Intel's Haswell CPUs

[encode]

Well, looks like we are in huge stagnation of PC hardware.

Upcoming Intel Core i7 4770K, according to specs, is just the same as 3770K or even 2700K. Probably 5% or less CPU performance gain. Yep, notable stronger iGPU, but again, I doubt I can play Battlefield 3 on it. Sad story.
As both 2700K and 3770K owner I see just a little difference - mainly 3770K has a better RAM management and is faster for about 200 MHz (on clock-per-clock basis) than 2700K. Notable stronger iGPU, yes, but with Z77 MoBos 2700K performs quite well too. (Intel cheaters!) I just compared same 2700K CPU on a few Z68, Z77 and H77 motherboards - they just did some chipset-based stuff and even same Sandy Bridge's iGPU is able to perform notable better. But anyway, the best game I can play on overclocked Ivy Bridge's 3770K iGPU is Unreal Tournament 3. Battlefield 2 at the settings I want (High, 1920x1200) is a slideshow!
As to overclocking, at least my 3770K is not that bad as expected: 4.7+ GHz stable on average air cooler is really nice. 4.7 is at least 4.9 GHz of 2700K. Yes, my 2700K was able to reach 5.3 GHz, but on water. Very soon I'll upgrade myself to Corsair H80i and that will be a fair comparison. Yes, better memory management means much higher freqs of RAM supported, but again ~1900 MHz is enough, so anyways.

Really sad...

I just need a proper video card and I may forget about upgrading for years. Tablets and Smartphones just destroyed PC market very quick! God damn! I'm writing this post via iPad...

[Piotr Tarsa]

Well, AMD isn't a competition (CPU performance wise) to Intel nowadays, except in some multithreaded tasks. On the other hand Intel's iGPUs are still notably slower than AMD's iGPUs (in APUs) in laptops and Haswell is probably going to change that. If Intel will continue developing new manufacturing processes as it does now, ie 14nm will be on track, then probably Intel could make even mid-range laptop GPU segment obsolete. And Intel is almost forced to not go far ahead of AMD, because anti-monopoly laws would cause Intel to split.

On the other hand we have quickly advancing ARM architecture which now has 64-bit version and its implementations are in development. Chromebooks are based on the current, slightly outdated, 32-bit ARM architecture and they are said to have an acceptable performance. It would be interesting for me to see ARM as a competitor to x86 in laptop and desktop segments.

[Matt Mahoney]

Too bad desktops and laptops are on the way out. All the computer makers know it. But I just can't see writing code on a touch screen.

[Bulat Ziganshin]

haswell may be much better than ivy because
1) it should be major architecture upgrade - with better perfromance per Mhz, like sandy was against i7-9x0
2) since 22 nm process is now debugged, it should allow higher frequencies

OTOH, it's still 4 cores, so Ivy-EP with 8 cores on desktop would be a serious competition. and Haswell-EP should be released only in 2014

[Shelwien]

http://en.wikipedia.org/wiki/Haswell...d_new_features

[encode]

haswell may be much better than ivy because...

I bet it won't. Such things looks great on paper, but in practice... Intel already improved its 22nm tech process with the latest 3770K batches - my CPU is a living proof - CPUs from first batches were unable to reach 4.7-4.8 GHz in general.
4770K's integrated VRM is questionable thing as well.
No reason for Intel to compete with themselves. Until AMD will not release a CPU that will vastly outperform Intel's, they will be really relaxed and throw away marketing slogans and only. And yes, they will ragingly try to enter mobile market...

[Piotr Tarsa]

There are some unofficial news about Haswell: http://oclab.ru/news/pervyie-rezulta...-prilozheniyah
Looks like there's nothing much to expect from desktop Haswells.

[nburns]

anandtech's review portrayed the architectural changes in Haswell as being primarily about aggressive power management, clearly as a response to the needs of mobile and the need to stay competitive with ARM. It sounds like Intel has accomplished a great deal in the way of power savings with the new architecture. As far as raw performance, you have to wonder how much more can be done.

