Intel Haswell CPUs Debut, Put To the Test

jjslash writes "Intel's Haswell architecture is finally available in the flagship Core i7-4770K and Core i7-4950HQ processors. This is a very volatile time for Intel. In an ARM-less vacuum, Intel's Haswell architecture would likely be the most amazing thing to happen to the tech industry in years. Haswell mobile processors are slated to bring about the single largest improvement in battery life in Intel history. In graphics, Haswell completely redefines the expectations for processor graphics. On the desktop however, Haswell is just a bit more efficient, but no longer much faster when going from one generation to another." Reader wesbascas puts some numbers on what "just a bit" means here: "Just as leaked copies of the chip have already shown, the i7-4770K only presents an incremental ~10% performance increase over the Ivy Bridge-based Core i7-3770K. Overclocking potential also remains in the same 4.3 GHz to 4.6 GHz ballpark."

Comments

Transactional Memory support

[rev0lt]

For me, this is by far the biggest architectural improvement I see in these line of processors (check http://en.wikipedia.org/wiki/Transactional_Synchronization_Extensions and http://software.intel.com/sites/default/files/m/9/2/3/41604 for more information). If it sticks, it will help solving a lot of multi-core shared memory software development issues.

Re:Transactional Memory support

[Z00L00K]

It's an interesting addition which can be useful for some.

But when it comes to general performance improvement it's rather disappointing. Looks like they have fine tuned the current architecture without actually adding something that increases the performance at the same rate as we have seen the last decades. To some extent it looks like we have hit a ceiling in increased performance with the current overall computer architecture and that new approaches are needed. The clock frequency is basically the same as for the decade old P4, the number of running cores on a chip seems to be limited too, at least compared to other architectures.

One interesting path for improving performance that may be useful is what Xilinx has done with their Zynq-7000 which combines ARM cores with FPGA, but it will require a change in the way computers are designed.

Re:Transactional Memory support

[PhrostyMcByte]

A more immediately useful feature is backwards-compatible hardware lock elision. Before taking a lock, you emit an instruction which is a NOP for older CPUs but causes Haswell to ignore the lock and create an implicit transaction. Instant scalability improvement to just about every app out there with contention, without having to distribute Haswell-specific binaries.

My favorite feature, though, is scatter/gather support for SIMD. Scatter/gather is very important because up until now loading memory from several locations for SIMD use has been a pain in the ass involving costly shuffles and often requires you to load more than you actually immediately wanted, possibly forcing you to spill registers. It's really not something you want to do, but sometimes there are no good alternatives. I'll be super interested to see benchmarks taking this into account.

Re:Transactional Memory support

[Hamsterdan]

I'm pretty sure if AMD would bring out something to be competitive, Intel would find a way to come out with faster processors, just like they did when the original Athlon kicked ass. Suddenly they were able to manufactures processors that were'nt just 33Mhz faster than the previous ones.

Re:Transactional Memory support

[TheRaven64]

The hardware lock elision stuff is going to be more than just a little bit useful. It means that software that uses coarse-grained locking can get the same sort of performance as software using fine-grained locking and close to the performance of software written specifically to support transactional memory. It will be interesting to see if Intel's cross-licensing agreements with other chip makers includes the relevant patents. If it's something that is widely adopted, then it is likely to change how we write parallel software. If not, then it will just make certain categories of code significantly more scalable on Intel than other CPUs.

Re:Transactional Memory support

[bored]

Just to reply again, some of these benchmarks are obvious bullshit. Like the AES one, the new cpu's have AES-NI instructions for accelerating AES.

So, yes if you happen to be doing AES, and your running code that can take advantage of AES-NI then the new CPU's are going to fly. But the whole benchmark is so tilted its not even funny. Why not use a benchmark that renders some SSL encoded web pages? Because the benchmark is going to be bottlenecked by the network stack and the rendering engine. Not the tiny percentage of the time the CPU spends doing AES.

