10-28-2013, 05:43 PM
My turn to play armchair analyst! I love this game.
Not true. Sandy bridge was a whopping 40% faster than the 1st generation models on average. Almost as big of a jump as pentium D to core 2 duo was (which was about a 50% jump).
That's because the netburst architecture was a shitty design based on misrepresented research. Anything was better than that. It was so bad that they had no choice but to abandon it completely on laptops back in 2003 with the introduction of pentium M. After pentium 4 prescott made minimal gains in performance over northwood and the architecture remained stagnant for 2 years after that Intel finally released a polished up version of pentium M to the desktop that they called core 2 duo in a last minute desperate attempt to shove netburst under the rug and pretend it never happened. Compared to netburst which had not seen any significant gains in 3 years at that point it looked like a big leap. However had they stuck with P6 and never gone down the netburst road to begin with we would have seen steady 10-20% gains every year from incremental minor improvements to P6 (following the tick-tock development model) instead of nothing for 3 years followed by a sudden 50% increase when they switched to the good architecture. And that's what we saw from AMD during this era who took that route instead.
Now as for why we've been stuck with 10-20% gains in single threaded performance for the last 13 years (closer to 10% recently) that's because of variety of factors in manufacturing and design of semiconductors including:
1. Increasing voltage leakage from smaller transistors
2. Increasing leakage and general power consumption from higher frequency signals
3. Increased power density and heat density. The net results of points 1 and 2.
4. Lack of increase in cooling efficiency. The net result of points 1, 2, 3, and 4 is a rise in temperature.
5. Insignificant increase in thermal tolerance.
6. The result of the above 5 points is extreme difficulty in increasing clock rate. Temperature will inevitably be driven upwards by points 1 through 4 but thanks to point 5 the temperature can't be increased without causing instability. Which makes clock rate increases impossible without significant power use optimizations, which are extremely hard to pull off. This forces designers to rely almost entirely on IPC gains to improve performance.
7. It's getting harder and harder for further IPC optimizations to be made as the "easy" ones have already been done. This means performance growth slows down at the core level as the only two remaining ways to increase performance (IPC and power optimizations) are both very hard to pull off at this point.
You can't expect them to keep making massive improvements to the efficiency of the design forever. There has to be a point where it becomes so efficient that there just isn't much more that can be done to make it any better. I believe we are quickly approaching this point. The only reason they've been able to maintain this 10-20% annual growth for this long is through spending billions of dollars annually on R&D to make these optimizations. Making the impossible tasks possible by just throwing mountains of cash at them (and an army of the best electronic engineers on the planet hired with that cash). They constantly do things that shouldn't be physically possible to get these "minor gains". Many of which are not even disclosed to the public. They've already implied that they have discovered some magic way around the quantum tunneling effect that will hit them in a few years.
Now this brings up a third point. So far all I've talked about is single threaded performance. What about multithreaded performance? Well they could have easily made significant gains there every year. As much as 45% annually. However doing so would have required sacrificing single threaded performance. Bringing the gains down to 0% at the core level. Reducing single threaded performance would allow for massive gains in multithreaded performance by freeing up a lot of power. Since these design tradeoffs would only benefit a small set of applications that could utilize a large number of cores efficiently they instead tried to strike a balance between the two in order to improve performance as much as possible in the range of applications that their typical users use. We haven't reached the point yet where a many core processor would be useful to a typical user. The number of things that it would be useful for is small and most of them are things that only engineers and researchers would be doing or things that are already being done on GPUs instead.
Maybe soon, maybe never. Who knows. Even Intel doesn't know that since their products don't always meet the engineers initial predictions once their fabricated. Probably would point towards "never" but nobody can really be sure.
Incorrect. Second gen. is sandy, third gen. is ivy, fourth gen. is haswell, 5th gen. is broadwell.
Not really. More on increasing power efficiency (ratio of performance to power consumption). Load power consumption goes up or down slightly every year but hasn't really changed much. Higher power efficiency enables them to make both products that achieve the same performance with less power and products that consume the same power with better performance. Filling different product niches, which is what they're doing. The actual microarchitecture itself isn't really aimed at either side of the scale though.
I would debate this as well. The vast majority of their budget and design emphasis still goes into increasing cpu performance. With the possible exception of some haswell variants IGP performance growth has remained fairly consistent over the last 10 years including in recent years. There doesn't appear to be any sudden shift towards GPU performance. Just normal growth that happens every year like clockwork. The die is still almost entirely cpu side stuff as well. That may change soon though.
I slightly disagree here. They sell more laptop cpus because there is more of a demand for laptops. Regardless of how much competition there is from AMD. Even if AMD only had marketshare in the mobile market laptop cpu sales for Intel would still greatly outnumber desktop cpu sales. Making them more inclined to focus on the mobile market.
While there does appear to be some increase in focus on mobile devices I think you might be exaggerated its effect on the microarchitecture. They do focus on improving power efficiency a lot but that's not just because of laptops. It's practically the only remaining means of improving performance regardless of platform. Virtually any optimizations they make to improve performance effect all platforms equally. The architectures design approach is likely fairly platform neutral as it will be used in everything from tablets all the way up to high end servers. They have made some changes that only benefit laptops but most of these were fairly simple changes and reallocating that time elsewhere to other optimizations likely wouldn't have boosted performance significantly. Hell even if they did focus on desktops you likely wouldn't notice any difference because like I said earlier both platforms need the same improvements for the most part to improve.
As far as actual products are concerned both laptop and desktop cpus have grown in performance equally. Your sentence implies reduction on desktops in favor of laptops.
As far as ultra mobile is concerned they are only just beginning to establish any sort of presence at all in that market.
