Yes another Haswell thread. Let's have a look at tock-to-tock IPC.

[Hulk]

So I’ve had a chance to step back and think about Haswell a bit and have some thoughts.

Let’s have a look at only “tock” performance. Ticks are really only intended to be die shrinks with a few small improvements. Here is the data from Anand’s Haswell review with clockspeeds equalized. I’ve tried to only include single threaded benches but needed to include a few additional to get a meaningful P4 analysis. But as we know the CS4 test isn’t heavily multithreaded.

Cinebench R10 – Single Threaded
Increase from Sandy Bridge to Haswell – 21.1%
Increase from Nehalem to Sandy Bridge – 14.1%
Increase from Conroe to Nehalem – 23.5%
Increase from Pentium 4 to Conroe – 46.1%

Sysmark 2007
Increase from Sandy Bridge to Haswell – 20.7%
Increase from Nehalem to Sandy Bridge – 1%
Increase from Conroe to Nehalem – 21.7%
Increase from Pentium 4 to Conroe – 52.5%

Adobe CS4 Artist Retouch
Increase from Sandy Bridge to Haswell – 13.7%
Increase from Nehalem to Sandy Bridge – 16.9%
Increase from Conroe to Nehalem – 33.6%
Increase from Pentium 4 to Conroe – 53.5%

X264 HD 5.0.1 First Pass
Increase from Sandy Bridge to Haswell – 16.7%
Increase from Nehalem to Sandy Bridge – 15.5%

X264 HD 5.0.1 Second Pass
Increase from Sandy Bridge to Haswell – 21.6%
Increase from Nehalem to Sandy Bridge – 11.1%

7-zip Benchmark – Single Threaded
Increase from Sandy Bridge to Haswell – 9.5%
Increase from Nehalem to Sandy Bridge – 10.7%
Increase from Conroe to Nehalem – 10.5%

Kraken Javascript Benchmark (Chrome)
Increase from Sandy Bridge to Haswell – 11.5%
Increase from Nehalem to Sandy Bridge – 32.8%

Looking at the averages from above it’s clear that P4 to Conroe was a HUGE jump in IPC. Conroe to Nehalem about half as much, and Nehalem to Sandy and Sandy to Haswell even a smaller IPC increase. But Intel has managed to increase IPC.

Averages
Increase from Sandy Bridge to Haswell – 16.4%
Increase from Nehalem to Sandy Bridge – 14.6%
Increase from Conroe to Nehalem – 22.3%
Increase from Pentium 4 to Conroe – 50.7%


Observation #1 – After enormous IPC gains from P4 to Conroe, IPC gains have leveled off at between 15 and 22% over the last three tocks.


Since there is only so much instruction level parallelism Intel can extract from single threaded code it will be interesting to see how much additional IPC Intel has left on the table. The execution engine is wider than ever, out of order engine refined over a few generations now, and plenty of additional clever tricks have been employed to move instructions through the pipe.

Observation #2 – If there was any pressure from AMD we would most likely see higher clockspeeds from Intel. I think we can all agree that if Intel wanted to they could release some top binned parts that could clock to 5GHz or close to it if they wanted to invest in the package/cooler improvements to get this done. Indeed we should be glad they don’t because if they did that they might vaporize the “K” series all together and send the message that if you want the fastest part you will have to pay dearly for it.

Observation #3 – With no pressure in absolute performance, IPC gains getting harder and harder to come by, and the market shifting to low TDP tablets and phones, it’s easy to see why Intel would put the Haswell’s emphasis on super high efficiency and low TDP.

Would I like to see some aggressively binned super high frequency parts at least available? Of course. If just to show us the capability of the parts Intel is producing. But if you look at the facts they have absolutely no reason to be aggressive in that part of the market. They own it. It’s in the sub 10W area where they need to focus.

Observation #4 – Haswell won’t be a failure by any means for Intel. It’s a better performing part than Ivy Bridge, and looking at tock-to-tock is actually a bigger improvement than the Sandy to Nehalem Tock. And very few people consider Sandy a failure. In addition, while maintaining IPC improvements Intel has also managed to dramatically improve efficiency and top of the line integrated GPU performance.

We don’t know if or by how much Intel could have tweaked the process and package to improve overclockability. But I bet if Haswell overclocked as well as Sandy, with equal or better thermals then the small but vocal minority of people disappointed in Haswell’s overclock potential would instead be fans.

 

[LogOver]

+100. Thanks for putting things together.