He really should have broken the benchmarks into two camps, benchmarks actually running the same code on both sets of CPU's and ones that can leverage some new instructions available only on the new CPUs.

Re:Transactional Memory support

[Z00L00K]

Of course the performance has improved since the P4, but the point is that it has been by tweaking things like cache, parallel execution with discard of unwanted branches, pre-fetch, various types of pipelining, out of order execution and so on.

All this means that in order to achieve high performance you have an architecture that does a lot of stuff that eventually is getting thrown away because it was done just in case. The catch here is that it costs energy and builds complexity. The benefit is of course that you can get a processor that executes multiple instructions per clock cycle which improves performance.

However there's a limit on the performance gain these improvements can provide.

Re:Transactional Memory support

[cnettel]

I wouldn't agree. The out of order depth has increased in later architectures, but not by that much. Rather, branch prediction is better, so we have fewer mis-predicts. We have wider execution units (thus necessitating the higher depth), but that's a "real" improvement in performance. The number of instructions retired per clock under ideal conditions per core, is higher. Hyperthreading has been kept, but it is implemented in a smarter manner with a more dynamic sharing of resources. All of this is happening in the area of "fat" cores and naturally we are reaching some point where there is little more that can be done, but typical energy use under idle and modest load has been decreasing, not increasing. If you just want the performance of a 65 W Core 2 Duo, you can certainly get that out of the new 35 W desktop SKU and get pretty close to it with the 17 W laptop ones under most conditions as well. It is "only" a decrease by 50-75 % in 6 years or so, but it is far from "nothing changes". And, yeah, by the way, you get integrated graphics within that 35 W TDP as well.

Re:Transactional Memory support

[gnasher719]

The clock frequency is basically the same as for the decade old P4, the number of running cores on a chip seems to be limited too, at least compared to other architectures.

P4 was an architecture where clock frequency was the design goal, with no regards for actual performance. Any non-trivial operation used lots of cycles, because one cycle was just too short to do useful work. A simple shift instruction was four cycles. An integer multiplication nine cycles. Since Banias, the design goal was performance, not clock speed. The amount of work done in a cycle is vastly increased. The clock speed of P4 and current processors is just not comparable.

On the Macintosh side, Apple shipped pre-release Intel Macs with 3.6 GHz P4s to developers. The first real hardware with 1.83 GHz Core Duos ran _faster_. But if you look at benchmarks, current high-end consumer Macs run about 15 times faster again!

To maybe make a stronger impression: If Apple replaces the processor in the fastest iMac with a Haswell chip, you'll get a computer that would make it into the top 100 of the June 2000 "Top 500 Supercomputer" list. That's how fast a modern Intel computer is, compared to a P4.

Re:Transactional Memory support

[PhrostyMcByte]

I see this problem time and again, where a programmer that just doesnt get it is trying to use SSE

Do you teach? I don't see that very often outside of SIMD newbies. Curiosities aside, as I said:

It's really not something you want to do, but sometimes there are no good alternatives.

There are some algorithms, like EWA/Jinc image resampling, which simply don't translate well to SIMD without gather support. Keeping with the image processing theme, sRGB gamma decompression will be lightning fast with a L1-friendly 1KiB LUT -- I bet faster than any low-precision approximation in use today.

This is not about trying to fit AoS into a SIMD world. This is about enabling some things in SIMD which simply weren't possible before. Newbies will no doubt use it improperly, but then those with experience and perhaps a little imagination will use it to make code that screams.

Better AoS support shouldn't really be discounted, either -- some standard data formats outside of the programmer's control are really unfriendly to SIMD, and scatter/gather might not give an 8x speedup in processing it but it will certainly help.

Re:Transactional Memory support

[bored]

The app I work on needs a couple GB/sec (not bit) of encryption bandwidth. So, I love having AES-NI.... But.... We have encryption/compression acceleration hardware in the machine. Which means we won't be using AES-NI because the CPU's haven't increased their compression speeds to match their new found encryption speeds. Plus, we (like many other things, including truecrypt) have always offered the user the ability to encrypt with algorithms that were more CPU friendly than AES.

