Broadwell to Skylake, will be similar to the transition from Prescott to Conroe

[Exophase]

I am pretty much certain that Prescott removed dual pumped ALUs. And anyway all stars had to be aligned perfectly for P4 to make use its dual pumped ALUs properly. It was glass jawed paper tiger.

I can't find a good source on it either way but I'll take your word for it. Bug half-cycle latency adds are really not that beneficial, I think because long dependency chains of adds aren't much of a thing in code. Especially considering how much you can reorder adds in a sum...

 

[BigDaveX]

If memory serves, the double-clocked ALUs in Willamette and Northwood were intended to help compensate for the pipeline stalls and get things going again quickly; Prescott (and Cedar Mill) had other ways of achieving the same thing that used less power. And god knows, it needed it.

In fact, looking back on the Prescott launch articles from 2004, Intel did actually incorporate a lot of genuine improvements to the core. Unfortunately, thanks to them nearly doubling the pipeline, said improvements instead ended up having to shore up the chip's performance (though to be fair, they probably weren't expecting their 90nm process to be such a huge misfire).

 

[Haserath]

Prescott still had the half-cycle ALUs.

http://www.anandtech.com/show/1230/7

The Pentium 4 has three Arithmetic and Logic Units (ALUs) that handle integer code (code that operates on integer values - the vast majority of code you run on your PC). Two of these ALUs can crank out operations twice every clock cycle, and thus Intel marketing calls them "double pumped" and says that they operate at twice the CPU's clock speed. These ALUs are used for simple instructions that are easily executed within 1/2 of a clock cycle, this helps the Pentium 4 reach very high clock speeds (the doing less work per cycle principle)

 

[BigDaveX]

Huh, that's odd; I checked Tech-Report's Prescott review and it said the same thing. Yet I distinctly remember at the time there was a lot of talk about how Prescott had gone back to single-speed ALUs because of power concerns and/or Intel thinking it would be redundant because the whole chip would be clocked at the 5-6GHz speeds the ALUs operated at in Northwood. Unless it was actually Tejas that was going to go back to single-speed...

In either case, that means that the ALUs in those Cedar Mills that people got "suicide shot" overclocks on using LN2 must have been operating at around 17GHz! Kind of blows the mind, thinking about it.

 

[JoeRambo]

Agner Fogg has instruction latency tables, and in them, Prescott no longer has dual pumped ALU. Latency goes from 0.5 to 1 just like expected.

 

[Maximilian]

Broadwell to Skylake, will be similar to the transition from Prescott to Conroe

When highly overlocked under the best conditions with a certain type of niche workload that uses some new instruction set

It'll be something like that

 

[Exophase]

Agner Fogg has instruction latency tables, and in them, Prescott no longer has dual pumped ALU. Latency goes from 0.5 to 1 just like expected.

All indicators are that they're still double pumped but removed the forwarding necessary to get 0.5 cycle latency, since they still get the same throughput (and Agner's uarch document says as much)

Netburst's double pumped ALUs were width-pipelined, where operations were performed over 2 stages (plus a third stage to generate flags). To get 0.5 cycle latency the second stage needed to forward to the first stage, which works for add/sub or logic ops because they can be done sequentially. But doesn't work for shifts, which is why Willamette/Northwood couldn't do shifts on either of its double pumped ALUs (and the slower ALU port forwarded to the SIMD unit to do it). But on Prescott, double pumped ALU1 could now do shifts. This was probably a driver for removing half-cycle forwarding, although they could have maybe kept it for ALU0. Adding 64-bit support may have also played a role.

If you had to pick between the two - half cycle simple ALU w/o shifts vs single cycle ALU with shifts, the latter seems like an obvious choice for a lot of code.

 

[mohit9206]

I seriously doubt that. Probably 10-15% performance improvement and another 10-20% more power efficient. Intel currently has no competition so why would they shoot for big improvements?

 

[Yuriman]

I seriously doubt that. Probably 10-15% performance improvement and another 10-20% more power efficient. Intel currently has no competition so why would they shoot for big improvements?

Because people currently see no reason to buy Haswell/Broadwell if they already have Sandy Bridge.

 

[mohit9206]

Because people currently see no reason to buy Haswell/Broadwell if they already have Sandy Bridge.

Yes that's a good reason. I too am on Sandy Bridge so for me it will be either Skylake /Cannonlake or Zen APU from AMD.

 

[NTMBK]

Because people currently see no reason to buy Haswell/Broadwell if they already have Sandy Bridge.

The majority of people would still see no reason, even if Skylake was twice as fast as Broadwell. Who cares if your CPU idles for 99% of the time instead of 98% of the time? Who needs 8 cores at 5GHz to run Office, Chrome and Spotify?

Outside of gamers, there is no killer app which is making people upgrade their desktops any more.

 

[Dave2150]

The majority of people would still see no reason, even if Skylake was twice as fast as Broadwell. Who cares if your CPU idles for 99% of the time instead of 98% of the time? Who needs 8 cores at 5GHz to run Office, Chrome and Spotify?

Outside of gamers, there is no killer app which is making people upgrade their desktops any more.

Obsiously we mean for gamers, not people who'd be best off with tablets....

 

[Headfoot]

You're playing semantics, so this will be my last post to you.

You're the one playing semantics here dude. Broadwell is coming to desktop, period. In BGA form. It might not be coming as Broadwell-K in socketed form, but that's not the same thing as not coming to desktop. 65w chip = desktop.

 

[Headfoot]

The majority of people would still see no reason, even if Skylake was twice as fast as Broadwell. Who cares if your CPU idles for 99% of the time instead of 98% of the time? Who needs 8 cores at 5GHz to run Office, Chrome and Spotify?

