Benchmarks : Intel Mobile Ivy Bridge: Worthy upgrade? (Turbo)

What is it?

Ivy Bridge (IVB) is the next generation Core CPU from Intel (v3 2012) replacing the current Sandy Bridge (SNB) (v2 2011) line-up in both desktop and mobile platforms. While socket compatible (LGA 1155 desktop), not all SNB laptops may be upgradeable - unlike desktops.

It does not introduce any new instruction sets - that will wait for Haswell - but it does bring a new integrated GPU unit (Series 4000, versus Series 3000/2000 on SNB) as well as an updated transcoding unit (QuickSync). Manufactured using a newer process (22nm vs 32nm), it consumes less power resulting in lower TDP at the same speed (77W vs 85W on the desktop) or higher Turbo headroom for both CPU and GPU at the same TDP (e.g. 25W or 35W on mobile).

In this article we test CPU, (GP)GPU and memory performance to see whether the upgrade from SNB to IVB is worth it for performance laptops/mobile workstations.

Hardware Specifications

We are comparing a SandyBridge i7-2720QM Mobile (2.2GHz) with the identically clocked IvyBridge i7-3620QM; note that its direct replacement IvyBridge i7-3720QM Mobile runs at 2.60GHz (18% faster). We wanted a clock-for-clock comparison.

Processing Performance

We are testing native, SIMD and cryptography performance using the highest performing instruction sets (AVX, AES, etc.). We are also testing virtualisation performance in .Net/CLR and Java/JVM.

Results Interpretation: Higher values (GOPS, MB/s, etc.) mean better performance.

Environment: Windows 7 x64 SP1, latest Intel drivers. Turbo / Dynamic Overclocking was enabled on both configurations. See here for Non-Turbo Performance.

Processing Benchmarks		Intel i7-3620QM (IvyBridge Mobile)	Intel i7-2720QM (SandyBridge Mobile)	Comments
	Native CPU Arithmetic (Dhrystone/Whetstone)	103.6 / 80.3 (91.2 GOPS) [+25%]	84 / 63 (72.8 GOPS) (baseline)	25% higher performance at the same TDP is a great improvement. Alternatively, by reducing Turbo power limit the same performance can be obtained at a lower TDP.
	Java Arithmetic (Dhrystone/Whetstone)	142.5 / 40 (75.5 GOPS) [+28%]	117.4 / 29.8 (59.1 GOPS) (baseline)	Java JVM performance improves by a similar amount, 28%, thus Java apps benefit as much as native ones by the upgrade.
	.Net Arithmetic (Dhrystone/Whetstone)	15.5 / 53 (28.6 GOPS) [+29%]	12.3 / 40 (22.2 GOPS) (baseline)	.Net CLR performance improves by a similar amount: Windows 8 WPF/WinRT apps benefit just as much which is great news for those upgrading to Windows 8 and beyond.

Simply replacing the SNB CPU to IVB, all performance (native, Java, .Net) improves by at least 25% - while rated speed, TDP and presumably price are maintained. As Turbo was enabled, this may not be due to core improvements - just better Turbo heardroom/more aggressive scaling. At the end of the day, the end-user would not care how better performance is achieved.

SIMD Benchmarks		Intel i7-3620QM (IvyBridge Mobile)	Intel i7-2720QM (SandyBridge Mobile)	Comments
	Native CPU Multi-Media SIMD	192.7 / 261.9 / 149 (224.6 Mpix/s) [+33%]	156.5 / 181.5 / 102.7 (168.5 Mpix/s) (baseline)	Using AVX SIMD, performance improves even further to 33% vs 25% - same TDP! Modern apps using SIMD (SSE2 and now AVX) will benefit even more from the IVB upgrade.
	Java Multi-Media	34.89 / 27.6 / 29 (31.8 Mpix/s) [+37%]	25 / 21.2 / 21.6 (23.2 Mpix/s) (baseline)	While Java code is not vectorised (sadly still not SIMD-like intrinsics), the performance improves even more: 37%!
	.Net Multi-Media	37 / 12.3 / 29 (32.8 Mpix/s) [+42%]	26.4 / 10 / 20 (23 Mpix/s) (baseline)	.Net does not have vectorisation yet either, but the CLR clearly benefits from IVB: a massive 42% improvement is observed simply by upgrading to IVB! Windows 8 users rejoyce! While performance may not get reach native code performance, at least new apps should run quite a bit faster on IVB than SNB.