Anyway, two points. In the high bandwidth web servers I've seen, the SSL public key handshake part of the initial negotiation was the problem with SSL, not the AES/etc. As such, i've seen accelerators for SSL setup too! Frankly, its possible to get ~100MB/sec out of even older CPU's doing software AES. So a 8x CPU machine can pretty much keep a 10Gb link saturated if all your doing is AES... Doesn't leave much time for anything else, but you can either pay for more CPU's or buy accelerators if you actually need the bandwidth. Frankly, adding a couple extra CPU's is not really going to be your problem if your transferring more than 10Gbit of web traffic.

But this is all xeon territory, not desktop CPU's. Which is where truecrypt comes in. If your running it on a recent SSD and you just have to have AES instead of something faster, AES-NI is going to be your thing. But its quite possibly the one desktop application with the largest performance difference between a modern CPU and one a couple generations back. Your not going to be pulling 10x+ performance increases on many other applications. So putting it in a general cpu comparison except as a footnote is a little disingenuous. Same for some of these other things where openCL/CUDA offloads have become the norm so the new AVX/etc instructions aren't necessarily such a big win. Not that I mind having them either!

Performance per Watt

[Technician]

Hmm, Performance per Watt seems to have been glazed over.

The possibility of a fanless media center PC, the ability of a server farm to eliminate over half the cooling cost, and long battery life in a gaming class laptop seems to not be the attention of the article.

Gee it's only X percent faster..

Re:Performance per Watt

[hpa]

I do kernel hacking on airplanes. 'nuff said.

Re:Performance per Watt

[Kjella]

Anandtech tested it, idle power is down probably due to the new voltage regulator (FIVR) but active power.... 113% the performance for 111.8% so performance per watt is essentially unchanged. If what you need is CPU power then you're better off waiting for a IVB-E hex-core in Q3, in threaded applications a quad-core Haswell won't touch a hex-core Ivy Brigde - it's trailing Sandy Bridge hex-cores as well. If you're not interested in the graphics or battery life, it's a giant yawn.

That said, the GT3e graphics for mobile looks to carve out a solid niche in the notebook market, the R-series desktop processors (GT3e graphics, BGA only) is probably compelling for AIOs that don't have room for graphics upgrades anyway and the lower idle wattage should be good for all laptops with Haswell graphics. None of the processors launched now have the new idle states for ultramobile/tablets, so the effect of those we'll have to wait to see. Anandtech tested the i7-4950HQ and it was impressive how a 47W mobile chip consistently beat AMDs A10-5800 100W desktop APU in gaming benchmarks. Of course it's going to sell in a price range of its own, but AMD just lost the crown here.

As a CPU in a regular tower with discrete graphics it's at best incremental but I think the full launch lineup hit all of Intel's main competitors - it's threatening AMD and nVidia's low end discrete card sales, it's threatening AMDs APU sales and the lower idle power is promising for their lower power parts that will compete with ARM. They're just not winning much against the i7-3770K but then they're also fighting against themselves in that market, the FX-8350 is not even close. The 8-series chipset finally brings 6 SATA3 ports, so the main AMD advantage chipset-wise also disappeared.

Re:Performance per Watt

[0123456]

Fact is, the emphasis is still on processing power over performance per watt. Sure, Intel gives it lip service, but not much beyond that.

So why does my new i7 system use half as much power as my old Pentium-4?

Software killed the PC, not hardware

[Anonymous Coward]

The lack of phenomenal hardware improvements may annoy the nerds, but the mass market PC is killed by the abysmal software environment. People are fleeing to tablets and phones and with that the cloud because maintaining a PC has become just about impossible for laymen. The slowness of a desktop that hasn't seen professional maintenance is astonishing, if it is still working at all. Viruses and trojans aside, every bit of software comes with its own updater, many of which are poorly implemented resource and attention hogs. If the updater doesn't do you in, it's the bundled adware, sometimes installed by the update "service". The PC is stiff and stone cold, a host overwhelmed and killed by its parasites. Time to put it 6ft underground.