Outside of gamers, there is no killer app which is making people upgrade their desktops any more.

Anyone doing anything with heavy duty professional software:
* Video editing
* CAD
* Rendering
* 3d modeling
* Photo editing
* Digital Audio Workstations, audio mixing, etc.
* Legacy apps which use only one thread but are too costly to re-code (many thousands of Java apps are like this)
* Legacy apps coded in the early-mid 2000's when programmers still said things like "I don't need to write efficient code, in 10 years CPUs will be so fast it wont matter how slow my code is"
* Real time medical imaging (there is a push to bring it off of render farms and onto desktops for greater responsiveness)
* Almost all GPU aware real world desktop apps can make use of more single thread CPU for greater throughput for submitting to the GPU(s)
* More that I'm sure I'm leaving out

 

[njdevilsfan87]

I remember Prescott to Conroe was hyped big even a good 6 months before Conroe was released. I'm not seeing any of the same hype for Skylake.

 

[Arachnotronic]

I remember Prescott to Conroe was hyped big even a good 6 months before Conroe was released. I'm not seeing any of the same hype for Skylake.

Intel was behind then, so they needed to generate buzz.

Intel is now out in front, so less hype required.

 

[crashtech]

You're the one playing semantics here dude. Broadwell is coming to desktop, period. In BGA form. It might not be coming as Broadwell-K in socketed form, but that's not the same thing as not coming to desktop. 65w chip = desktop.

I haven't seen anything that indicates the two LGA1150 Broadwell models shown in the VR-Zone article won't be released. So likely BGA and LGA; for SFF, desktop, and maybe even a higher-end AIO.

 

[NTMBK]

Anyone doing anything with heavy duty professional software:
* Video editing
* CAD
* Rendering
* 3d modeling
* Photo editing
* Digital Audio Workstations, audio mixing, etc.
* Legacy apps which use only one thread but are too costly to re-code (many thousands of Java apps are like this)
* Legacy apps coded in the early-mid 2000's when programmers still said things like "I don't need to write efficient code, in 10 years CPUs will be so fast it wont matter how slow my code is"
* Real time medical imaging (there is a push to bring it off of render farms and onto desktops for greater responsiveness)
* Almost all GPU aware real world desktop apps can make use of more single thread CPU for greater throughput for submitting to the GPU(s)
* More that I'm sure I'm leaving out

And for all but 1 or 2 of those things, the answer is "moar cores". Which is why Intel is happy to sell you an 18 core Xeon if you need lots of CPU performance. :thumbsup:

 

[erunion]

You're the one playing semantics here dude. Broadwell is coming to desktop, period. In BGA form. It might not be coming as Broadwell-K in socketed form, but that's not the same thing as not coming to desktop. 65w chip = desktop.

Actually, I'm not. I already explained what I meant by canceled. I've layed out the roadmap in this thread, feel free to agree to disagree on my assessment.

Obviously you are though, because your argument just hinged on telling me BGA still count as desktops.( I.e. arguing over the definitions of words) We could debate that issue at length, but it doesn't relate to what actual broadwell chips are being released.

Broadwell-k is in fact not cancelled, I never said it was. Broadwell-K is the name Intel is using the broadwell GT3e 65w parts.

 

[IntelUser2000]

I remember Prescott to Conroe was hyped big even a good 6 months before Conroe was released. I'm not seeing any of the same hype for Skylake.

I am thinking...

4GHz Haswell vs 3GHz Skylake = Skylake 5-10% faster

That's probably the best case scenario. Because 22nm and 14nm processes offer zero clock speed increases, lowering clocks to increase IPC offers them few years of slight increase in clocks every generation to get back to 4GHz.

 

[aigomorla]

For whatever reason I remember conroe decimating the athlons worse than those charts show.

thats cuz people were overclocking them like MAD!
:thumbsup:



stock to stock, it wasnt that bad... now if u factor in the NOS.. err overclocking...

And the Q6600 just about decimated everything that it came in touch with...
it was literally GAME OVERS.... :wub:

I remember Prescott to Conroe was hyped big even a good 6 months before Conroe was released. I'm not seeing any of the same hype for Skylake.

The hype came from the mobile p3's they converted to desktop.

People knew conroe was a P3 variant upgraded for desktop use, hence they already had a ballpark idea on how it would perform.

 

[Fjodor2001]

Still miss the boiler functionality on Pentium 4 Prescott:

https://www.youtube.com/watch?v=sG3wceb05Sg

What are we going to cook pasta on now?

 

[erunion]

do you really have to ask?

http://www.newegg.com/Product/Product.aspx?Item=N82E16819113347

 

[mikk]

I am thinking...

4GHz Haswell vs 3GHz Skylake = Skylake 5-10% faster

That's probably the best case scenario. Because 22nm and 14nm processes offer zero clock speed increases, lowering clocks to increase IPC offers them few years of slight increase in clocks every generation to get back to 4GHz.

There are Stepping 1 ES samples clocking already 2.6 Ghz and even BDW 65W with GT3e aimed for AIO clocks 3.3-3.8 Ghz for the fastest SKU. I doubt 3 Ghz SKL is a best case scenario. Also the removal of FIVR might be a good thing for higher clocking variants.

 

[IntelUser2000]

There are Stepping 1 ES samples clocking already 2.6 Ghz and even BDW 65W with GT3e aimed for AIO clocks 3.3-3.8 Ghz for the fastest SKU. I doubt 3 Ghz SKL is a best case scenario. Also the removal of FIVR might be a good thing for higher clocking variants.

Sure ok.

4GHz Haswell vs 3.5GHz Skylake = Skylake 5-10% faster

Better?