Haswell model specs leaked

[shadow_k]

How can you tell whether Broadwell will be promising or not by just looking at that table?

Look at the last Broadwell chip <10 TDP. This should hopefully mean we should see more tablets with Broadwell inside and also Broadwell is going to have a newly designed arch in the graphics

A Intel worker just did a IAMA on Reddit
http://www.reddit.com/r/IAmA/comments/15iaet/iama_cpu_architect_and_designer_at_intel_ama/

 

[RussianSensation]

Unless I'm mistaken, the performance numbers you are citing and linking are based on multithreaded applications...which is not what the Intel graph is about (single-threaded IPC).

Am I missing something?

I understand what you are saying but do you use CPUs in a theoretical sense or in a practical sense? What games are single-threaded? What programs are single threaded now? ITunes? Just because you are testing a dual-threaded game like Starcraft 2 on a quad-core CPU, it doesn't mean you are not stressing the IPC of the CPU in that particular app. I can't play single-threaded Starcraft 2 game since there is no such game.

You are missing IPC increases in real world apps, not in some chart Intel made without telling us what programs, how many of such programs and from what year the apps they used to arrive at their results. I am sure you can feel comfortable that IPC increases from Conroe to Nehalem at least 10% because Intel has made such a claim. However, if in 20-30 modern applications the average is more like 17-20%, as users of such processors we see 17-20%, not 10%.

I linked for you IPC increases based on real world applications people may use who buy $200-300 CPUs. I have no idea how Intel derived those numbers but since I've been following IPC increases from real world reviews from Conroe to Nehalem to SB, the #s you posted don't match any review I've ever read on this subject. The example you cited only has a 10% IPC increase from Conroe to Nehalem and just 7% from Nehalem to SB. This is grossly inaccurate from the performance users see right now if they were to upgrade from a Core 2 Quad to a Core i5 2500K and tested the performance at the same clocks. Are you suggesting that going from Core Q6600 @ 3.2ghz to Core i5 2500K @ 3.2ghz only improves overall performance by 1.10 (C2Q -> Nehalem) *1.07 (Nehalem -> SB) = 17.7%? Q6600's IPC is only on par with Phenom II X4 955-965.

Now look at what happens in CPU limited games between i5 2500K and X4 955 at similar clocks:

Perhaps Intel may have used outdated programs/benchmarks when it arrived at those numbers. If you want a true reflection of IPC how a user experiences it, you should benchmark modern apps that have modern code which runs faster on new generation of CPUs due to their inherent architectural enhancement that are better catered for executing modern code.

For example, how did Intel arrive at IPC increase from Conroe to Nehalem? Did they use 2011-2012 programs or 2008 ones when Nehalem just came out? If they tested games, did they use a 2008 GPU with Nehalem? If so, maybe Nehalem was GPU bottlenecked. Furthermore, why would you ignore IPC increase in multi-threaded apps when discussing overall increase in IPC from one generation to the next? With a relatively modern GPU, you can exploit Nehalem's CPU to the fullest and reveal serious bottlenecks in older architectures such as lack of sufficient or slow cache. If you run a 4-threaded app on a quad core CPU, are you not testing IPC? If there are inefficiencies in shared L2/L3 cache of the CPU architecture and its ability to fully feed all the 4 cores in a 4-threaded app, it will show up as a performance hit.

Trying to graph Intel's IPC increases in limited theoretical single threaded apps is not a true reflection of IPC increases users actually see and feel in modern apps. For that real world benchmarks should be used. Based on real world usage, IPC from Core 2 Quad to Nehalem seems to have increased by 20%, Nehalem to SB is an increase of 14-16%, and SB to IVB is 3-4% --- all at the same clock speed.

Maybe I don't understand the definition of IPC from a technical perspective you are describing and confusing Performance Per Core with IPC? Sorry if I did . The way I look at it, if I get a Haswell CPU, how much faster will it be on average in modern apps @ 4.5ghz compared to an i5 2500K @ 4.5ghz? If it's only 10%, it would end up being the lowest real world increase in IPC/per core performance compared to C2Q --> Nehalem --> SB.

Granted by the same account, Haswell's IPC increase could grow beyond 10% if more programs start using AVX2 and benefit from doubling of the floating point operations. Perhaps I am jumping the gun then and not giving Haswell the chance to shine in more modern apps that will take advantage of its architecture. Maybe Intel just has a tendency to underestimate their real world IPC increases? I guess I'll wait for real world tests before burring Haswell's 10% claim as disappointing as of now. Fair enough.