Re:Software killed the PC, not hardware

[PopeRatzo]

Time to put it 6ft underground.

This was the giveaway.

What do you care if there are still people who would rather use a desktop PC that's not behind a garden wall and actually get work done? Why do you insist that the PC platform has to be killed off? Isn't there room in the world for more than one set of computing needs?

This notion, that only the most popular form of anything should exist pops up strangely often around here. The iPad is phenomenal, so Android tablets should just disappear from the market. The iPhone is popular so no Windows phones can be allowed. That sort of thing.

Friend, I can understand that you'd rather work on a tablet and have someone else make decisions about what you can and cannot have, what you can and cannot do, but why in the world are you so insistent that no one else be able to make their own choice?

I don't get you.

Re:Software killed the PC, not hardware

[etash]

do you know how many people have declared the PC to be dead. It's usually either people who have "better" solutions to offer, or the useful idiots who believe them. Ever tried doing some real work on a tablet ? like video editing, image editing, mundane tasks like excel, word editing. how about video games ?

Re:Software killed the PC, not hardware

[Seumas]

The 40 or 50 some-odd million Steam gamers and the often 6 million concurrent online Steam gamers would beg to differ with you and, anecdotally, the new people I hear every day saying that they're building a PC gaming rig for the first time in their life is surprising and seems to be growing. There are a ton of gamers out there and they have always been a heavy push for the PC market. They aren't going anywhere. They continue to be just as important and just as numerous, even in a world full of consoles and tablets and cell phones.

Also, most people who would ONLY use a phone or a tablet are not people who would have bought a desktop, anyway. They MIGHT have otherwise had a low power laptop... MAYBE... but they were never going to be users going out and building/buying meaningful desktop systems. These aren't losses from one platform fleeing to another.

Re:Software killed the PC, not hardware

[Seumas]

I don't buy the "consoles are killing PCs" argument for one second (also, people have spent the last two years claiming that tablets and phones are killing consoles, so there you go).

Steam *alone* has something like 40 or 50 million users and between 2.5 and 6 million concurrent users playing a game at any one moment.

We don't even need to talk about the depth and breadth of games available on PC that simply don't exist on other platforms. Just with the user numbers alone -- and only the Steam ones, here -- compare those to consoles and you simply can not argue that PC games are dying. I think that a lot of factions (including publishers, themselves) are doing every fucking thing they can to invalidate, complicate, and ultimately kill PC gaming and that has been the case for a few years . . . but it isn't changing the fact that it's pretty fucking vibrant and charging forward.

How does this compare

[maroberts]

With AMDs CPU/GPU solutions?

Re:How does this compare

[beelsebob]

Nope, the Iris 5200 wipes the floor with the A10. You're probably referring to the low end HD 4600 in the i7 (which they put there because they expect no one buying an i7 to be using integrated graphics).

Re:How does this compare

[Rockoon]

It's also only in $400+ mobile i7s.

Intel fanboys never want to discuss price, at least not really. Sometimes they pretend, but never want to make an honest price bin comparison.

Re:How does this compare

[Kjella]

Intel fanboys never want to discuss price, at least not really. Sometimes they pretend, but never want to make an honest price bin comparison.

So let's compare single threaded CPU performance at equal price points then shall we. No? Ah right, the comparison of choice lately has been gaming performance at equal price points. Choose the scenario where you win and pretend it's the only scenario that matters, it's a powerful form of self-delusion. But that aside, they don't compete directly just yet but Intel just proved they have an integrated graphics solution that is faster and uses less power than anything AMD can come up with. That is what most people would call "technical superiority" where you can extract tons of profit by offering products your competitors can't. Intel has a lot of that, AMD just lost one of the last strong cards in their deck.