SIMD and mathematical code in both Java and .Net improves even more by at least 33% on IVB: as this kind of code is usually the most time-consuming and most optimised (the reason it is vectorised/SIMD'd/multi-threaded in the first place) it is unusual to see large improvements from newer revisions - never mind from just a die shrink!

With future operating systems (like Windows 8) favouring .Net (either WPF or WinRT) or Java software, IVB's overall improvement will only increase, from 25% all the way up to 40%!

Cryptography Benchmarks		Intel i7-3620QM (IvyBridge Mobile)	Intel i7-2720QM (SandyBridge Mobile)	Comments
	Native CPU Crypto (Encryption/Decryption/Hashing)	6.25 / 1000 (2.6) [+23%]	6.26 / 730 (2.11) (baseline)	While AES crypto performance has not improved (memory bandwidth limited), AVX hashing performance is enough to improve the performance by 23% overall.

Most Popular Processors, Video Adapters

Наиболее популярные процессоры, протестированные пользователями (за последние 30 дней):		Наиболее популярные видеокарты, протестированные пользователями (за последние 30 дней):
1. 7% Intel Core i7-3930K 597,20 USD 2. 4% Intel Core i7 920 414,41 USD 3. 4% Intel Core i5-2500K 265,59 USD 4. 3% Intel Core i7-2600K 523,94 USD 5. 3% AMD Phenom II X4 965 99,99 USD		1. 6% AMD Radeon HD 7900 289,99 USD 2. 6% AMD Radeon HD 6800 214,85 USD 3. 4% AMD Radeon HD 6900 363,98 USD 4. 3% NVIDIA GeForce GTX 580 517,94 USD 5. 3% NVIDIA GeForce GTX 560 Ti 267,30 USD
For a complete list of statistics, check out the Most Popular Hardware page. For a list of more products, see SiSoftware Shopping.

Memory Performance

We are testing the memory bandwidth, cache bandwidth (L1D, L2, L3) as well as cache and memory latencies using all the access patterns supported by Sandra (in-page random, full random and sequential/linear access patterns).

Results Interpretation: Higher values (MB/s) mean better performance.

Base 2 Multipliers: 1MB/s = 1024kB/s, 1kB/s = 1024bytes/s, 1byte = 8bits, etc.

Environment: Windows 7 x64 SP1, latest Intel drivers. See here for Non-Turbo Performance.

Memory and Cache Benchmarks		Intel i7-3620QM (IvyBridge Mobile)	Intel i7-2720QM (SandyBridge Mobile)	Comments
	Memory Bandwidth	20.55 GB/s (+0.2%)	20.5 GB/s (baseline)	No increase in memory bandwidth or efficiency. Memory controller seems unchanged.
	OpenCL (GPCPU) Memory Bandwidth	13.2 GB/s (10%)	12 GB/s (baseline)	OpenCL is still not as efficient as Sandra in transferring memory thus improves a bit on IVB; most likely due to the improved memory prefetchers. This comes handy in APU mode when using both CPU and GPU to process data.
	Cache Bandwidth (L1D / L2 / L3)	466 / 260 / 152 GB/s (+15%)	401 / 222 / 132 GB/s (baseline)	Cache performance has increased at all levels, with around 15% better performance at levels. Whether they are genuinely faster or run faster due to Turbo remains to be seen.
	Memory and Cache Latency - In-Page Random (L1D / L2 / L3)	25.4ns 2.7 / 8.2 / 16 clocks (+7%)	23.8ns 2.7 / 8.2 / 12.6 clocks (baseline)	While L1D and L2 latencies are the same, L3 and memory latencies have increased from SNB: we are told these are "phantom" latency increases due to the updated prefetchers and should not occur in "normal usage". Let's hope that's true!
	Memory and Cache Latency - Full Random (L1D / L2 / L3)	64.9ns 2.7 / 8.2 / 21.1 clocks (-12%)	73.2ns 2.7 / 8.2 / 20.2 clocks (baseline)	IVB does better out-of-page than in-page: either the page-walk hit is lower or the updated prefetchers are better in handling random accesses. In any case, software using large memory blocks but not large pages (due to Windows issues) should perform faster.