 

[cbn]

http://www.techpowerup.com/177817/Intel-Haswell-and-Broadwell-Silicon-Variants-Detailed.html

Seems like desktop computers won't be getting GT3 :|

Broadwell looks very promising.

Well, it looks like Broadwell on LGA1150 could exist, just not in a dual core format.

 

[jpiniero]

Well, it looks like Broadwell on LGA1150 could exist, just not in a dual core format.

If they are going to go ahead with Broadwell instead of the faster Haswell for LGA, it wouldn't surprise me if it was only one model - 95W 4 Ghz no turbo.

 

[fixbsod]

no 6 core broadwell?

 

[ShintaiDK]

no 6 core broadwell?

Its a different platform. on LGA11xx, no. Never intended.

 

[RU482]

http://www.techpowerup.com/177817/Intel-Haswell-and-Broadwell-Silicon-Variants-Detailed.html

Seems like desktop computers won't be getting GT3 :|

Broadwell looks very promising.

Is there any interest in a GT3 ULT or ULX board for desktop?

Not being skeptical, but rather trying to find out if there is any interest..

 

[ShintaiDK]

Is there any interest in a GT3 ULT or ULX board for desktop?

Not being skeptical, but rather trying to find out if there is any interest..

NUC, Mac Mini etc?

 

[cbn]

Is there any interest in a GT3 ULT or ULX board for desktop?

Not being skeptical, but rather trying to find out if there is any interest..

Can you add the following request to your "ULT/ULX for Desktop" list:

A way to get true desktop speeds from ULT or ULX when the mobile device they are installed in is connected to AC power. (eg, remove battery and replace with supplemental cooling for a clock speed boost).

To me that would be more interesting than a specific ULT/ULX desktop board per se. Although with that said, I have to admit it would be interesting to see what kind of potential could be unlocked from the GT3 iGPU if the much faster desktop DDR3 DIMMs (on a true desktop board) were used in place of the slower SO-DIMMs normally found in laptops.

 

[IntelUser2000]

Look at the last Broadwell chip <10 TDP. This should hopefully mean we should see more tablets with Broadwell inside and also Broadwell is going to have a newly designed arch in the graphics

I'm not hoping that much for Broadwell. Just like Nehalem, Sandy Bridge, and Haswell, its the Tock that's exciting.

 

[Fjodor2001]

I'm not hoping that much for Broadwell. Just like Nehalem, Sandy Bridge, and Haswell, its the Tock that's exciting.

The Tick Tock model does not seem to be so strict anymore. For example IB was not only a node shrink, but brought a new iGPU arch and IPC improvements too.

 

[Idontcare]

Maybe I don't understand the definition of IPC from a technical perspective you are describing and confusing Performance Per Core with IPC? Sorry if I did . The way I look at it, if I get a Haswell CPU, how much faster will it be on average in modern apps @ 4.5ghz compared to an i5 2500K @ 4.5ghz? If it's only 10%, it would end up being the lowest real world increase in IPC/per core performance compared to C2Q --> Nehalem --> SB.

RS, that is the source of the confusion. You are comparing products, SKUs if you will, full-fledged multi-core chips which is not the same thing and will not tell you the IPC improvements.

IPC is intended to represent single-core single-thread performance, and it is obviously heavily dependent on the instruction mix of the specific application in question.

There are reasons people care to quantity IPC, same reasons people care to normalize clockspeeds for comparisons and so on. But to be sure IPC doesn't tell you how a final product is going to perform, but if you know IPC, and you know target clockspeeds, and you know thread-scaling performance, then you can compute the final product performance which is what you, as the end-user, are going to experience or read about in a review.

I think you are mixing the two and that is what is leading to the confusion.

If I had asked you how much better Nehalem was to Conroe the first question you'd have to ask is "which nehalem? and which conroe?".

But if I asked you for the IPC of each then that eliminates core count, cache arrangement (mostly, but not entirely), and clockspeed as variables.

At that point there is only one questions "IPC for which set of instructions?"

IPC for superPI is going to be different than the IPC for Cinebench R11.5 because those two apps use very different instructions in the code.

That is where we rely on Intel to know what they are doing, to have no reason to deceive or manipulate our expectations, when they state "using a broad range of consumer applications".

