Core i7-4770K: Haswell's Performance, Previewed
Table of contents
A recent trip got us access to an early sample of
Intel’s upcoming Core i7-4770K. We compare its performance to Ivy
Bridge- and Sandy Bridge-based processors, so you have some idea what to
expect when Intel officially introduces its Haswell architecture.
We recently got our hands on a Core i7-4770K, based on Intel's Haswell micro-architecture. It’s not final silicon, but compared to earlier steppings (and earlier drivers), we’re comfortable enough about the way this chip performs to preview it against the Ivy and Sandy Bridge designs.
Presentations at last year's Developer Forum in San Francisco taught us as much as there is to know about the Haswell architecture itself. But as we get closer to the official launch, more details become known about how Haswell will materialize into actual products. Fortunately for us, some of the first CPUs based on Intel's newest design will be aimed at enthusiasts.
According to Intel’s current plans, you’ll find dual- and quad-core LGA 1150 models with the GT2 graphics configuration sporting 20 execution units. There will also be dual- and quad-core socketed rPGA-based models for the mobile space, featuring the same graphics setup. Everything in the table above is LGA 1150, though. All of those models share support for two channels of DDR3-1600 at 1.5 V and 800 MHz minimum core frequencies. They also share a 16-lane PCI Express 3.0 controller, AVX2 support, and AES-NI support. Interestingly, four of the listed models do not support Intel's new Transactional Synchronization Extensions (TSX). We're not sure why Intel would want to differentiate its products with a feature intended to handle locking more efficiently, but that appears to be what it's doing.
The
much-anticipated GT3 graphics engine, with 40 EUs, is limited to
BGA-based applications, meaning it won’t be upgradeable. Intel will have
quad-core with GT3, quad-core with GT2, and dual-core with GT2 versions
in ball grid array packaging. GT3 will also make an appearance in a
BGA-based multi-chip package that includes a Lynx Point chipset. That’ll
be a dual-core part, though.
In addition to the processors Intel plans to launch here in a few months, we’ll also be introduced to the 8-series Platform Controller Hubs, currently code-named Lynx Point. The most feature-complete version of Lynx Point will incorporate six SATA 6Gb/s ports, 14 total USB ports (six of which are USB 3.0), eight lanes of second-gen PCIe, and VGA output.
Eight-series chipsets are going to be physically smaller than their predecessors (23x22 millimeters on the desktop, rather than 27x27) with lower pin-counts. This is largely attributable to more capabilities integrated on the CPU itself. Previously, eight Flexible Display Interface lanes connected the processor and PCH. Although the processor die hosted an embedded DisplayPort controller, the VGA, LVDS, digital display interfaces, and audio were all down on the chipset. Now, the three digital ports are up in the processor, along with the audio and embedded DisplayPort. LVDS is gone altogether, as are six of the FDI lanes.
Although Dhrystone isn’t necessarily applicable to real-world performance, a lack of software already-optimized for AVX2 means we need to go to SiSoftware’s diagnostic for an idea of how Haswell’s support for the instruction set might affect general integer performance in properly-optimized software.
The Whetstone module employs SSE3, so Haswell’s improvements over Ivy Bridge are far more incremental.
Sandra’s Multimedia benchmark generates a 640x480 image of the Mandelbrot Set fractal using 255 iterations for each pixel, representing vectorised code that runs as close to perfectly parallel as possible.
The integer test employs the AVX2 instruction set on Intel’s Haswell-based Core i7-4770K, while the Ivy and Sandy Bridge-based processors are limited to AVX support. As you see in the red bar, the task is finished much faster on Haswell. It’s close, but not quite 2x.
Floating-point performance also enjoys a significant speed-up from Intel’s first implementation of FMA3 (AMD’s Bulldozer design supports FMA4, while Piledriver supports both the three- and four-operand versions). The Ivy and Sandy Bridge-based processors utilize AVX-optimized code paths, falling quite a bit behind at the same clock rate.