Memory performance is a mixed bag; cache performance bandwidth does increase at all levels (L1D, L2, L3) but that is due to Turbo. The ports are still 128-bit and we need to wait for Haswell for true 256-bit memory transfers. Some latencies go up, others down - overall the updated prefetchers should help but it is a small change.

While SNB crushed its predecesors in cache and memory performance at all levels (either desktop or mobile), IVB can only overtake SNB by running Turbo longer/more often. Otherwise performance is very similar.

Graphics (GPGPU) Performance

Ivy Bridge comes with an enhanced internal DirectX 11 compliant GPU that, for the first time, includes GPGPU capabilities: OpenCL and DirectX ComputeShader. While Sandy Bridge's DirectX 10.1 compliant GPU did report GPGPU capabilities - they were actually emulated in software thus running on the CPU!

Ivy Bridge is effectively an "APU", going head-to-head with AMD designs, most likely "Llano" at the low-end and future "Trinity" (Bulldozer successor FX Series with integrated GPU) at the high-end. By harnessing both CPU and GPU processing power, it can outperform Sandy Bridge simply by using the processing performance of the GPU in addition to the CPU!

Here we test Ivy Bridge's GPGPU and GPAPU performance against Sandy Bridge's GPCPU performance.

Results Interpretation: Higher values (MB/s) mean better performance.

Base 2 Multipliers: 1MB/s = 1024kB/s, 1kB/s = 1024bytes/s, 1byte = 8bits, etc.

Environment: Windows 7 x64 SP1, latest Intel drivers. See here for Non-Turbo Performance.

GPU/APU Processing Benchmarks		Intel 4000 Series (IvyBridge Mobile)	Intel 3000 Series (SandyBridge Mobile)	Comments
	Video DirectX Shading	132.1 / 38.4 (71.3) [+2.47x]	77.8 / 10.7 (28.9) (baseline)	IVB's series 4000 GPU's DirectX 11 shaders are at around 2.5x faster than SNB's DirectX 10.1 ones. While SNB shaders do not support GPGPU, this test shows their level of performance if they were to have had the capability.
	GPGPU DirectX ComputeShader	249 / 51.6 (113.4 MPix/s) [+6x]	62.5 / 4.2 (16.1 Mpix/s)* (baseline)	While SNB emulates GPGPU on the CPU, IVB's native GPGPU support soars to 6x faster performance! The CUs support for native double (FP64) support means that it would be faster than dedicated mainstream GPUs that lack native double support.
	GPGPU OpenCL	230.2 / 16.5 (62 Mpix/s) [+3.9x]	n/a	Unfortunately the OpenCL driver does not support native double (FP64) so while the native float (FP32) performance is similar to DirectX ComputeShader, the overall score is much lower. It is still competitive against SNB's OpenCL CPU score (62 vs. 67) - which may dissapoint some people but for integrated GPU it is just fine.
	GPAPU (CPU+APU) OpenCL	297 / 60.4 (134 Mpix/s) [+8x]	n/a	Harnessing both CPU and GPU increases performance even further: if you consider SNB's OpenCL CPU performance as baseline, IVB APU is 2 times faster! For dual-core CPUs, the difference would be much greater.