Once you know the IPC delta for say Haswell, now you need to factor in core count differences, if any, and clockspeed differences, if any, turbo-boost differences, if any, and you have your single-threaded performance gain, if any.

Why is single-threaded performance relevant? Ask anyone who has heard of bulldozer and they'll tell you just how much it matters.

Multithreaded performance may well be increasing all the more, but that is a product of more than just IPC improvements. The final performance improvement in multithreaded apps will be dependent on clockspeeds and core counts.

 

[ShintaiDK]

The Tick Tock model does not seem to be so strict anymore. For example IB was not only a node shrink, but brought a new iGPU arch and IPC improvements too.

Penryn also gave IPC improvements over Conroe. Westmere over Nehalem.

The only part that differs is the GPU.

 

[IntelUser2000]

The Tick Tock model does not seem to be so strict anymore. For example IB was not only a node shrink, but brought a new iGPU arch and IPC improvements too.

Yes, but a far less advancement than Tock chips like Nehalem and Sandy Bridge.

Even the feature that is quoted as "Tick +", which is graphics, only brings 40-60% gain when Sandy Bridge brought 1.5-2.5x gain over Westmere along with the CPU IPC gain and power efficiency improvements.

Penryn = 5-7% IPC gain
Westmere = Additional instructions(like AES, but none specifically for general purpose performance), and zero IPC gain
Ivy Bridge = 3-5% IPC gain

 

[IntelUser2000]

IPC is intended to represent single-core single-thread performance, and it is obviously heavily dependent on the instruction mix of the specific application in question.

Exactly.

I wanted to say something in line with what IDC said but I am too late.

I'd like to add that Haswell's "10%" is for everything, while Nehalem's "20%" was a mix of single and multi-threaded applications.

People were actually disappointed with Nehalem at first because the single threaded performance increase was only 5-10%. It was made up with higher clocks using Turbo, and 35-40% you got in multi-threaded applications with Hyperthreading.

If you had an application that a) Wasn't sensitive to memory bandwidth b) Didn't take advantage of extra threads

Then you'd end up with 5-10% gain. Some reported that in some code the gains were more like 0-5%.

Instructions Per Clock is also much harder to achieve. AMD for example could also probably get easy 30%+ in multi-threading if they have SMT like Hyperthreading too. But getting it up in everything is much harder to do.

 

[IntelUser2000]

Fudzilla claims that Haswell GT2 will be 3x faster in 3D graphics than Ivy Bridge GT1.

That sounds similar to what Intel claimed with Ivy Bridge. They said back then Ivy Bridge GT2 will be 3x faster than Sandy Bridge GT1.

But because in the Ivy Bridge generation they were comparing it to Core i7 2600's GT1(which clocks at 1350MHz), rather than the regular GT1(which is at 1150MHz), it means Haswell's gain will be that much less.

Conclusion: I expect Haswell GT2 to end up being 20-30% faster than Ivy Bridge GT2. Perhaps upper bound of 30% is achievable in average, but expect no more than that.

 

[Fjodor2001]

Yes, but a far less advancement than Tock chips like Nehalem and Sandy Bridge.

Even the feature that is quoted as "Tick +", which is graphics, only brings 40-60% gain when Sandy Bridge brought 1.5-2.5x gain over Westmere along with the CPU IPC gain and power efficiency improvements.

Penryn = 5-7% IPC gain
Westmere = Additional instructions(like AES, but none specifically for general purpose performance), and zero IPC gain
Ivy Bridge = 3-5% IPC gain

Sure. But if you also bring the chipset into the equation, with IB came the Z77 chipset that brought native USB3, DDR3L support, PCIE3, 1600 MHz DDR3 support etc. And not to mention the lower TDP that comes with 22 nm.

So it's not like Tocks are all good and Ticks are pointless anymore. Somehow I feel that Intel is going in a direction where both Ticks and Tocks should bring similar amounts of added value to the end user.

Then it's of course individual what each end user value the most; improvements in CPU arch, iGPU arch, I/O, lower power consumption, or something else. For example I guess a lot of people already feel that IB is fast enough, so the CPU arch improvements that come with Haswell will maybe not be of much value to them (especially since the performance difference compared to IB is not that high). But others may think differently.

 

[Idontcare]

So it's not like Tocks are all good and Ticks are pointless anymore. Somehow I feel that Intel is going in a direction where both Ticks and Tocks should bring similar amounts of added value to the end user.