At one time, Moore's law drove predictable increases in clock speed. Pentium 4 was the last architecture of that era. After P4, AMD came out with Athlon 64 and Opteron, which brought things like serial interconnects, movement of more components on-die -- most significantly the memory controller -- and 64-bit. That gave an edge in performance to AMD, probably due mostly to the integrated memory controller and the fact that Pentium 4 was saddled with a flawed design. Intel's following architecture, Core 2, managed to leapfrog Athlon64 without matching all of its architectural improvements, apparently mainly by improving the cache subsystem and getting smarter about prefetching and such. By now, of course, Intel's latest cpus have achieved full parity with AMD's feature-wise, as well as continued to improve throughput by, the way I imagine it, a relentless process of combing through the architecture looking for ways to tune and tweak the prefetchers, branch predictors, instruction decoders, etc.

The problem of extracting increasing IPC through architectural changes these days, as I understand it, is not unlike data compression. It comes down to producing a probabilistic model for the stream of instructions so you can predict memory accesses, pipeline stalls, etc. It seems to me that that's a problem for which no amount of investment can guarantee improvements on a predictable schedule.

[Piotr Tarsa]

http://semiaccurate.com/forums/showthread.php?t=7006

My comments after skimming through the Tom's Hardware article (in addition to linked post):
- memory & L3 performance looks lower,
- L1 bandwith 2x higher as promised,
- GT2 in Haswell noticeably faster than HD 4000 - it suggests that GT3 that will come to laptops will bring a huge speedup and can make mobile AMD APU's look silly,
- Haswell is mainly about optimizations for laptops (ie power) and AVX2, which seems a bit weird combination for me,
- speedup is very small on single threaded workload, more on multithreaded ones - this suggests that they improved HyperThreading scalability,


Power draw measures haven't been posted but Intel has promised much lower power consumption.

Overall, to recap:
- very low improvement in high-end chips - not worth to wait,
- probably big improvement in low-power chips,

[Bulat Ziganshin]

plus TSX - disabled though in K and some other chips. AVX2 is also very importnat - it means 2x speed for any SSE integer code, including hashes, reed-solomon coding and so

[Piotr Tarsa]

Depends on what is bottlenecking you. As seen on this page http://www.tomshardware.com/reviews/...ce,3461-2.html AES-NI performance on Haswell noticeably decreased due to decreased memory bandwidth. But that's rather an exception.

[nburns]

"The mobile space is where Intel?s efforts should become more apparent?and it has something for that market we anticipate will give AMD?s and Nvidia?s entry-level GPUs a serious run." -- If Intel made AMD's graphics look silly, that would be a major turn of events. It doesn't sound like that's what THG is anticipating. They're comparing Haswell to AMD's and nVidia's entry-level GPUs, and only predicting that it could be fairly competitive. Or, perhaps I misunderstood your point?

I'm a little concerned that the cpu industry is running out of ideas for how to increase single-threaded cpu performance. As I tried to explain in my last post, increasing clock speed was the easy way to boost performance (perhaps not easy for the electrical and materials science engineers, but the benefits were across-the-board and easy to reap); and I'm afraid that there can only be so many of the hard ways left, like making the circuits ever more complex, and making cpus better at prediction: memory access patterns, branching, etc.

It's hard to imagine that AVX2 is going to be a tremendous improvement. I'm a novice at vectorization. But from looking through the history of MMX, SSE, and AVX, I got the impression that Intel chooses which groups of instructions to add based on how likely they are to be useful. That would explain why, for instance, floating point vector instructions tend to get added a generation or two before similar integer instructions. That suggests that the integer instructions and other things that Intel has saved until now were probably toward the back of the list. I've tried to use vectorization, and my, admittedly, novice efforts only managed to slow things down. For instance, I was playing with Eugene Shelwien's (impressive) BWTMix, and I thought it might benefit from vectorization. I was able to force clang to generate vector instructions (verified from the disassembly), and the resulting code ran slower. It was probably the wrong application (BWTMix mainly had a bunch of shifts and additions, maybe too lightweight in the first place, or perhaps there were not enough of them in a pipeline). On another occasion, I tried using a C++ vectorization library on something I was working on, with the same result: negative performance gain. It's apparent that vector instructions are not a solution for general every-day kinds of performance. They are specialized tools that benefit specific applications, and are probably beyond the capabilities of regular professional programmers (like me) and need to be applied by experts in specialized niches.

