Intel "Haswell" Speculation thread

[ShintaiDK]

Well, I'm sending this P4(478)Prescott out to pasture when Haswell comes out. Unless they decide to keep the paste under the IHS like Ivy. Then I may decide to vote with my dollar and support the underdog. Say what you want about Prescotts... at least they have a soldered lid. Lasted a good long time of faithful service over the years of daily use.

So you base your buy on whatever paste is used, seriously?

 

[BenchPress]

not directly related to this post but to a past discussion of ours about the MIC-AVX convergence, one more clue it's coming: http://software.intel.com/en-us/forums/showpost.php?p=190704

"Reconciliation of MIC instruction sets with future AVX is a publicized goal"

Awesome. Thanks for the heads up.

 

[Ajay]

So you base your buy on whatever paste is used, seriously?

Yeah, while somewhat stupefying, I wouldn't not get Haswell because it used paste instead of 'solder'. The base IPC improvement should be worth it, and once other architectural enhancements (AVX2 especially) are utilized, it'll just blow SB and IB away.

I might pick up an 8 core Piledriver CPU when they come out, but that's to replace a dual core Athlon II so that I can use it for folding (hopefully, Piledriver will fold much better than Bulldozer).

 

[Homeles]

Seriously, if you have that much of a problem with the IHS not being soldered, remove it.

 

[Foamy]

Hey guys, got a question for you if you don't mind, i'm building a new pc or at least trying to and i'd like to know if the difference from an ivy to a haswell processor would be that great.

I'm coming from a e6300 @ 1,83 and the upgrade is long overdue but i wouldn't like to buy something right now that would be absolutely blown out of the water as soon as Haswell ships

I've looked through this thread but i haven't come to decisive conclusion since most of it is full of tech words that mostly have no understanding for me. ( sorry for being dumb)

Thanks anyway, if you could help me with the build as well i'd be grateful. Just point me to the right topic

 

[KingRaptor]

Hey guys, got a question for you if you don't mind, i'm building a new pc or at least trying to and i'd like to know if the difference from an ivy to a haswell processor would be that great.

I'm coming from a e6300 @ 1,83 and the upgrade is long overdue but i wouldn't like to buy something right now that would be absolutely blown out of the water as soon as Haswell ships

I've looked through this thread but i haven't come to decisive conclusion since most of it is full of tech words that mostly have no understanding for me. ( sorry for being dumb)

Thanks anyway, if you could help me with the build as well i'd be grateful. Just point me to the right topic

Intel Haswell processors should use a new architecture compared to Sandy Bridge and Ivy Bridge. If everything goes according to plan, the new architecture should be more efficient and handle new instructions better. If you really need a system now, Ivy Bridge is the way to go; otherwise, Haswell should be worth the wait.

 

[BenchPress]

I'm coming from a e6300 @ 1,83 and the upgrade is long overdue but i wouldn't like to buy something right now that would be absolutely blown out of the water as soon as Haswell ships

Don't just think about the performance gain on the day you buy it. Think about how long it will last too...

Ivy Bridge fulfills the needs of today's applications, but as the workloads increase it will start feeling slower. Haswell offers little or no advantage for today's applications (over Ivy Bridge), but offers many new features which will allow future applications to run at much higher performance than on Ivy Bridge. So it will last many years longer.

 

[Foamy]

Don't just think about the performance gain on the day you buy it. Think about how long it will last too...

Ivy Bridge fulfills the needs of today's applications, but as the workloads increase it will start feeling slower. Haswell offers little or no advantage for today's applications (over Ivy Bridge), but offers many new features which will allow future applications to run at much higher performance than on Ivy Bridge. So it will last many years longer.

Well i'm looking for a ~3 year investment here, would an 3570k hold me till the tick of the Haswell? Or would you guys think it would become obsolete way before that time?

 

[thm1223]

Don't just think about the performance gain on the day you buy it. Think about how long it will last too...

Ivy Bridge fulfills the needs of today's applications, but as the workloads increase it will start feeling slower. Haswell offers little or no advantage for today's applications (over Ivy Bridge), but offers many new features which will allow future applications to run at much higher performance than on Ivy Bridge. So it will last many years longer.

Would you mind elaborating? Sorry if what your referring to has already been stated.

Edit: And Shintai what's your take on this? I garner that your general view of Haswell is lukewarm in terms of its expected performance leap. However, a modest increase albeit a much more future proof chip relative to previous Intel releases would actually be pretty awesome.

 

[boxleitnerb]

What kind of IPC improvement do you expect? I hope for another 15% with 15% higher clocks.

 

[ShintaiDK]

Well i'm looking for a ~3 year investment here, would an 3570k hold me till the tick of the Haswell? Or would you guys think it would become obsolete way before that time?