Take a grounded reality check - is AMD's pricing because they're this undiscovered treasure chest of excellent processors or because they're selling inferior solutions they won't sell any other way? I think you can certainly rule out that it's a strategic plan from AMD's side, given how their latest financial statements look. Intel is not pricing Haswell aggressively at all, if anything they're increasing prices so that maybe AMD can manage to not go bankrupt and get Intel in all sorts of scrutiny over their x86 monopoly. And still AMD is failing, really if they can't manage to undersell Intel at the prices Intel have now maybe AMD doesn't deserve to be in business. Sad for the consumer, but true.

Re:How does this compare

[beelsebob]

Wrong. We discuss it. But when we start talking about price you people go back to this. Exactly this.

The i7-4770 costs 4+ times as much (currently if you pre-order) as the most expensive AMD A10 chip, which also arent even the fastest AMD chips available.

i7 4770k –$339
A10 5800k - $129.99
So we're looking at a chip that's 2.6 times the price, but 4 times the CPU performance.

Price was brought up, so you responded by not mentioning price at all, and proceeded to compare the highest performing Intel chip ever with a mid-range AMD part.

No, the assertion that the intel chip was 4 times more expensive was brought up, and neglected to mention the context that the intel chip was also 4 times faster (and also greatly exaggerated the 2.6 times price). I merely noted that there's a reason for the intel chip being significantly more expensive – that it's significantly faster.

You're right, the AMD chip is indeed a mid range chip, and it should never even have been compared to the i7. Unfortunately, AMD has no chip that can reasonably be compared to the i7, so all the review sites chose the fastest CPU AMD has on offer that includes an IGP. If you want to push it a bit further, lets ignore the IGP completely, and take the fastest desktop chip AMD has on offer:
FX 8350 –$199.99
i7 4770k –$339
So, we're at 1.9 times the price, and around 1.8 times the performance (see here).
But, of course, we all know that price increases exponentially with performance when it comes to high end processors, so lets just scale back that intel part a bit:
FX 8350 –$199.99
i5 4670 –$213
So now we're at 1.06 times the price (effectively equal), but about 1.5 times the performance (see here).
If you really want to make the price war completely in intel's favour, then look at
FX 8350 –$199.99
i5 4570 - $192
Unfortunately Anand do not list this in their benchmarks, but it's the exact same die as the 4670, just clocked 6% lower, so we can extrapolate the result, and say that this i5 will be around 1.4 times the speed of the FX 8350 Oh, and by the way, it throws in a decent IGP, while the FX 8350 does not. It also consumes only 67% of the power under load, and that's including the IGP.

So conclusion. Haswell is pretty effectively blowing AMD out of the water. The A10 5800 remains reasonably unscathed only because Haswell i3s have not yet been released. When the i3 4220 appears, I would fully expect the A10 to lose to it in every respect in the same way as the FX 8350 loses to the i5 3570 in pretty much all ways.

Re:Progress

[ArcadeMan]

The hype is really getting repetitive.

Re:Graphics..

[timeOday]

You must have said that before looking at the benchmarks? Looks to me like AMD is toast. Intel's integrated graphics beat AMD's on every game in the AnandTech test.

The new Intel even beats the discrete mobile GPU (Geforce GT 650M) on a couple tests. On most the Intel is somewhat slower but using around half the power.

Need to wait a few years

[gnasher719]

In the last few years, Intel has been adding new instructions that will give major performance gains when they are used. For example, Haswell can do two fused multiply-adds with four double or eight single precision operands per cycle per core, but no current code will use this. We'll get the advantage when HPC code is recompiled (in a few months time), and when general code assumes that everyone has this feature (in five years time). But on the other hand, we _now_ get the advantages of features they added five years ago.

Re:Need to wait a few years

[Anonymous Coward]

This is absolutely the most overlooked aspect of Haswell. As an incremental improvement, its less than stellar, but in certain areas, it effectively doubles performance over the previous generation. Aside from FMA for floating point apps, the integer 256bit SIMD pipeline is now effectively feature complete.