It's good that Intel has managed to lower power consumption. Fan speeds can be reduced, noise can be lowered, cases can shrink, and less electricity will be wasted, even on the desktop. Obviously, a good thing for mobile.

http://semiaccurate.com/forums/showthread.php?t=7006

My comments after skimming through the Tom's Hardware article (in addition to linked post):
- memory & L3 performance looks lower,
- L1 bandwith 2x higher as promised,
- GT2 in Haswell noticeably faster than HD 4000 - it suggests that GT3 that will come to laptops will bring a huge speedup and can make mobile AMD APU's look silly,
- Haswell is mainly about optimizations for laptops (ie power) and AVX2, which seems a bit weird combination for me,
- speedup is very small on single threaded workload, more on multithreaded ones - this suggests that they improved HyperThreading scalability,


Power draw measures haven't been posted but Intel has promised much lower power consumption.

Overall, to recap:
- very low improvement in high-end chips - not worth to wait,
- probably big improvement in low-power chips,

[Piotr Tarsa]

If Intel made AMD's graphics look silly, that would be a major turn of events. It doesn't sound like that's what THG is anticipating. They're comparing Haswell to AMD's and nVidia's entry-level GPUs, and only predicting that it could be fairly competitive. Or, perhaps I misunderstood your point?

Yeah, you've missed the point. THG compared GT2 to entry-level desktop GPU's, while I'm talking about comparing GT3 to AMD mobile APUs. GT3 has in general twice of majority of units of GT2, so it should be much faster than GT2. Intel has much lower power hungry CPUs so they can allocate more TDP for iGPU.

It's hard to imagine that AVX2 is going to be a tremendous improvement. I'm a novice at vectorization. But from looking through the history of MMX, SSE, and AVX, I got the impression that Intel chooses which groups of instructions to add based on how likely they are to be useful.

I have an impression that Intel chooses instructions that will get them the best score in popular benchmarks and work with authors of most commonly benchmarked programs to utilize that instructions. Otherwise, there are rarely any benefits from anything newer than SSE2.

I'm a little concerned that the cpu industry is running out of ideas for how to increase single-threaded cpu performance. As I tried to explain in my last post, increasing clock speed was the easy way to boost performance (perhaps not easy for the electrical and materials science engineers, but the benefits were across-the-board and easy to reap); and I'm afraid that there can only be so many of the hard ways left, like making the circuits ever more complex, and making cpus better at prediction: memory access patterns, branching, etc.

Well, it's rather logical that they can't increase the clock infinitely. Current leakage (and therefore power consumption)/ mm^2 is growing and that makes cooling the chip harder. And high temperature forbids high clocks.

The future is in parallelization, but today's approaches are not very comfortable to program for. GPGPUs requires simple control flows, ie very few branches or very skewed branches. Threads on GPGPUs are grouped and if threads in one group branch differently then both branches have to be executed. CPUs on the other hand have relatively high price for synchronization and even then CPUs can run very few threads in parallel, compared to GPGPUs at least. In addition, x86 requires rather strong memory consistency between CPU cores and that increases the price of making many core x86 CPUs. AFAIR Intel removed most of cache consistency semantics within Larrabee, because with so much CPU cores (about 50 AFAIR) cache consistency requirements would kill the whole project (too much power and area dedicated to cache consistency). With the new Xeon Phi's it seems Intel has reintroduced full cache coherency, but Xeon Phi isn't now as impressive. Specifications are here: http://ark.intel.com/products/71992/...53-ghz-60-core 60 cores x 1 GHz on 22nm, while Radeon 7970 is 32 cores (compute units) * 1 GHz on 28nm, so not much more increase in flexibility. Additionally it's rather possible that 22nm server Haswell can have 20 cores at 3 GHz, thus giving comparable peak compute strength and being much easier to program.

I'm keen on message passing and event driven asynchronous architectures but unfortunately there are no hardware acceleration for such things in contemporary CPUs. Going low-level with locks and semaphores always gives me headaches.