Ivy would last you to Skylake. Unless you really feel the urge to upgrade.

 

[ShintaiDK]

Would you mind elaborating? Sorry if what your referring to has already been stated.

Edit: And Shintai what's your take on this? I garner that your general view of Haswell is lukewarm in terms of its expected performance leap. However, a modest increase albeit a much more future proof chip relative to previous Intel releases would actually be pretty awesome.

I´m not lukewarm of Haswell. I just see it more realistic and not turning it into Phenom/Bulldozer type of hype.

Posting AVX2 and TSX everywhere all the time doesnt make it better. Its not gonna be more future proof due to that. Even AVX2 would only benefit games at maybe 5% in average vs AVX. And AVX is barely used today outside of more specialized apps.

Haswell will offer up to 20% as far as I am aware now. But consider that IB already gave you 5% IPC+3% from clock if you compare 2500K with 3570K in average.

Some of Haswells main features are VRM integration and iGPU improvements. As I see it the CPU part keeps the relatively evolutional approach. Still very good, but its not some kind of revolution you need to be on the other side for.

People buying K8s in 2002 to play 64bit games didnt exactly win on that account either.

 

[tweakboy]

That's awfully disappointing if true, but I wouldn't doubt it. At least my Gulftown will be modern for a very long time...

Look at me with my 2007 CPU. Trust me that Sandy 2600k will go for years and years to come. Unless all you do is video editing and render. After work my dad messes with Adobe Premiere, , hes improved, make videos and effects. so we got him a 2600k 128GB SSD as drive for everything.... ,,,

Intel hasnt hyped up Haswell and its supposed to come out March 2013 ... from what I have seen. Im sure they improved the CPU side as well as the GPU side. But GPU inside good or not, we use our video cards ()

 

[BenchPress]

Well i'm looking for a ~3 year investment here, would an 3570k hold me till the tick of the Haswell? Or would you guys think it would become obsolete way before that time?

Nothing ever becomes "obsolete" in such a short time span. But it largely depends on what you use your system for to determine which is the best value. If you do lots of video editing or content creation then you'll see Haswell pulling ahead of Ivy Bridge quite considerably and well within that time span (those applications are preparing for Haswell's features right now). But considering that you've had an E6300 for many years you probably don't run many cutting-edge applications that always hunger for higher performance.

 

[BenchPress]

What kind of IPC improvement do you expect?

Very little. They already have a very strong architecture so there's not much headroom. Any drastic changes would compromise other aspects. Power efficiency is critically important for the emerging markets, and increasing IPC can easily cost power efficiency when being too aggressive. IPC gains from the hypothetical improvement on macro-op fusion, which would be favorable because it has the potential to decrease power consumption, would only be in the lower single digits.

There might be a third AGU to support two loads and one store each cycle, but that would be a tough compromise power-wise so I wouldn't bet on it. Any IPC improvement from it would again be in the lower single digits.

I can't think of much else that would increase IPC. And either way they've put a ton of effort into AVX2 and TSX to improve performance through other means. So I don't expect they've made bold changes to increase IPC for legacy applications.

 

[tweakboy]

They will release Haswell 6 core 12 threads // the 500 dollar chip...

 

[BenchPress]

I´m not lukewarm of Haswell. I just see it more realistic and not turning it into Phenom/Bulldozer type of hype.

Except Haswell isn't a hype. It is known to double the peak throughput per core. And unlike Bulldozer there isn't a single reason to believe it would compromise anything to achieve that, since they just have to merge/extend the SIMD execution stacks in the same way they did with Sandy Bridge. So it's impossible to turn this into a Phenom/Bulldozer "type of hype".

Posting AVX2 and TSX everywhere all the time doesnt make it better. Its not gonna be more future proof due to that.

Please do explain. How does doubling the throughput by enabling the SPMD on SIMD programming model (the same technology that makes GPUs so fast), and adding highly praised multi-threading enhancements, not make it more future proof?

Even AVX2 would only benefit games at maybe 5% in average vs AVX. And AVX is barely used today outside of more specialized apps.

What are you basing that number on?

And yeah, AVX1 is barely used. But that's because it's limited to floating-point and lacks the bandwidth to sustain peak throughput. So it's no surprise the vast majority of developers are skipping it. But you can't conclude anything from that for Haswell's fate. It adds integer support, doubles the bandwidth, adds gather support, adds FMA support, adds missing vector equivalents of scalar instructions, and adds permutation instructions. That's everything developers ever dreamed of in a vector instruction set, and it won't even require assembly programming!

Haswell will offer up to 20% as far as I am aware now.

Do you have anything at all to back that up?

People buying K8s in 2002 to play 64bit games didnt exactly win on that account either.

K8 was launched in 2003.