 

[blackened23]

First, thanks for putting that together. Would definitely +rep you if I could, that was very informative and interesting.

But I do have to comment on one of your points: Out of curiosity with regard to point 2: why would intel care about AMD? The fight is for mobile, , The real competitors that intel has in mind are qualcomm, nvidia, and samsung. Not AMD.

I guess AMD putting pressure on intel on the desktop could have an effect, but i'm not entirely convinced. Desktop sales are diminishing very rapidly.

 

[Hulk]

First, thanks for putting that together. Would definitely +rep you if I could, that was very informative and interesting.

But I do have to comment on one of your points: Out of curiosity with regard to point 2: why would intel care about AMD? The fight is for mobile, , The real competitors that intel has in mind are qualcomm, nvidia, and samsung. Not AMD.

I guess AMD putting pressure on intel on the desktop could have an effect, but i'm not entirely convinced. Desktop sales are diminishing very rapidly.

You have a good point. It's just my opinion that Intel holds the fact that they are #1 in microprocessor performance in high regard. And while there might be some corporate pride at play, there is more to it than that.

Technology flows from the top down. If you have the fastest part, not only does your marketing department get to "sell" that fact throughout the line up, but that technology trickles into your lower performance parts. Think about how nice it is for Intel to simply "unleash" their fastest parts, architecture-wise to beat all AMD competition. They can play it safe with clockspeed and kill them on efficiency as well. All the while they have literally hundreds of MHz in reserve if need be.

AMD stung Intel quite painfully with the Athlon. Intel had to recant clockspeed is king, Intel had to recant Netburst and long pipes, in short Intel had to eat crow and reverse course abruptly. It was bad all around for the company and a costly miscalculation. They don't want to repeat that.

If Intel looks vulnerable at the high end, an area they have dominated for 40 years how are they going to get big buys in the tablet and phone market? It seems to me it would just make the sell harder.

While I doubt Intel is "worried" about AMD at this point considering the road they have put between them, I bet they are still taking an occasional glance in the rearview mirror. Just in case...

First rule in business. Protect what you have and then expand.

And finally while the desktop market is diminishing, as process size continues to shrink and TDP along with it, what we once thought of as desktop processors will be tablet and phone processors. It's not like process size and TDP are getting larger with each generation, they are getting smaller. So ARM has to scale up while Intel simply has to tweak what is already the fastest architecture and let physics do it's thing and reduce TDP as process size decreases. At 10nm we could conceivably see 3W Intel parts with the compute of today's desktop parts. That's why the desktop parts are important besides just raw desktop sales. Those desktop CPU's also go into laptops (today), tablets (just starting), and phones (perhaps a tick or tick-tock away).

Intels "top down" approach may well be better positioned for the future than ARM's "bottom up." Until they get there with the big boy chips Atom will do battle with ARM.

 

[frozentundra123456]

Very impressive collection of data. But in previous generations clock speed was also increasing steadily, as was overclocking headroom, so the overall performance improvement was much better.

Haswell is especially dissapppointing in this regard because it is the first time we have not seen a cloockspeed increase over 2 generations. I am not saying it is a bad chip like some of the Intel bashers in this forum, but I am disappointed in this regard.

 

[Hulk]

Very impressive collection of data. But in previous generations clock speed was also increasing steadily, as was overclocking headroom, so the overall performance improvement was much better.

Haswell is especially dissapppointing in this regard because it is the first time we have not seen a cloockspeed increase over 2 generations. I am not saying it is a bad chip like some of the Intel bashers in this forum, but I am disappointed in this regard.

That is most definitely "the other side of the coin." And it's a point I should have mentioned. We definitely have seen a stall in clockspeed increases over the last few years. It seems to be in part do to the physics of the situation. Intel has most likely tuning their processes for low leakage, low voltage, without too much regard for ultimate clock. Because quite simply they are already the fastest available.

Let's say they could tune the process for more volts and higher clocks. And rework the package for better heat transfer, and pair it with a better cooler. I for one have no doubt Intel could release a 5GHz Haswell in short order. But what would it get them? They are losing zero sales for lack of a 5GHz part.

But they are losing lots of sales at the other end of the market. So that is where they are putting their effort.

As we know all too well this is a point of diminishing returns when clocking a chip. You start adding more volts and in return you get more heat and less increase in clock.