[nburns]

Yeah, you've missed the point. THG compared GT2 to entry-level desktop GPU's, while I'm talking about comparing GT3 to AMD mobile APUs. GT3 has in general twice of majority of units of GT2, so it should be much faster than GT2. Intel has much lower power hungry CPUs so they can allocate more TDP for iGPU.

I see.

I have an impression that Intel chooses instructions that will get them the best score in popular benchmarks and work with authors of most commonly benchmarked programs to utilize that instructions. Otherwise, there are rarely any benefits from anything newer than SSE2.

That might not make too terrible a proxy for usefulness, though. Assuming there are any useful instructions left to add. Maybe SSE2 exhausted the potential.

Well, it's rather logical that they can't increase the clock infinitely. Current leakage (and therefore power consumption)/ mm^2 is growing and that makes cooling the chip harder. And high temperature forbids high clocks.

Absolutely.

The future is in parallelization, but today's approaches are not very comfortable to program for. GPGPUs requires simple control flows, ie very few branches or very skewed branches. Threads on GPGPUs are grouped and if threads in one group branch differently then both branches have to be executed. CPUs on the other hand have relatively high price for synchronization and even then CPUs can run very few threads in parallel, compared to GPGPUs at least. In addition, x86 requires rather strong memory consistency between CPU cores and that increases the price of making many core x86 CPUs. AFAIR Intel removed most of cache consistency semantics within Larrabee, because with so much CPU cores (about 50 AFAIR) cache consistency requirements would kill the whole project (too much power and area dedicated to cache consistency). With the new Xeon Phi's it seems Intel has reintroduced full cache coherency, but Xeon Phi isn't now as impressive. Specifications are here: http://ark.intel.com/products/71992/...53-ghz-60-core 60 cores x 1 GHz on 22nm, while Radeon 7970 is 32 cores (compute units) * 1 GHz on 28nm, so not much more increase in flexibility. Additionally it's rather possible that 22nm server Haswell can have 20 cores at 3 GHz, thus giving comparable peak compute strength and being much easier to program.

I'm keen on message passing and event driven asynchronous architectures but unfortunately there are no hardware acceleration for such things in contemporary CPUs. Going low-level with locks and semaphores always gives me headaches.

I haven't had much occasion to write ultra-performance-sensitive code, but my first impulse would be to exhaust all single-threaded optimizations before resorting to multi-threading. There is almost always some performance that's been left on the table, and it's more than likely easier to get than any multi-threaded benefits, which opens up a huge can of worms.

Another thing new in Haswell is transactional memory. It will be interesting to see what that's like.

I remember reading a paper while in school that found that, at least for databases, the answer to parallel scaling clearly had to be shared-nothing. Sharing memory, or anything else, led to too much overhead. This is probably just as true for multi-core; there will be a limit to how many cores you can use effectively when they're all sharing resources.

[Piotr Tarsa]

Haswell has came. Still doesn't seem like a strong upgrade from Sandy Bridge, but at least the performance increase between Haswell and Ivy Bridge seems higher than the performance increase between Ivy Bridge and Sandy Bridge.

On mobile side there's Crystalwell included in some very expensive CPU models and it improves the performance of Intel's integrated graphics quite a bit.

[nburns]

Haswell has came. Still doesn't seem like a strong upgrade from Sandy Bridge, but at least the performance increase between Haswell and Ivy Bridge seems higher than the performance increase between Ivy Bridge and Sandy Bridge.

Anandtech did a 7-zip benchmark, and the improvement over Ivy Bridge is actually much smaller than Ivy Bridge over Sandy Bridge:

http://www.anandtech.com/show/7003/t...4560k-tested/6

[Piotr Tarsa]

From what I saw there could be some performance penalty of having a L3 cache on separate clock domain. It adds a 2 clocks or so to access time to L3 cache and to main memory. On the other hand, in Haswell, there can be more parallel loads and stores but that's only good when there is parallelism that can be extracted by the processor. It seems that many benchmarks are limited by L3 and memory controller performance, so Haswell doesn't gain much there, but there are many cases where improvements are relatively nice (in comparison to power consumption increase).

