Analysis: Haswell For Gaming - When Is It Appropriate to Upgrade?

Written by  Tuesday, 04 June 2013 23:47
  •  

Haswell's officially been out of embargo/NDA for a few days now, TigerDirect has launched its product offerings, and the internet is abuzz about Intel's 4th Gen processors. After reading through the vast majority of reviews on the web, comparing benchmark results, and studying specs, overclocking, and potential shortcomings, we've built up enough confidence to officially comment on the abounding "is Haswell worth it?" question.

intel-haswell-slider

 

As always, Intel's released around a dozen SKUs of their new CPU, two of which are K-SKU OC chips, the rest are scattered between S-series powersaving or no-suffix "casual" computing. Everything looks great on paper:

 

  • The new Z87/Haswell combo has improved memory bandwidth and Haswell has a better memory controller, which means ~3GHz memory overclocks are natively supported by the CPU.
  • Lower idle power consumption (TDP) than ever before, despite increased desktop TDP (from 77W to 84W).
  • Higher IPC (instructions per cycle).
  • New sleep-state tech: DevSleep drive 'deep slumber' tech means SSDs can run at ~0.165mW, using micro-stutters to improve mobile battery life substantially.
  • And more - to include additional BCLK OC options and insanely high multipliers (x80).


Note: for purposes of this article, "mobile" includes all non-desktop computing (tablets, laptops, phones, etc.). We're also assuming that any upgrades would be performed in an equivalent  fashion - i.e., 3770k to 4770k.

It sounds great, and that's without even getting into the IGP discussion. The mobile Iris chip (5200) has been an incredibly promising improvement upon Intel's IGP technology, to say the least; it's not mind-blowing for gamers, but it's noteworthy, and should no longer be shunned. We frequently stated that the HD4000 was very impressive given Intel's previous attempts at graphics accelerators (hint: they sucked), and though we never adamantly recommended gaming on it over Trinity or discrete units, it's still a good step forward. With the new Iris 5200 and other mobile IGP SKUs, we've seen nearly double performance, to the point where some of the higher-end Intel IGPs perform equally to discrete GT 640M GPUs. Definitely impressive, and at the low power demand and smaller physical space requirements, we see diminished thermal generation -- important in laptops for a number of reasons (battery life and chip preservation, to name a few). The desktop-side HD4600 isn't quite as impressive, but it's safe to assume you're going to be using a dedicated card for desktops.

But is Haswell "worth it" for your gaming desktop build? Is it worth upgrading from Ivy Bridge, Sandy Bridge, or Nehalem? This article arms you with the ammunition to answer that question.

Haswell i7-4770K, i5-4670K, and Other 4th Gen Specs 

Model Cores External Clockrate Boost L3 Cache IGP TDP Availability
i7-4770K 4/8 3.5GHz 3.9GHz 8MB HD 4600 84W $350
(TigerDirect Link) 
i7-4770 4/8 3.5GHz 3.9GHz 8MB HD 4600 84W $320
(TigerDirect Link) 
i7-4770S 4/8 3.1GHz 3.9GHz 8MB HD 4600 65W $320
(TigerDirect Link) 
i5-4670K 4/4 3.4GHz 3.8GHz 6MB HD 4600 84W $250
(TigerDirect Link) 
i5-4670 4/4 3.4GHz 3.8GHz 6MB HD 4600 84W $230
(TigerDirect Link) 
i5-4570 4/4 3.2GHz 3.6GHz 6MB HD 4600 84W $210
(TigerDirect Link) 
i5-4570S 4/4 3.0GHz 3.6GHz 6MB HD 4600 65W $200
(TigerDirect Link) 
i5-4430 4/4 3.0GHz 3.2GHz 6MB HD 4600 84W $190
(TigerDirect Link) 

 

Haswell's suffix/naming scheme is fairly straight forward:

  • As with previous generations, i7 is the hyperthreaded chip (ideal for enthusiasts or those performing render operations). i5s are the quad-core chips (ideal for gamers). i3s (not shown) are the dual-core chips, which work for gaming or desktop use.
  • The number works like this: The first digit is the generation (4th Gen), the next three digits are the generic SKU numbers. Any suffixes also mark important items: K is for the unlocked/overclocking class CPUs. You won't get BCLK modifications and high multipliers without K-SKU chips. S is for power saving.
 
As for general buying advice: Prices should normalize at MSRP within coming weeks. Retailers have the procs listed a bit higher than MSRP now due to initial demand/enamored buyers.