The real question is if Intel can maintain a 15% IPC during the next tock? I believe they are kind of waiting for software development to catch up so that increasing performance is more a matter of adding cores rather than increasing IPC or clockspeed. While at the same time I think they also realize that for 90% or more of their customers their processors are... dare I even write it here? "Fast enough."

 

[Zucker2k]

Thanks for this thread. IvyBridge was a great "Tick," as people might find out if they do what you've done here with "Tocks." This is actually how I view Haswell; as being 19%+ points ahead of Sandy Bridge, the previous tock, as per Anand's 4770k review, and in that regard, plus the robust gpu and AVX2, it's frankly a great chip for enthusiasts and people after raw processing power on mainstream desktop.

 

[Maximilian]

Highly interesting numbers OP! :thumbsup:

 

[AtenRa]

Thx for the info, but one observation

You can only count IPC with the single core benchmarks like Cinebench Single Core and 7zip Single Core. The averages of all the benchmarks you have are Performance only increases(higher IPC + higher MT performance and higher Frequency) and not for IPC increase.

I will say that average IPC increase (across a lot of legacy applications) from Sandy to Haswell is about 10-12%.

 

[Idontcare]

Thx for the info, but one observation

You can only count IPC with the single core benchmarks like Cinebench Single Core and 7zip Single Core. The averages of all the benchmarks you have are Performance only increases(higher IPC + higher MT performance and higher Frequency) and not for IPC increase.

I will say that average IPC increase (across a lot of legacy applications) from Sandy to Haswell is about 10-12%.

Which, coincidentally, is exactly what Intel told us it would be (and they told us quite some time before Haswell was released).

 

[AtenRa]

Ahh nice, didnt remember that. IDC, where is that slide from ??

 

[krumme]

Which, coincidentally, is exactly what Intel told us it would be (and they told us quite some time before Haswell was released).

Standard marketing spin.

Now show me the numbers on the x axis?

Secondly how should Intel be able to pan out the ipc of Haswell years ago?

The entire striving for more ipc is bad for the development of the entire industri and is eroding the entire pc market. But typically Intel, the paranoid decease aquired in the P4 age, they think just for the short term selling the idea of ipc to consumers thereby beating AMD even more.

Now we are left in a situation with loads of ultrabook reviews on Anandtech that only a minority buys, and a Haswell arch that might have 7% better IPC than IB, but with performance potential that is perhaps even lower than the arch before.

No wonder the desktop market is going backwards.

If Intel, instead of making silly figures pretending everything is going as planned, had strived for excellent performance overall instead of IPC, we would have been in another situation today. 8 cores, full steam avx and perhaps avx2 support and the likes. Now that would make a difference for the possibilities fx. in gaming, giving unique experiences, and prevent people to switch people from pc to xbox.

Now the pc is just a dying dog giving no unique user benefit. History shows this is just a dead end.

I dont know if the marketing department is sleeping, but its standard knowledge, when you are a big player in the market to focus on the total consumption in the market, even if it helps your competitors. Thats why fx. Colgate told consumers to brush their teeths.

Intel long term strategy have failed. They should have protected their core markets, instead of next years profit. Now they can take their Capex to arm land and learn it the hard way.

 

[Idontcare]

Standard marketing spin.

I don't follow

Plotting data of the IPC for various microarchitectures is "marketing spin"?

Now show me the numbers on the x axis?

The x-axis is already clearly visible in the graph. What more is there to show?

Secondly how should Intel be able to pan out the ipc of Haswell years ago?

Standard project management in engineering. How do people build bridges and know in advance they aren't going to collapse? Or buildings? Or airplanes?

Like anything in engineering, it can be done if it is done correctly. But simply claiming you've done it doesn't mean you did it correctly (see bulldozer, "IPC does not go down").

 

[Vectronic]

I probably spent about 10 minutes just trying to figure out "fx." Reading previous posts by Krumme, going through acronyms, etc. Perhaps by my own stupidity, or just laziness, I have now concluded that it's meant as "for example", ie: eg.

My skull IPC needs increased performance.

 

[JimmiG]

Very impressive collection of data. But in previous generations clock speed was also increasing steadily, as was overclocking headroom, so the overall performance improvement was much better.

Haswell is especially dissapppointing in this regard because it is the first time we have not seen a cloockspeed increase over 2 generations. I am not saying it is a bad chip like some of the Intel bashers in this forum, but I am disappointed in this regard.