[Fallon]

Too bad desktops and laptops are on the way out. All the computer makers know it. But I just can't see writing code on a touch screen.

Maybe not in the way you think. It's true that Mobile handhelds are everywhere. And we see a microsoft surface now. Their present shot at the future is a small Tablet with Laptop functions. It can lay flat, for touch and games, and it can act like a laptop. According to Microsoft it's supposed to be for 'people who are underway, use clouds and want to deliver a modern attractive experience to their customers'. For Laptop mode there is a touch pen.

This will hardly convince many professionals out there. Especially when these work in office space under rooftops, where they have to meet dead lines on PC's with mouses and big quality screens, convenient to look at for hours at a time and suitable for team work.
The present 'ms surface' is somewhat innovative, but not good enough to replace desktops. Windows 8 is more innovative and also not good enough.

If this will be it, I think you will be good with a desktop for years to come. Samsung has come up with a 27 inch full-HD touch-screen monitor optimized for Windows 8, which they would not have done when desktops would be 'on the way out'.

A Desktop PC, midi tower or Barebone, has some characteristics that can hardly be beaten, like the flexible build that allows the customer to exchange and update parts of it, while not loosing the whole device. Especially the separate monitor screen, is a big advantage. It's detached and can be any size. On a notebook/laptop it should be detachable, which is not the case yet, as far as I know. And don't forget that mobile devices sooner or later need peripherals. They are not the answer to all and everything.

So, what will a PC have to look like, in the future?

I wouldn't mind a keyboard that can withstand a cup of coffee thrown over it. Or a screen that can be exchanged quickly when it it gets damaged. Screen size is a common thread of interest in electronic devices we use.
What screen and resolution do we want to watch, on the road, at home, and in the office? It's a mix!

On the road, smartphone and tablets get all the emphasis now. Maybe an ultimate PC would be a mobile handheld, that can logon to any screen that will allow it. Any option to change a screen for a bigger one, would be interesting.
And Oleds can bend. Bendable Oled displays are already coming to future Samsung smartphones and tablets. Also Microsoft has presented a bendable prototype. Software, Windows 9, should -in my view- handle exchange options as well, including the one to skip 'touch' altogether, with an updated mouse, not only a touch pen. (Btw. Will you replace a battery on the Iphone 5 some day? Good luck with that.)

At home, we know we like big screens as well, 27 inch for PC, upwards to whatever for Television. To watch a 3D movie, TV has a bigger comfort zone than a smartphone. Maybe the Desktop PC will get another name when fusion with TV will come full circle. No doubt, big screens will stay.

At work, again, the whole screen size palette is present. An architect may walk around on a project with a tablet. For a presentation, he or she may want a big screen. Let's imagine a stock market with brokers shouting at each other, like in the old days, and now looking through their 'google glasses'. (The future of mobile? A world of twitching fingers of computerized people?). Even if it happens, they will want their big screens just the same, whenever there is a discussion in a crowd.

I was at my PC store last week. A guy turned in his notebook, because of a scratch on the screen. A second time repair. He had been without his computer for months already. I was there to buy a new keyboard, the one with lasered keys. No more worn off keys for me. A new keyboard. Easy to exchange. But only with a Desktop PC or a barebone, a device with exchangable parts.

Bottom line. For now, a desktop PC, big tower, midi tower or Barebone, is convenient and flexible in any place where a mobile device is not necessary or unwanted.

[encode]

Despite the fact that I'm a huge fan of PC upgrading, I'll ignore that HAssWell! It's a Dollar Vote and I'm tired of that Intel keeps releasing same things again and again! It's enough! According to many user tests, IvyBridge is still the King - since AssWell is a bad overclocker and has same heat issues...

[Black_Fox]

They still use the thermal paste under the heatspreader instead of soldering it -> 2 different layers instead of 1 -> harder to conduct heat away from the silicon.

[Piotr Tarsa]

Despite the fact that I'm a huge fan of PC upgrading, I'll ignore that HAssWell! It's a Dollar Vote and I'm tired of that Intel keeps releasing same things again and again! It's enough! According to many user tests, IvyBridge is still the King - since AssWell is a bad overclocker and has same heat issues...