Intel Haswell i7-4770K Die Shot

intel-haswell-die-blockOur article on how to read & understand a CPU die block diagram.

History Lessons: Intel's Traditional CPU Advancements and Modern 'Vertical' Domination

A decade or two ago, it wasn't uncommon to hear AMD or Intel boast a 30% or 50% gain in floating-point or INT performance. Conroe was an example of this, where IPC gains easily exceeded 10% of its predecessor. When AMD paved the way for 64-bit adoption, we saw massive potential for the future of computing architecture and its capabilities. The market was volatile, the technology was still relatively untapped, and Intel and AMD were competing with great ferocity (trading blow-for-blow, for a couple of years). Depending on how far you go back, we were still on the verge of inventing highly sophisticated silicon fabrication processes used today, which have significantly enhanced production efficiency and die yield.

Even Nehalem -- much more recent, c. 2009 -- saw 20%-30% gains in some aspects of performance, depending on what you were measuring. That was the First Gen of the Intel i7-series family, laying the groundwork for what we've seen in the past 3 generations of Intel CPUs.

4770k-box

Gains have dwindled since then, though, and a lot of commenters on the 'net have been wondering why. For each generational gap -- Nehalem to Sandy, Sandy to Ivy, Ivy to Haswell -- we're seeing gains (IPC or otherwise) in the 7-13% spectrum. There's good reason for what seems like slow growth, though. Aside from the technology simply being so far beyond what we used to see, which is naturally more difficult to improve upon, we're also seeing a shift of focus by Intel (and AMD) to developing markets. It's no longer necessary to 'simply' make a fast desktop CPU. As far as "normal" consumers and biz-client users (non-production professionals, not extreme gamers) are concerned, they don't need a faster CPU.

As the desktop market continues to bleed down to the skeletal structure of enthusiast- and enterprise-class computing, we see a shift in focus from high-performance to high-mobility; plainly-put, everyday consumer and biz-client users don't need much more computational prowess. They've got what they need. As the market majority's demand for speed subsides, we see new demands emerge. I'm not teaching you something you don't know: If you have a need that's fulfilled, that need is taken for granted, ignored, and superseded by something more immediately attractive and seemingly unattainable. Whether or not the average user comprehends their demands against reality, of course, is another matter; but from a technical perspective, having a faster CPU doesn't mean the average user will have a faster computing experience -- they're going to be throttled by data downstream, input, or I/O bottlenecks first.

As huge a sway on the market as gaming is, we're not alone in the industry. And right now, that "immediately more attractive and unattainable" element—at least for biz-client and consumers, who make up a huge portion of the market—is what we define as "mobility." In general, mobility consists of a few key concepts: Battery life, physical portability (size), silence, and scaled processing power to that of a desktop (within reason).

cadex-labs-chartFrom our Busting Battery Myths article. The fewer charge cycles the battery undergoes, the longer its lifespan.

In and of itself, battery life is largely dictated by the CPU's ability to mitigate power draw, thermal footprint, and integration of advanced power stepping and states. As heat is increased, we see a significant system-wide impact: Fans spin-up to higher RPMs (if possible), nearby components warm up, heatsinks max-out their thermal dissipation capacity, silicon takes a beating, reliability degrades, and thermal interfaces lose their effectiveness.

In practical examples, many of these points are ineptly demonstrated by some of HP's consumer-grade laptops, which are quite possibly some of the most thermally-unsound systems I've ever had the misfortune of testing; when the GPU needs to be reflowed to correct the solder (and thermal compound re-applied), there's a problem. That's not what you want.

Why is any of this relevant to Haswell and enthusiasts, though?

Intel is a big player in the market. The biggest, actually, with a market cap that exceeds nearly all the other public consumer-class hardware manufacturers combined. Like any good company, they're planning for the future: Intel wants to dominate their vertical, which will inevitably lead to a higher shipment volume of mobile SOCs (where Intel builds/dictates the board and the non-socketed CPU), and as Intel moves increasingly more motherboard functions to the chip (VRM now merged, along with the memory controller of boards past), they'll continue to expand their grip on the industry's direction. It's not out of malice, either: Intel picks up where OEMs/ODMs are unable or ill-equipped. They're optimizing, improving efficiency, and making an overall more desirable product. It's good business sense, but it does potentially threaten market diversity. Such is the trade-off.