Your point about waiting for recompiles is a good one - all the more reason we should be moving to adaptable metaprogramming systems for HPC, rather than a constant manual reworking of codebases. Projects like Terra (http://terralang.org) are particularly promising in this regard

Heat Dissapation

[stms]

How hot do these chips get. I have to throttle my (not overclocked) 3770k when I do media encodes because it gets too damn hot. I've been meaning to get a better cooler I just haven't gotten around to it.

Why would Intel care

[zrelativity]

Why would Intel care about raw CPU performance. They have no competition from AMD in CPU performance. The GPU performance may not be as good as A10, but it has improved and that's what matters for Intel.

Intel for a little while has correctly perceived that their risk to business is from shift in computing to mobile devices and they are addressing this issue. One thing Intel has always been very good at, and I'm a great admirer of them for that, when they perceive a risk, they are extremely good at steering their giant ship very rapidly into the headwind and tackle that threat. Their process technology lead also gives them a huge advantage.

Over the next couple of years, the battle front will be the mobile and server devices, the desktop processors will become a second class citizen. Maybe this will give some lifeline to AMD, but AMD is so far behind on performance.

Get A Clue, Intel

[Jane Q. Public]

While I am all for advances in CPUs, I seriously wish Intel would go back to a naming scheme for its CPUs that made any kind of sense to the average buyer (or even the technically-oriented buyer). I have grown really weary of having to look at tables of CPU specifications every time I shop around for computers.

Intel's naming scheme -- expecially in recent years -- has been a mishmash of names and numbers without any obvious coherence. Get a clue Intel. You're hurting your own market.

If I didn't have to run OS X in my business, I'd buy AMD just for that reason. Their desktop CPUs may not be quite up to the latest Intel, but they are certainly adequate and the price is better.

Re:Progress

[Samantha Wright]

The hype may be the most amazing thing to happen to the tech industry in years.

Re:AMD still a LOSER for me

[fnj]

For me. I couldn't care less what the majority of desktop users care about. What I care about is a CPU with high performance per watt and graphics good enough to scroll a text display (editor, eclipse) with some graphics (browser) fast enough not to be annoying, and to watch HD video without any pauses. Sandy bridge was more than good enough in the graphics department for anything I would ever want to do on current displays. Haswell will probably more than maintain this on the highest resolution displays coming down the road in the next few years.

Dollar cheapness of the CPU within limits does not impress me at all. I am generally content with a system for at least 4 years to spread out that cost, and there are many other cost centers in a system besides CPU.

I have never used a single AMD system and see no reason to believe that they will ever make anything that would change my mind.

I don't really begrudge any slack jawed gamers their massive nuclear power plant busting AMD systems with the absurd overkill of space heater SLI graphics; it's just not anything that has the slightest relevance to anything I could ever care about.

Only on some Haswells, not on "K" processors

[niceworkthere]

For some yet unknown reason the unlocked "K" Haswells (and maybe others) were so far all listed to come without TSX.

Depends on your field

[Frobnicator]

If all you care about is the perspective of the boring desktop business app, then this processor doesn't have much to excite you. Of course, that's just one field. Sending a few database queries over the wire or updating your text boxes doesn't exactly saturate a quad-core box. Business desktop apps don't really see much no matter what.

For data-heavy, cache-intensive, and parallel-intensive programs the processor looks to offer quite a lot. HPC developers like that.

For notebooks and low-power devices the processor is wonderful. If you are paying the power bill for a data center, the energy use will add up. Accountants and laptop users will like that.

The option to have graphics integrated to the chip this way means better SOC options. Embedded developers will like that.

Many fields will see great things out of this chip.

If you are fixated on the world of desktop business software, you still get an incremental ~10% improvement. Unlike technologies such as SIMD, you get it without changing a line of code. So now you can add 10% more text boxes to fill out, or maybe pick up some more wasteful coding habits.