The integrated GPU of IVB might not seem powerful enough for some people, but with native double GPGPU support (FP64) it can trade blows the best quad-core mobile CPU on the market (yes, we're talking about Ivy Bridge); if you consider the dual-core version, it would be 3x faster!

As mentioned, many mainstream (middle-range) dedicated GPUs from either AMD or nVidia lacks double native (FP64) support, the integrated IVB GPU may well be faster in double precision than these dedicated GPUs! Even desktop dedicated GPUs that do support double native (e.g. latest nVidia Kepler aka GeForce 680 GTX) do so with such few shaders (1/24 of FP32) than they may just not bother. It seems that GPGPU is moving backwards, with users required to buy expensive workstation GPUs (e.g. nVidia Tesla) for GPGPU - with mainstream (even high-end) GPUs relegated to games.

In APU (GPU+CPU) mode, IVB is 2x faster than SNB's CPU. Considering TDP, speed and price are the same as SNB, this is a huge improvement from a minor architecture revision.

GPU/APU Cryptography Benchmarks		Intel 4000 Series (IvyBridge Mobile)	Intel 3000 Series (SandyBridge Mobile)	Comments
	GPGPU DirectX ComputeShader	822 / 1000 (941 MB/s) [+2.8x]	132 / 832 (331 MB/s)* (baseline)	Against SNB's CPU emulation, IVB's native GPGPU is almost 3x as fast; for such algorithms (crypto/hash) memory performance also matters which is why we don't see the 6x improvement we saw with GPGPU processing performance. 3x is still a massive improvement whatever the reason
	GPGPU OpenCL	744 / 1000 (881 MB/s) [+2.66x]	n/a	OpenCL is not as fast as ComputeShader - yet - but still 2.66x faster, again, a massive improvement from SNB's CPU.
	GPAPU (GPU+APU) OpenCL	1230 / 2000 (1540 MB/s) [+4.6x]	n/a	In APU mode, even sharing the same memory controller, IVB is almost 5x faster than SNB. Whatever the algorithm, APU mode is compelling.

The integrated GPU of IVB impresses whatever the algorithm; in APU mode, IVB is between 5x to 8x faster in GPGPU mode than SNB's CPU - all for the same TDP, speed and cost.

GPU/APU Memory Benchmarks		Intel 4000 Series (IvyBridge Mobile)	Intel 3000 Series (SandyBridge Mobile)	Comments
	Video Memory Bandwidth	22.2 / 5.6 (11.2 GB/s) [+1.83x]	14.8 / 2.5 (6.1 GB/s) (baseline)	Using the same memory installed and same timings, IVB manages to improve internal memory bandwidth by 50% while transfers (though not zero-copy) are twice as fast. Let's hope future drivers will be able to improve effciency further.
	GPGPU OpenCL Bandwidth	22.7 / 13 (17.2) [+43%]	12 / 12 (12)* (baseline)	The increase in internal memory bandwidth makes even OpenCL overall bandwidth 43% better than SNB. GPGPU applications working with lots of memory can definitely appreciate these improvements.

While we have seen native memory bandwidth unchanged - the new IVB GPU has higher internal and transfer bandwith than the SNB GPU; while it would be nice and support zero-copy (since both CPU and GPU share the same memory) - the GPU now matches the native memory bandwidth which is no mean feat for an APU design.