Actually, clock speed has "stalled" at 3.5-ish GHz for the past 10 years. The Pentium 4 first hit 3.6 GHz in 2004, and only ever hit a max of 3.8 GHz. That CPU was designed from the ground up for high clock speeds (Intel was aiming for 10 GHz in 2010). The Core architecture literally turned back the clock, and clock speed again slowly began to creep up towards the mid-3.x GHz range.

Now we're at the same point again. Clearly, something makes it nonviable to release chips with stock clocks above 3.5 - 3.8 GHz - in 2004 and in 2013.

The vast overclocking headroom of SB (people seem to suffer from selective memory when it comes to IB) was a happy accident. 50% overclocks have never been the norm.

 

[krumme]

What i object is the lack of dates on the x-axis and the drawing of a straight line at the same time.

As shown by OP its getting harder to get the IPC for obvious reasons. Its not a straight line its a flattening curve. And then we are not even talking performance.

The intention of this figure was to sell the idea of ipc on a broad workload as important. Why not a figure of what mattered - broad workload performance?

For all i cared this figure could be perfectly valid, it just adressed an unimportant problem.

 

[mrmt]

Intel long term strategy have failed. They should have protected their core markets, instead of next years profit. Now they can take their Capex to arm land and learn it the hard way.

I wouldn't call it quit until we see 14nm on mobile. It can turn tables.

 

[krumme]

I probably spent about 10 minutes just trying to figure out "fx." Reading previous posts by Krumme, going through acronyms, etc. Perhaps by my own stupidity, or just laziness, I have now concluded that it's meant as "for example", ie: eg.

My skull IPC needs increased performance.

Sorry yes, you are right - lol , english is not my native language

 

[krumme]

One of the nicest recent boost i had from upgrading, was from Pentium M to Core duo. I think that qualify as a tick?

- but the performance was double, at the same or even lower power consumption compared to Pentium M.

I have a Haswell comming today by mail. Thats a quadcore. Now next time, just give me the same at 14nm and 8 cores with some small tweaks. Then i will stop complaining

...at least for 2 months.

 

[krumme]

I wouldn't call it quit until we see 14nm on mobile. It can turn tables.

We perfectly agree the 14 nm is what matures their current mobile strategy, but there was no reason to nearly abandon the desktop market. A market they practically own now. They could have defended the size of it - like they defend ultrabooks.

 

[mrmt]

We perfectly agree the 14 nm is what matures their current mobile strategy, but there was no reason to nearly abandon the desktop market. A market they practically own now. They could have defended the size of it - like they defend ultrabooks.

They will pratically own the desktop market too. Whatever AMD field in desktop won't be able to deal with Broadwell, neither in graphics nor in CPU power, so why wouldn't they own that market too?

 

[Ancalagon44]

The entire striving for more ipc is bad for the development of the entire industri and is eroding the entire pc market. But typically Intel, the paranoid decease aquired in the P4 age, they think just for the short term selling the idea of ipc to consumers thereby beating AMD even more.

Uhhh how is an IPC increase possibly a bad thing?

 

[krumme]

Uhhh how is an IPC increase possibly a bad thing?

That looks like a strawman to me?

I think i have explained why the focus was wrong?
If yes - You can refer to the arguments there. They might be wrong but we can discuss from there.

 

[krumme]

They will pratically own the desktop market too. Whatever AMD field in desktop won't be able to deal with Broadwell, neither in graphics nor in CPU power, so why wouldn't they own that market too?

Yes they will. And thats excactly why they have failed.
They own a dying market. A market where there is very solid profit for Intel.

My thesis is, had they instead of focusing on expanding and development the market around performance, instead of beating AMD, they would be in a far better position today. They shot themselves in the foot.

Now i will go and asemble my Haswell system for my boy. He doesnt even want to open the package that just arived. He doesnt even want to look at it ! Man - and here his father does everything for him He just want to get rid of the wireless and have a 1g cable. Damn kids theese days, they dont care for tech.

And perhaps thats exactly the point. Intel just lost a new generation.

 

[blackened23]

I wonder if any website has done any similar generation to generation statistical/powerpoint comparisons in terms of performance-per-watt or efficiency? I'm really interested in how that has progressed. I know that in particular has increased exponentially within the past 5 years.

On a side note: I know this may be an unpopular view, but i'm REALLY excited about Haswell ultra portables. Hearing about the macbook air with 11 hours of battery life - nearly double last years model is particularly impressive. I have to say, mobile computing has really grown on me - and while I still have applications that require desktop performance (not to mention games), I love being able to do stuff on the go.