As Intel moves toward these technologies, we're going to see diminished gains in advanced desktop processing tech as innovation shifts to IGPs and highly-efficient power draw. It's not all about moving toward mobility, though -- we're also hitting a steep hill (it's not quite a 'performance wall') in CPU potential. As former Intel CTO Pat Gelsinger told Anand Shimpi about ten years ago, "you can only integrate a memory controller once." Then it's done. You need to find another innovation, and that's not easy. The unified PCH was another similar step, moving the VRM to the CPU an even newer sample.

The fact that our CPUs are growing more efficient, more integrated, and more powerful means that even overclocking begins to exhibit a slumped performance curve, yielding smaller gains than we're typically used to. We even have a clear path to 8nm fab process - at which point fabrication switches over to more precise ion cannons - but the bolstered transistor count doesn't necessarily produce linear gains in processing power. It depends what those transistors are being used for.

Haswell's Performance vs. Ivy Bridge, Sandy Bridge, and Nehalem

nehalem-dieNehalem is still a damn impressive architecture, but it's a far cry from Haswell. Compare this block diagram with Haswell's -- the differences are clear as day, especially when considering that nearly 1/3 of the Haswell die is dedicated to its IGP.

We're currently inundated with SSD testing -- and as many of our recurring visitors are aware, we're slowly expanding into other review spaces (started in cases, moved to cooling, now moving to SSDs). CPUs and GPUs are complex enough that we really want to make sure our test methodology hits a homerun, so we're not quite there yet; that said, I've read through numerous reviews and have been able to aggregate and mentally process the results, hopefully filtering for more straight-forward information for novice and intermediate builders.

Regarding Intel-only testing, a lot of you have asked whether upgrading from your existingSandyBridgeorIvyBridgemachines is "worth it for gaming." The short answer? No. Saved you some time there -- now let's talk about why.

The iterative jumps from Nehalem to Sandy, Sandy to Ivy, and Ivy to Haswell are all very close in overall performance/IPC gains. As stated earlier, it's consistently in the 10-13% range, generally making upgrades from one-generation-old architecture financially inadvisable. For enthusiasts? Sure, it could be worth it for the fun. But for the average system builder or gamer? Not really.

Backing up my statement, we look at the benchmarks performed by Anandtech and Tech Report:

tech-report-haswellSource: Tech Report's preview. Check their full review here. anandtech-haswellSource: Anandtech's review.

 

If you're on Nehalem or older hardware, it's probably worth considering an upgrade. That's a massive jump, with the i7-950 being improved upon by nearly 35% for multithreaded Cinebench rendering performance. Single-threaded performance is also promising, though decreasingly relevant, marked as a 33% improvement.

As forSandyBridge? Maybe not quite as impressive. The 4770k over the 2700k sees ~14% improvement in multithreaded Cinebench rendering.

Then there'sIvyBridge, where the 4770k over the 3770k sees ~7% improvement in render performance. Things are looking good for SB and IB owners, but it's not all about Cinebench and rendering performance.

Cinebench is a great synthetic benchmark for rendering performance, and as prolific as video production is, most gamers aren't really concerned about that -- they're concerned about gaming performance.

For anyone doing home-brew videos, YouTube work, or streaming, Haswell is a worthwhile consideration. There are noticeable gains in render-out performance that equate to increased efficiency when dealing with a high-volume of videos being produced; that said, if you're working in any sort of professional capacity, there are more render-focused enterprise/performance chips built for that.

Realistically, Cinebench will stress the CPU significantly more than most (or all) games on the market, but it's important to have "real-world" data to put things into perspective.

Is Haswell "Worth It" For Gaming? The CPU Side of Things. 

Once again, the answer to this depends on your current system configuration. Are you onIvyBridgeorSandyBridge? I'll save you the time: Probably not. Nehalem? It's questionable -- that platform is old enough that other things start holding you back anyway, like the shoddy SATA III controllers that fail to produce proper 6Gbps throughput (throttling your SSDs). Still, if you don't care that much about limited SATA throughput (or have one of the few boards that had a good controller), you're still going to be more GPU-bound in most games than CPU-bound.

Let's look at some benches performed by the outlets equipped for CPU cross-analysis, starting with Tech Report's average FPS bench for Metro: Last Light, then leading into Anand's Civ 5 bench (which is CPU-bound, and should provoke bigger differences).

You should go check out the full Tech Report 4770k review if you haven't yet; they've got frametime measurements, full analysis, spec tables spanning multiple platforms, and some of the best analysis we've seen on the web. Anandtech's review is also impressive, though slightly less detailed. Check that over here.

anand-civ5Source: Anandtech. tech-report-metroSource: Tech Report.