Efficiencies

Because not all things in life are evaluated to their true value, the next measurements will take into consideration various efficiency aspects:

CPU Power Efficiency (this measures the efficiency of power design, or TDP)		Intel i7-3620QM (IvyBridge Mobile)	Intel i7-2720QM (SandyBridge Mobile)	Comments
	Native CPU Arithmetic Efficiency	91.2 GOPS / 45W 2.03 GOPS/W [+25%]	72.8 GOPS / 45W 1.63 GOPS/W (baseline)	As both CPUs are rated for the same TDP, efficiency increases in line with performance: we see 25% increase with the move to 22nm.
	Native CPU Multi-Media SIMD Efficiency	224.6 Mpix/s / 45W 4.99 Mpix/sW [+33%]	168.5 Mpix/s / 45W 3.74 Mpix/sW (baseline)	IVB improves on the SNB's already impressive performance per Watt by 33% making competitors lifes more difficult.
	Native CPU Crypto Efficiency	2.6GB/s / 45W 57.8MB/sW [+23%]	2.1GB/s / 45W 46.7MB/sW (baseline)	The smallest improvement of all benchmarks, it is still respectable: 23%.
GPU/APU Power Efficiency (this measures the efficiency of power design, or TDP)		Intel 4000 Series (IvyBridge Mobile)	Intel 3000 Series (SandyBridge Mobile)	Comments
	GP APU/GPU Processing	134 Mpix/s / 45W 2.98 Mpix/sW [+8.3x]	16.1 Mpix/s / 45W 0.36 Mpix/sW (baseline)	SNB's emulation of GPGPU features on CPU means low performance and thus low power efficiency. It may seem unfair comparing it with IVB's APU mode, but if you were to run GPGPU software this is the kind of improvement you would see. The result exceeds what APU competition can muster (~2.39Mpix/sW AMD Llano mobile) by around 25%!
	GP APU/GPU Crypto	1540 MB/s / 45W 34.2 MB/sW [+4.6x]	331 MB/s / 45W 7.36 MB/sW (baseline)	Again, APU mode on IVB utterly demolishes what little SNB could achieve in GPCPU mode. Against the APU competition (~28.31MB/sW AMD Llano mobile), IVB remains 21% more efficient.

IVB finally allows Intel to be power efficient against APU competition; while the SNB CPU, by itself, was still competitive even against a CPU+GPU (e.g. AMD Llano), IVB's GPGPU now ensures it is more efficient by at least 20% (on the quad-core mobile devices platform).

CPU Cost Efficiency (this measures 'bang-per-buck', i.e. cost effectiveness)		Intel i7-3620QM (IvyBridge Mobile)	Intel i7-2720QM (SandyBridge Mobile)	Comments
	Native CPU Arithmetic Efficiency	91.2 GOPS / $378 241 MOPS/$ [+25%]	72.8 GOPS / $378 192 MOPS/$ (baseline)	As both CPUs have launched at the same price, efficiency increases in line with performance: we see the very 25% increase with the move to 22nm.
	Native CPU Multi-Media SIMD Efficiency	224.6 Mpix/s / $378 0.59 Mpix/s$ [+33%]	168.5 Mpix/s / $378 0.44 Mpix/s$ (baseline)	IVB 33% better value for money, or 33% more for FREE!
	Native CPU Crypto Efficiency	2.6GB/s / $378 6.87 MB/s$ [+23%]	2.1GB/s / $378 5.55 MB/s$ (baseline)	23% is the smallest increase so far, but every improvement helps!
GPU/APU Cost Efficiency (this measures 'bang-per-buck', i.e. cost effectiveness)		Intel 4000 Series (IvyBridge Mobile)	Intel 3000 Series (SandyBridge Mobile)	Comments
	GP APU/GPU Processing	134 Mpix/s / $378 0.35 Mpix/s$ [+8.3x]	16.1 Mpix/s / $378 0.04 Mpix/s$ (baseline)	With competitor CPUs (e.g. AMD Llano 0.42 Mpix/s$) being about 1/2 price even the APU cannot make IVM more cost effective, but at least they are competitive while SNB had no chance. High end CPUs (or APUs here) have always commanded a price premium and they have never meant to be cost effective.
	GP APU/GPU Crypto	1540 MB/s / $378 4.07 MB/s$ [+4.6x]	331 MB/s / $378 0.87 MB/s$ (baseline)	Again, the competitors are much cheaper (e.g. AMD Llano 5.66 MB/s$) and thus more cost effective but again at least IVM is competitive.