Today's modern microarchitectures are so mind-bogglingly complex that anything less than a tick-tock "baby step" type micro-evolutionary approach to enhancing them would assuredly end in disaster.

Not many people can even imagine the sort of complexity that goes into attempting to effectively manage and direct a team of 5,000 or so engineers working towards a single goal. It drives managerial evolution in its own right, let alone the sorts of tool evolution that must occur for the engineers to actually get work done that isn't a waste of time or resources.

So many left hands that must talk with so many other right hands, and in the end it just takes one or two missteps and you get a crippling TLB bug, or FDIV bug, or a 6 month delay, or a $1B over-budget program.

Some folks know to marvel at the miracle of modern science that is a 32nm or 22nm process node, or the billion xtor IC that is produced on those miniature nodes, but few folks know that the organizational level and methods that are necessary to create those chips and nodes in the first place also had to be created beforehand.

The tools to effectively harness and manage the combined output of thousands of people while minimizing collisions and conflicts (and organic processor of sorts, complete with layout/design/validation in its own right) had to be developed before a functional product could ever be the resultant output.

It is this behind-the-scenes evolution in the ways humans manage to work together that the tick-tock model serves. Otherwise you will figuratively have design teams and program managers biting off more than they can chew, setting goals that are unattainable because the organizational mechanisms necessary for success have not been invented in the meantime.

 

[Fjodor2001]

Today's modern microarchitectures are so mind-bogglingly complex that anything less than a tick-tock "baby step" type micro-evolutionary approach to enhancing them would assuredly end in disaster.

Well, partly true. But you work by breaking down the huge complex task into smaller problems which by themselves are possible to grasp and handle. Then you define "interfaces" between each area and connect it all together. Basic engineering mindset.

Same with software. For example the 3.2 release of the Linux Kernel had 14,998,651 lines of code. At first sight that may seem hugely complex and impossible to maintain and develop. But it's broken down into separate logical areas that are completely manageable (even though it still can be complex at times of course).

 

[Idontcare]

Well, partly true. But you work by breaking down the huge complex task into smaller problems which by themselves are possible to grasp and handle. Then you define "interfaces" between each area and connect it all together. Basic engineering mindset.

Same with software. For example the 3.2 release of the Linux Kernel had 14,998,651 lines of code. At first sight that may seem hugely complex and impossible to maintain and develop. But it's broken down into separate logical areas that are completely manageable (even though it still can be complex at times of course).

If it wasn't clear from my post, obviously the management of these projects is possible as evidenced by the existence of the products themselves.

That, in and of itself, was not at all what I was attempting to highlight with my post. But...I will accede that my failing is of my own doing, wall of text and all.

 

[IntelUser2000]

Then it's of course individual what each end user value the most; improvements in CPU arch, iGPU arch, I/O, lower power consumption, or something else. For example I guess a lot of people already feel that IB is fast enough, so the CPU arch improvements that come with Haswell will maybe not be of much value to them (especially since the performance difference compared to IB is not that high). But others may think differently.

Which is everything Haswell does better. Radical improvement in power management, greater improvement in graphics, better integration, noticeably better CPU.

 

[Fjodor2001]

Which is everything Haswell does better. Radical improvement in power management,

Yes, but not low lower TDP (for the desktop CPUs).

greater improvement in graphics,

HD4000->GT2 is expected to be less improvement than HD3000->HD4000. GT3 is another story though, but it will only be included in a few selected mobile Haswell models.

better integration,

Please clarify.

noticeably better CPU.

Depends on what you mean by noticeably. In some specific use cases perhaps. But 10-15% IPC improvement is not that spectacular in my world.

Finally, note that I'm not saying Haswell is a crappy CPU. But I'm saying that SB->IB brought about the same amount of increased end user value (especially if the new Z77 chipset is included in the evaluation), depending on what kind of improvements you value most that is.

 

[Homeles]

Can someone clairfy something for me? Is desktop Haswell getting the on-package VRM? Or is that only GT3 Haswell?

 

[ShintaiDK]

Can someone clairfy something for me? Is desktop Haswell getting the on-package VRM? Or is that only GT3 Haswell?

All SKUs gets the VRM.

 

[Fjodor2001]

Can someone clairfy something for me? Is desktop Haswell getting the on-package VRM? Or is that only GT3 Haswell?

Do you mean on-package VRM or RAM? I think the latter has been associated with the GT3, but it's still not confirmed that there will be on-die RAM.