Why do doubles seem to speed up so much more than floats on Haswell? The code path for FMA3 is actually latency-bound. If we were to turn off FMA3 support altogether in Sandra’s options and used AVX, the scaling proves similar.
All three of these chips feature AES-NI support, and we know from past reviews that because Sandra runs entirely in hardware, our platforms are processing instructions as fast as they’re sent from memory. The Core i7-4770K’s slight disadvantage in our AES256 test is indicative of slightly less throughput—something I’m comfortable chalking up to the early status of our test system.
Meanwhile, SHA2-256 performance is all about each core’s compute performance. So, the IPC improvements that go into Haswell help propel it ahead of Ivy Bridge, which is in turn faster than Sandy Bridge.
The memory bandwidth module confirms our findings in the Cryptography benchmark. All three platforms are running 1,600 MT/s data rates; the Haswell-based machine just looks like it needs a little tuning.
We already know that Intel optimized Haswell’s memory hierarchy for performance, based on information discussed at last year’s IDF. As expected, Sandra’s cache bandwidth test shows an almost-doubling of performance from the 32 KB L1 data cache.
Gains from the L2 cache are actually a lot lower than we’d expect though; we thought that number would be close to 2x as well, given 64 bytes/cycle throughput (theoretically, the L2 should be capable of more than 900 GB/s). The L3 cache actually drops back a bit, which could be related to its separate clock domain.
It still isn’t clear whether something’s up with our engineering sample CPU, or if there’s still work to be done on the testing side. Either way, this is a pre-production chip, so we aren’t jumping to any conclusions.
The Core i7-4770K plants itself right between Intel's four-core Core i7-3770K and six-core -3970X.
More aggressive threading places Haswell ahead of Ivy Bridge in this OCR-based test, but quite a ways behind the flagship CPU representing Sandy Bridge-E.
Like LAME on the previous page, our iTunes benchmark is single-threaded. In this test, however, we're leaving Turbo Boost enabled. Higher IPC throughput pushes Core i7-4770K into a lead. The Sandy Bridge-E-based Core i7-3970X doesn't benefit from its six cores, and slides back to third place. It's only faster than the Core i7-2700K because of a higher single-core Turbo Boost frequency.
Haswell is certainly faster than Ivy and Sandy Bridge, but our Core i7-4770K cannot catch the Core i7-3970X in a benchmark able to utilize all processing resources.
Again, Core i7-3970X proves the viability of an older Sandy Bridge-E configuration in threaded software. The more mainstream processors trail behind.
Expert
Thanks for reading!
We recently got our hands on a Core i7-4770K, based on Intel's Haswell micro-architecture. It’s not final silicon, but compared to earlier steppings (and earlier drivers), we’re comfortable enough about the way this chip performs to preview it against the Ivy and Sandy Bridge designs.
Presentations at last year's Developer Forum in San Francisco taught us as much as there is to know about the Haswell architecture itself. But as we get closer to the official launch, more details become known about how Haswell will materialize into actual products. Fortunately for us, some of the first CPUs based on Intel's newest design will be aimed at enthusiasts.
According to Intel’s current plans, you’ll find dual- and quad-core LGA 1150 models with the GT2 graphics configuration sporting 20 execution units. There will also be dual- and quad-core socketed rPGA-based models for the mobile space, featuring the same graphics setup. Everything in the table above is LGA 1150, though. All of those models share support for two channels of DDR3-1600 at 1.5 V and 800 MHz minimum core frequencies. They also share a 16-lane PCI Express 3.0 controller, AVX2 support, and AES-NI support. Interestingly, four of the listed models do not support Intel's new Transactional Synchronization Extensions (TSX). We're not sure why Intel would want to differentiate its products with a feature intended to handle locking more efficiently, but that appears to be what it's doing.