 

We see even smaller gains here. When two decimal places are required to differentiate processors, it becomes evident exactly how overpowering our hardware is; the games aren't taking advantage of increasingly more power like we'd hope, and as everything shifts to the GPU, that's not going to change much. CryEngine 3 is one of the first engines that natively supports eight simultaneous threads (though no games currently utilize that tech fully), running three threads by default: A game logic thread, rendering thread, and physics thread. Other threads are spawned and destroyed as necessary, but the games aren't optimized enough where it really matters much beyond the first three logical threads.

Anandtech's test uses a 7970 to ensure the tests aren't GPU-bound, so what you're seeing are true "CPU only" limitations. Take a look at the Q9400 -- pre-Nehalem -- and compare it against the modern CPUs. Sure, you get about twice the performance in Haswell, but you're still above 40FPS.

More has to be taken into account than pure FPS, of course. Anything older than Nehalem (and even Nehalem) won't produce results for modern SSDs, you'll be limited in memory bandwidth, overclocking support, new software-driven technologies (and co-development with gamedevs, like Grid 2's OIT and self-shadowing smoke), and other chipset-driven features.

So is it all worth it? Let's draw out a couple of scenarios:

  • You're a gamer first-and-foremost. You don't care about much else than gaming, at least for the system in question.
  • You're a gamer, but you also do a good deal of video production or other professional tasks (like compiling).
  • You're new to PC gaming systems or otherwise have a need to build a completely new computer. You don't currently have a system that you rely on for gaming.

 

If you're building a brand new gaming PC and don't currently have something running on Nehalem or newer tech, it's almost definitely worth getting Haswell. There's just no real reason to opt for a 3570k when Haswell's 4670k is fresh on the market. Sure, it's a few bucks more ($15-$30 more, depending on where you buy it), but Haswell's going to be more supported going forward, it has undeniably better performance for non-gaming tasks (if only just), and has great power stepping features and future tech support. Overclocking is a matter of contention right now, as overclockers continue to exploit Haswell and determine its limits; from what I've read thus far, OCing on air seems to run significantly hotter than Ivy Bridge and exhibits odd heat spikes, but the performance yields from lower clock frequencies more than make-up for lowered external clockrate.

If you're just gaming and are currently on SB or IB, it's my advice to just skip Haswell. You have no reason to invest in it. Your current setup is powerful enough that the money is best spent elsewhere, like on a more powerful GPU or SSD. Nehalem users have an interesting choice: If you're on one of the RAID controllers that can't push true SATA III speeds, then that alone may make an upgrade worthwhile -- but only if you're doing a lot of IO-intensive tasks. If you're on magnetic drives and/or don't do a lot with IO, then jumping ship for Haswell doesn't seem immediately beneficial (GPU upgrades will output more noticeable gains). Anything older than Nehalem? Yeah, it's probably time to upgrade.

For those who mix gaming tasks with streaming, video production, or other more core- or INT-intensive applications, it's worth considering Haswell. OnIvyBridge, the ~6-10% gains in render performance aren't really worth it -- the money is better saved for render-specific cards or RAID arrays.SandyBridgebegins looking pretty worthwhile, and the 35% improvement over Nehalem is easily worth the jump.

So to recap, anyone building a new Intel-driven rig should be looking at Haswell. Unless you're on a restrictive budget (in which case, talk to us on the forums for help), there's just no good reason to ignore the small gains over Ivy Bridge - especially given support for future Intel-driven technologies. Look at Grid 2, for example: Codemasters worked closely with Intel to develop self-shadowing smoke and integrate order-independent transparency (OIT), lending more depth to smoke and trees, but that tech is only available on Haswell.

It's easy to be disappointed by Haswell - really, it is.IvyBridgewas damn good, so wasSandyBridge, and Nehalem was an incredible stepping stone. Haswell has a lot to live up to, and as we progress in CPU tech, it's not going to get any easier. The IGPs are definitely impressive in their own right, though I've mostly neglected them in favor of desktop focus. For mobile systems and HTPCs, though an entirely different topic, I'll simply say that it's worth investigating Haswell. AMD puts up a damn good fight in the gaming HTPC market, though, if power draw isn't an issue. As for now? It's worth a look - especially for new system builds.

- Steve "Lelldorianx" Burke.

Last modified on Wednesday, 05 June 2013 01:30

From Around the Web