SiSoftware Official Ranker Scores

• Intel(R) Core(TM) i7-3920XM CPU @ 2.90GHz (IvyBridge)(4C 8T 3.6GHz Turbo, 3.6GHz IMC, 4x 256kB L2, 8MB L3)
• Intel(R) Core(TM) i7-3820QM CPU @ 2.70GHz (IvyBridge)(4C 8T 3.7GHz Turbo, 4x 256kB L2, 8MB L3
• Intel(R) Core(TM) i7-3720QM CPU @ 2.60GHz (IvyBridge)(4C 8T 3.6GHz Turbo, 4x 256kB L2, 6MB L3)
• Intel(R) Core(TM) i7-3612QM CPU @ 2.10GHz (IvyBridge)(4C 8T 3.1GHz Turbo, 4x 256kB L2, 6MB L3)

Final Thoughts / Conclusions (Turbo performance)

Turbo IVB CPU performance - at the same TDP, speed and cost - is over 20% up to 40% better than SNB. The IVB GPU that now supports GPGPU allows it to operate in APU mode which further increases the performance delta - 2x better - making dedicated GPUs obsolete (save the ridiculous multi-GPU mobile monsters). All from a die shrink (32nm > 22nm) and some optimisations - not a new Core version.

IVB also brings quad-core performance to the 35W level (i7-3612QM) that can be packed into thinner and smaller laptops, e.g. 11-12" (like Alienware M11x) - unheard of just a few years ago! Its GPU may even be enough for low-resolution gaming and GPGPU applications without the need for a dedicated GPU! So from a dual-core + GPU for 50W we have quad-core inc. GPU for 35W!

Let's remember than SNB itself obsoleted previous Core 1 designs both desktop (Nehalem/Lynnfield/Westmere) and mobile (Westmere) - now IVB raises the bar even higher. SNB's GPU itself was a significant upgrade over previous Intel integrated GPUs (Westmere) - and now Intel brings GPGPU capabilities to all platforms, from tablets to laptops/netbooks and desktop.

While we are not comparing its performance with the competition in this article, it is safe to say that the gap has widened further in Intel's favour; the only weak link, the GPU, is now competitive and even has some tricks up its sleeve (e.g. internal cache and double native support).

If your system can support IVB and it does not have a dedicated GPU, then it would be a worth-while upgrade. While most SNB laptops (6X series chipset) should be upgradeable to IVB, most are not - either due to BIOS or VRM limitations. Unlike desktop platforms where vendors have released IVB compatible mainboards - we have not seen mainstream laptops claiming compatibility. Certainly, the laptops we tried (Alienware M14X and Lenovo T420) did not boot with the IVB CPU installed.

We also need to remember that its direct replacement (3720QM) clocks 18% faster (2.6GHz vs. 2.2GHz), thus adding the 20% clock-for-clock advantage, IVB would be 42% faster - a huge improvement.

If you are looking to upgrade to a new IVB laptop (or Ultrabook) - and you do not have a SNB laptop, then it is a no-brainer. If you do have a SNB laptop already without a dedicated GPU then, again, IVB is a worth-while upgrade; if it has a dedicated GPU then IVB's upgraded GPU won't matter. As a platform, IVB (and thus chipset 7X) does not bring much to the table compard to SNB: native USB 3.0 (XHCI) support. Even then, SNB laptops already have a 3-rd party USB 3.0 controller or an ExpressCard slot that can host an USB 3.0 adapter. Older laptops, though are not likely to have SATA3/600 support, a SSD or fast DDR3 memory - to sustain the bandwidth an USB 3.0 device.

There you have it: SNB users with dedicated graphics might as well wait for Haswell; everybody else will appreciate IVB's features and performance.