The
much-anticipated GT3 graphics engine, with 40 EUs, is limited to
BGA-based applications, meaning it won’t be upgradeable. Intel will have
quad-core with GT3, quad-core with GT2, and dual-core with GT2 versions
in ball grid array packaging. GT3 will also make an appearance in a
BGA-based multi-chip package that includes a Lynx Point chipset. That’ll
be a dual-core part, though.In addition to the processors Intel plans to launch here in a few months, we’ll also be introduced to the 8-series Platform Controller Hubs, currently code-named Lynx Point. The most feature-complete version of Lynx Point will incorporate six SATA 6Gb/s ports, 14 total USB ports (six of which are USB 3.0), eight lanes of second-gen PCIe, and VGA output.
Eight-series chipsets are going to be physically smaller than their predecessors (23x22 millimeters on the desktop, rather than 27x27) with lower pin-counts. This is largely attributable to more capabilities integrated on the CPU itself. Previously, eight Flexible Display Interface lanes connected the processor and PCH. Although the processor die hosted an embedded DisplayPort controller, the VGA, LVDS, digital display interfaces, and audio were all down on the chipset. Now, the three digital ports are up in the processor, along with the audio and embedded DisplayPort. LVDS is gone altogether, as are six of the FDI lanes.
Results: Sandra 2013
Although Dhrystone isn’t necessarily applicable to real-world performance, a lack of software already-optimized for AVX2 means we need to go to SiSoftware’s diagnostic for an idea of how Haswell’s support for the instruction set might affect general integer performance in properly-optimized software.
The Whetstone module employs SSE3, so Haswell’s improvements over Ivy Bridge are far more incremental.
Sandra’s Multimedia benchmark generates a 640x480 image of the Mandelbrot Set fractal using 255 iterations for each pixel, representing vectorised code that runs as close to perfectly parallel as possible.
The integer test employs the AVX2 instruction set on Intel’s Haswell-based Core i7-4770K, while the Ivy and Sandy Bridge-based processors are limited to AVX support. As you see in the red bar, the task is finished much faster on Haswell. It’s close, but not quite 2x.
Floating-point performance also enjoys a significant speed-up from Intel’s first implementation of FMA3 (AMD’s Bulldozer design supports FMA4, while Piledriver supports both the three- and four-operand versions). The Ivy and Sandy Bridge-based processors utilize AVX-optimized code paths, falling quite a bit behind at the same clock rate.
Why do doubles seem to speed up so much more than floats on Haswell? The code path for FMA3 is actually latency-bound. If we were to turn off FMA3 support altogether in Sandra’s options and used AVX, the scaling proves similar.
All three of these chips feature AES-NI support, and we know from past reviews that because Sandra runs entirely in hardware, our platforms are processing instructions as fast as they’re sent from memory. The Core i7-4770K’s slight disadvantage in our AES256 test is indicative of slightly less throughput—something I’m comfortable chalking up to the early status of our test system.
Meanwhile, SHA2-256 performance is all about each core’s compute performance. So, the IPC improvements that go into Haswell help propel it ahead of Ivy Bridge, which is in turn faster than Sandy Bridge.
The memory bandwidth module confirms our findings in the Cryptography benchmark. All three platforms are running 1,600 MT/s data rates; the Haswell-based machine just looks like it needs a little tuning.
We already know that Intel optimized Haswell’s memory hierarchy for performance, based on information discussed at last year’s IDF. As expected, Sandra’s cache bandwidth test shows an almost-doubling of performance from the 32 KB L1 data cache.
Gains from the L2 cache are actually a lot lower than we’d expect though; we thought that number would be close to 2x as well, given 64 bytes/cycle throughput (theoretically, the L2 should be capable of more than 900 GB/s). The L3 cache actually drops back a bit, which could be related to its separate clock domain.
It still isn’t clear whether something’s up with our engineering sample CPU, or if there’s still work to be done on the testing side. Either way, this is a pre-production chip, so we aren’t jumping to any conclusions.