You can try removing the IHS and applying eg Coollaboratory Liquid Pro :] That voids warranty, but in the result you should be able to gain a few hundred MHz.

From what I saw Haswell generally is better than IvyBridge in "real" tests. In synthetic tests it can lose because of power optimizations in the architecture (eg decoupled L3), but in typical usage those deficiences should be overcome by increased execution engine size. OTOH, added logic consumes additional power at full load, so maybe that would mean even smaller (if any) performance improvements when overclocking.

[encode]

It might be better at stock speed... But user feedback on overclocked 4770K makes me cry... Probably it's about first "not so good" batches, probably not. Anyway, I will not pay for a tiny boost, if any. Or else Intel will continue releasing nearly the same CPUs with different labels - it's a Dollar Vote like I said...

[Piotr Tarsa]

IvyBridge-E was just released. From what I read on reviews, it overclocks slightly worse than SandyBridge-E. Performance gains clock-for-clock are (generally consistently) minimal and (surprisingly) power savings are somewhat disappointing. IvyBridge-E has soldered IHS, just like SandyBridge-E, so packaging is not inferior. Seems that power density is really a problem - IvyBridge-E is somewhat close to 2x smaller than SandyBridge-E when it comes to die size (SandyBridge-E was harvested/ partially disabled 8-core while IvyBridge-E is native 6-core, but that doesn't change much I think). Smaller manufacturing nodes will allow for more cores, but we need to lower the clocks to stay within reasonable heat flux.

[Bulat Ziganshin]

i just paid for 4770 box. it should be shipped in a few days

[moisesmcardona]

I upgraded to a Haswell laptop back in August and at least for Video Editing there is a difference. My old system was an Ivy Bridge i7-3600QM. My new system is i7-4700MQ and according to Cyberlink, the use of AVX2 means encoding is 3 times fasters. I have proved this and there seems to be a difference between the two systems enconding a video. Also editing is much faster too. The system is even more responsive than the Ivy Bridge and I cloned the hard disk of the old system to the new one. So far I'm liking Haswell more than the Ivy Bridge.

[encode]

i just paid for 4770 box. it should be shipped in a few days

Not a "K" edition?

[Bulat Ziganshin]

i want to play with TSX

[Sebastian W]

Do you think, the R-version of the Haswell might be interesting for data compression? 4570R 4670R 4770R

The 3rd level cache of all R-processors is 2 MB smaller (only 4MB on i5 and 6 MB on i7). Instead there is a 128 MB 4th level cache on the CPU-board, shared between CPU and GPU, if the on-board-GPU is enabled. I think I have read, that the transfer-rate is a few times higher than main-RAM and the access times are much faster.

What do you think? Is a fast 128 MB (or less if onboard-GPU is enabled) cache good for the memory-access-statistic of compressors?

[Piotr Tarsa]

For compressors that actively use at most 128 MiB of RAM during a non-instantaneous period then it should give a healthy boost if the compressor does a lot of random memory access (well, that statement is trivial, but I felt I have to write that ).

In case of compressors that use much more memory than 128 MiB and in a random access fashion then it would depend if the L4 is really organized like normal cache (ie multiple way set associative; or something). If that's something dumb that eg simply discards old cache entries always (instead of basing the decision on frequency of usage), then it would give less performance boost.


I haven't seen any big boost due to L4 in benchmarks (except benchmarks that stresses IGP). But no benchmarker run PAQ8 :]



I don't see L4 gaining popularity soon, so playing with it wouldn't benefit a lot of people. OTOH it could cut development time because of reduced testing time. I wouldn't expect much, though.

[encode]

The upcoming Intel i7-4790K with it's 4.0 GHz base clock and 4.4 GHz turbo looks promising! I guess 5+ GHz is a modest OC for it!
In addition, Intel claims an improved TIM. If so, it's my next killer ITX build (in pair with Z97 Impact ITX from ASUS, since the currently announced ASUS Z97I-PLUS looks like the lower-end MoBo, aimed to pitch sales for not-yet-announced Maximus VII Impact. Uh, ASUS)