Anyway, you're again making a completely irrelevant comparison. 64-bit GPRs are merely useful to extend the address space, with no significant effect on game performance. Anyone claiming it would was completely ignorant. So don't try to make any parallels with Haswell, because there are none.

Extending the SIMD execution units to 128-bit back in 2006 on the other hand had a huge impact on performance per clock for games. And yes, clock frequency took a hit, but note that Haswell will have 256-bit SIMD support (with gather and FMA to boot) and there won't be any drop in clock frequency!

 

[Tuna-Fish]

Even AVX2 would only benefit games at maybe 5% in average vs AVX. And AVX is barely used today outside of more specialized apps.

No. The main draw of AVX2 is that it allows the use of vectorized instructions on a lot of loads where you cannot use them today, specifically because it will make automatic vectorization by the compiler be much more efficient. Because of this "5% average" is *way* lowballing it -- I'd fully expect 2x average gain over a large set of different workloads. For recompilation only.

Of course, it will be years before people start to actually target AVX2. But gather absolutely is a big deal.

 

[ShintaiDK]

They will release Haswell 6 core 12 threads // the 500 dollar chip...

Not for LGA1150.

 

[BenchPress]

Of course, it will be years before people start to actually target AVX2.

"Target" as in using it as the minimum system requirement, yes, but many applications will have AVX2 optimized paths long before then. All major compilers already have support for AVX2 intrinsics, and auto-vectorization is on the way. Intel will also no doubt have its OpenCL framework support AVX2 on the day of Haswell's launch. And I'd be very surprised if they weren't handing out engineering samples to key software companies who create multimedia applications (especially those commonly used in benchmarks).

Because developers can get twice the throughput with little effort, the adoption of AVX2 should be much faster than any previous vector instruction set extension.

 

[Fjodor2001]

Because of this "5% average" is *way* lowballing it -- I'd fully expect 2x average gain over a large set of different workloads.

If AVX2 really is that good, how come it hasn't been added years ago? Is it hard to implement in HW, or what has been holding it back?

 

[Tuna-Fish]

If AVX2 really is that good, how come it hasn't been added years ago? Is it hard to implement in HW, or what has been holding it back?

Gather is notoriously hard/expensive to implement with cache coherency and a strong memory ordering system like the one in x86.

It (and it's twin, scatter) has been in widespread use in a lot of systems, from supercomputers to modern GPUs, but afaik x86 will be the first arch with strong memory ordering and coherency guarantees that will get gather.

 

[BenchPress]

If AVX2 really is that good, how come it hasn't been added years ago? Is it hard to implement in HW, or what has been holding it back?

First of all, it's not like there wasn't anything before AVX2:

1996 - Pentium MMX: 64-bit integer (INT) vector instructions, 64-bit INT execution units
1999 - Pentium 3: 128-bit floating-point (FP) vector instructions, 64-bit execution units
2001 - Pentium 4: 128-bit FP + INT vector instructions, still 64-bit execution units
2006 - Core 2: 128-bit FP + INT vector instructions, 128-bit execution units, but lower frequency
2011 - Sandy Bridge: 256-bit FP vector instructions, 256-bit FP execution units
2013 - Haswell: 256-bit FP + INT vector instructions, 256-bit FP + INT execution units, gather support, FMA support

You have to think of doubling the vector width as almost an alternative to doubling the number of cores. If Haswell didn't have AVX2 but came with 8 cores by default, it would have been a pretty great chip too but nobody would question "why didn't they think of this before". Likewise AVX2 is great from a performance perspective for a certain range of applications, but not all that extraordinary from a technological point of view to question why they haven't done it before. They've been adding vector capabilities to x86 since 1996, and the transistor budget has been the main limiting factor.

So why did they choose AVX2 over doubling the number of cores? It's all about balance. Software contains varying levels of task parallelism (TLP) and data parallelism (DLP). TLP is extracted using more cores/threads, while DLP is best extracted using wider vectors. DLP can actually also be extracted using more cores, but programming for many cores is hard while using wide vector instructions is relatively easy. Last but not least, wider vectors are also more power efficient than more cores.

What does set AVX2 apart from every previous vector instruction set extension is gather support; the ability to read multiple memory locations in parallel. This is a big deal because it enables the compiler to easily vectorize scalar code. So why haven't they added gather before then? Well it only starts making a real difference when the vectors are quite wide, so it wasn't worth the transistors before. Wide vectors and gather kind of belong together, but once you have both they enable a whole new programming model.

 

[ShintaiDK]

Just a small revisit to an old part.

One input to rule them all.

 

[Sweepr]

Fudzilla talks about 10+% IPC increase over Ivy Bridge:

http://www.fudzilla.com/home/item/28318-haswell-to-be-10 -percent-faster-than-ivy-bridge