Results: OpenCL Performance
Intel enabled OpenCL 1.1 support on its Ivy Bridge-based
processors with HD Graphics 4000 and 2500, giving developers an option
to exploit the graphics component’s execution units for general-purpose
workloads. Popular desktop applications like WinZip and Photoshop now
offer sometimes-substantial performance gains on platforms able to more
granularly parallelize workloads that would have previously been handled
by fewer processing cores. With Haswell, support is being expanded to
OpenCL 1.2.
Our Photoshop CS6 benchmark is most effective at showing the difference between processors that lack OpenCL support and those with it. The Core i7-2700K tackles this workload using its four Hyper-Threaded cores, while the -3770K and -4770K get their HD Graphics components involved.
The Haswell-based Core i7-4770K is slightly faster than its predecessor, likely due to a combination of additional EUs, more bandwidth, and higher IPC.
We run our WinZip test with and without OpenCL enabled on all processors, and you can clearly see there isn’t as much differentiation as there was in Photoshop. The explanation is easy enough, though. WinZip 17 is really well-threaded (much more so than 16.5 was). So, the CPU cores are taxed, even without OpenCL support. With OpenCL turned on, WinZip only offloads compression for files larger than 8 MB. So, if our 1.3 GB folder of files is full of documents, PowerPoint presentations, PDFs, and music (which it is), acceleration isn’t going to help much.
We do observe small speed-ups from the Core i7-4770K and -3770K, whereas the -2700K actually slows down when we try turning OpenCL on. The moral of the story? OpenCL is only going to register as a benefit insofar as the tasks you run are well-suited to heterogeneous computing. The Photoshop benchmark represents one end of that spectrum, and our WinZip test illustrates the other.
LuxMark 2.0 quantifies the speed-up from HD Graphics 4000 to 4600, simultaneously reminding us that the Core i7-2700K, for as capable as it is, doesn’t help in OpenCL-enabled software. As a side note, AMD's A10-5800K registered 225,000 samples per second, less than the Core i7-3770K.
Now, with that said, is OpenCL always going to be the performance win that each of our tests seems to show? Not necessarily. As we see in Sandra 2013’s GP Processing module, FP32 math is significantly faster on Intel’s HD Graphics engine than its x86 cores. However, doubles have to be emulated on all three processors, and the Sandy Bridge-based Core i7-2700K turns in better results there. It turns out that Intel’s powerful x86 cores emulate those results faster than Ivy Bridge or Haswell can on the GPU.
At least in our single-threaded LAME conversion test, Haswell is just over 3% faster than Ivy Bridge and over 5% faster than Sandy Bridge.
In Blender 2.64, the quad-core Haswell-based part is more than 7% quicker than Core i7-3770K and 14% faster than Core i7-2700K. At the same time, a stock Core i7-3970X is still more than 23% faster than Core i7-4770K. If you were hoping that Haswell would offer an inexpensive quad-core substitute for what will be the two-generation-old Sandy Bridge-E architecture, it’d appear that the six-core design will continue to make sense for anyone with a true workstation.
Our Visual Studio-based Chrome compile benchmark would seem to concur. Think the -3770K and -2700K seem too close together? That’s what I thought at first, until I looked back to this image from our Core i7-3770K launch and saw the same thing. In comparison, Haswell has a huge impact in performance, pretty much cutting the gap between Sandy Bridge and Sandy Bridge-E in half. The six-core chip still reigns supreme in workstation-oriented tasks, but the Core i7-4770K’s >13% advantage over -3770K is stellar.
Our Photoshop CS6 benchmark is most effective at showing the difference between processors that lack OpenCL support and those with it. The Core i7-2700K tackles this workload using its four Hyper-Threaded cores, while the -3770K and -4770K get their HD Graphics components involved.
The Haswell-based Core i7-4770K is slightly faster than its predecessor, likely due to a combination of additional EUs, more bandwidth, and higher IPC.
