspect that it does and that is where the majority of the single threaded CPU performance increase comes from. I strongly suspect that the single threaded performance advantage is far narrower in legacy code that doesn’t take advantage of AVX.

    I’m glad the 2012 model was tested with a Radeon 7950. The ability to upgrade GPU’s matters and it’ll keep the 2012 model competitive for awhile. The system will support future video cards that come in from the PC side of things. With UEFI on video cards now, there is little difference between a Mac and PC version. For what it is worth, I have stuck an EVGA GTX 770 into a 2012 Mac Pro without issue and no modification on the video card or OS X drivers. It just works.

    A bit of a random note is that the GPU connector used in the Mac Pro isn’t new to Apple: they used it for the G4 class daughter cards form 15 years back.

    The PLX chip doesn’t have to do any port switching as a single GPU can drive up to 6 surfaces. That would imply the six DP signals from one GPU are routed in pairs to each of the Falcon Ridge controllers for encapsulation.

    One shocking thing is that wall power draw exceeds that of the PSU’s DC rating. That is worrying as the system itself has only a 450W rated power supply. Due to the AC to DC conversion, there is an efficiency factor but the system has to be running close to its DC limit. Performing several file transfers over powered Thunderbolt devices could put the power draw beyond the rated DC limit. I wonder if Apple has implemented throttling based upon raw power consumption of the system as a whole in addition to temperature and power consumption of individual parts. Perhaps testing the system on a 240V AC circuit would alter things here as it is more efficient power delivery?

    One aspect not accounted for is memory expansion. The 2009/2010/2012 Mac Pro’s will work with registered ECC memory which brings their maximum capacity up to 128 GB. Memory bandwidth too is superior in the dual socket 2010/2012 models: six channels of 1333 Mhz memory does have more bandwidth than four channels at 1866 Mhz. Going multi-socket does carry some overhead but still a bit of a disappointment that the theoretical number didn’t improve.
  • Bill Thompson - Wednesday, January 1, 2014 - link

    My guess is the nVidia-based iMac is faster with After Effects and Premiere because of CUDA.

    Davinci Resolve has been updated for OpenCL, but I don't think Octane or Adobe apps have.

    BTW, FCP X 10.1 displays multiple 4K streams in real time without rendering. It's a serious app.

Log in

Don't have an account? Sign up now