We run our WinZip test with and without OpenCL enabled on all processors, and you can clearly see there isn’t as much differentiation as there was in Photoshop. The explanation is easy enough, though. WinZip 17 is really well-threaded (much more so than 16.5 was). So, the CPU cores are taxed, even without OpenCL support. With OpenCL turned on, WinZip only offloads compression for files larger than 8 MB. So, if our 1.3 GB folder of files is full of documents, PowerPoint presentations, PDFs, and music (which it is), acceleration isn’t going to help much.
We do observe small speed-ups from the Core i7-4770K and -3770K, whereas the -2700K actually slows down when we try turning OpenCL on. The moral of the story? OpenCL is only going to register as a benefit insofar as the tasks you run are well-suited to heterogeneous computing. The Photoshop benchmark represents one end of that spectrum, and our WinZip test illustrates the other.
LuxMark 2.0 quantifies the speed-up from HD Graphics 4000 to 4600, simultaneously reminding us that the Core i7-2700K, for as capable as it is, doesn’t help in OpenCL-enabled software. As a side note, AMD's A10-5800K registered 225,000 samples per second, less than the Core i7-3770K.
Now, with that said, is OpenCL always going to be the performance win that each of our tests seems to show? Not necessarily. As we see in Sandra 2013’s GP Processing module, FP32 math is significantly faster on Intel’s HD Graphics engine than its x86 cores. However, doubles have to be emulated on all three processors, and the Sandy Bridge-based Core i7-2700K turns in better results there. It turns out that Intel’s powerful x86 cores emulate those results faster than Ivy Bridge or Haswell can on the GPU.
Results: Performance Teaser, Per-Clock Perf And Threaded Apps
Per-Clock Performance
First, let's take a look at how Haswell fares against Ivy and Sandy Bridge at a constant 3.5 GHz, with power-saving features and Turbo Boost disabled.At least in our single-threaded LAME conversion test, Haswell is just over 3% faster than Ivy Bridge and over 5% faster than Sandy Bridge.
Threaded Performance
How about when we turn on all of the chip’s features, and let the Core i7-4770K stand up against Ivy Bridge, Sandy Bridge, and Sandy Bridge-E?In Blender 2.64, the quad-core Haswell-based part is more than 7% quicker than Core i7-3770K and 14% faster than Core i7-2700K. At the same time, a stock Core i7-3970X is still more than 23% faster than Core i7-4770K. If you were hoping that Haswell would offer an inexpensive quad-core substitute for what will be the two-generation-old Sandy Bridge-E architecture, it’d appear that the six-core design will continue to make sense for anyone with a true workstation.
Our Visual Studio-based Chrome compile benchmark would seem to concur. Think the -3770K and -2700K seem too close together? That’s what I thought at first, until I looked back to this image from our Core i7-3770K launch and saw the same thing. In comparison, Haswell has a huge impact in performance, pretty much cutting the gap between Sandy Bridge and Sandy Bridge-E in half. The six-core chip still reigns supreme in workstation-oriented tasks, but the Core i7-4770K’s >13% advantage over -3770K is stellar.
Results: More Common Desktop Apps
The Core i7-4770K plants itself right between Intel's four-core Core i7-3770K and six-core -3970X.
More aggressive threading places Haswell ahead of Ivy Bridge in this OCR-based test, but quite a ways behind the flagship CPU representing Sandy Bridge-E.
Like LAME on the previous page, our iTunes benchmark is single-threaded. In this test, however, we're leaving Turbo Boost enabled. Higher IPC throughput pushes Core i7-4770K into a lead. The Sandy Bridge-E-based Core i7-3970X doesn't benefit from its six cores, and slides back to third place. It's only faster than the Core i7-2700K because of a higher single-core Turbo Boost frequency.
Haswell is certainly faster than Ivy and Sandy Bridge, but our Core i7-4770K cannot catch the Core i7-3970X in a benchmark able to utilize all processing resources.
Again, Core i7-3970X proves the viability of an older Sandy Bridge-E configuration in threaded software. The more mainstream processors trail behind.
Expert
Thanks for reading!