Yeah, no. This I strongly disagree with. Intel has spent 40 billion dollars annually every year for the last 5 years. The vast majority of that (excluding expenses that they can't spend freely like taxes) goes to R&D for design and manufacturing towards improving cpu performance. AMD could be big, small, or dead and this wouldn't change. They can't sell new cpus if they aren't better than the old ones.
If AMD was more competitive they wouldn't be spending more on improving their architecture because there isn't any additional money they could be spending! Their profits are surprisingly thin due to their super high operating expenses. In fact they would have worse cpu performance because increased market share for AMD would mean less money for them to spend on R&D.
Your second sentence makes it sound like they just waited around until their cpu performance was faster than the competition then stopped improving it in order to focus on other things. Which is not true at all. All three of the things you listed have continued to improve at a fairly constant rate.
They're barely going to make any money off of that. Since they're not handling the manufacturing they're likely only going to make a couple hundred million dollars over the next few years off of the royalties according to what the financial industry analysts are saying. Which is a drop in the bucket to them.
Actually that's exactly what they did with llano. They did exactly what Intel did before them. They just decided to give it a fancy marketing name for no good reason.
In the future they may begin to become more than a typical IGP as HSA begins to take shape. But for now that's all they are.
And that's all I have time for today. I'm not waiting until tomorrow to finish my post. I'll get to you later DarkeoX.
drhycodan Wrote:Ever since the original Core i7, these newer 2nd, 3rd, and 4th gen i7's barely improve upon the original i7 in performance.
drhycodan Wrote:Then Core i7 came out and it also offered a substantial improvement over the Core 2's. But since then, the newer Core i7's barely offer any improvement.
Not true. Sandy bridge was a whopping 40% faster than the 1st generation models on average. Almost as big of a jump as pentium D to core 2 duo was (which was about a 50% jump).
drhycodan Wrote:I remember when the Core 2 came out, it was like day and night compared to a Pentium D/4.
That's because the netburst architecture was a shitty design based on misrepresented research. Anything was better than that. It was so bad that they had no choice but to abandon it completely on laptops back in 2003 with the introduction of pentium M. After pentium 4 prescott made minimal gains in performance over northwood and the architecture remained stagnant for 2 years after that Intel finally released a polished up version of pentium M to the desktop that they called core 2 duo in a last minute desperate attempt to shove netburst under the rug and pretend it never happened. Compared to netburst which had not seen any significant gains in 3 years at that point it looked like a big leap. However had they stuck with P6 and never gone down the netburst road to begin with we would have seen steady 10-20% gains every year from incremental minor improvements to P6 (following the tick-tock development model) instead of nothing for 3 years followed by a sudden 50% increase when they switched to the good architecture. And that's what we saw from AMD during this era who took that route instead.
Now as for why we've been stuck with 10-20% gains in single threaded performance for the last 13 years (closer to 10% recently) that's because of variety of factors in manufacturing and design of semiconductors including:
1. Increasing voltage leakage from smaller transistors
2. Increasing leakage and general power consumption from higher frequency signals
3. Increased power density and heat density. The net results of points 1 and 2.
4. Lack of increase in cooling efficiency. The net result of points 1, 2, 3, and 4 is a rise in temperature.
5. Insignificant increase in thermal tolerance.
6. The result of the above 5 points is extreme difficulty in increasing clock rate. Temperature will inevitably be driven upwards by points 1 through 4 but thanks to point 5 the temperature can't be increased without causing instability. Which makes clock rate increases impossible without significant power use optimizations, which are extremely hard to pull off. This forces designers to rely almost entirely on IPC gains to improve performance.
7. It's getting harder and harder for further IPC optimizations to be made as the "easy" ones have already been done. This means performance growth slows down at the core level as the only two remaining ways to increase performance (IPC and power optimizations) are both very hard to pull off at this point.
You can't expect them to keep making massive improvements to the efficiency of the design forever. There has to be a point where it becomes so efficient that there just isn't much more that can be done to make it any better. I believe we are quickly approaching this point. The only reason they've been able to maintain this 10-20% annual growth for this long is through spending billions of dollars annually on R&D to make these optimizations. Making the impossible tasks possible by just throwing mountains of cash at them (and an army of the best electronic engineers on the planet hired with that cash). They constantly do things that shouldn't be physically possible to get these "minor gains". Many of which are not even disclosed to the public. They've already implied that they have discovered some magic way around the quantum tunneling effect that will hit them in a few years.
Now this brings up a third point. So far all I've talked about is single threaded performance. What about multithreaded performance? Well they could have easily made significant gains there every year. As much as 45% annually. However doing so would have required sacrificing single threaded performance. Bringing the gains down to 0% at the core level. Reducing single threaded performance would allow for massive gains in multithreaded performance by freeing up a lot of power. Since these design tradeoffs would only benefit a small set of applications that could utilize a large number of cores efficiently they instead tried to strike a balance between the two in order to improve performance as much as possible in the range of applications that their typical users use. We haven't reached the point yet where a many core processor would be useful to a typical user. The number of things that it would be useful for is small and most of them are things that only engineers and researchers would be doing or things that are already being done on GPUs instead.
drhycodan Wrote:When is Intel going to release something that offers as much an improvement in IPC as the original i7 did?
Maybe soon, maybe never. Who knows. Even Intel doesn't know that since their products don't always meet the engineers initial predictions once their fabricated. Probably would point towards "never" but nobody can really be sure.
MaJoR Wrote:Uh, wrong. The first gen is Nehalem-Westmere, the second gen Sandy Bridge-Ivy Bridge, the third gen is Haswell-Broadwell. Sandy Bridge was a HUGE leap in performance. So you're wrong there.
Incorrect. Second gen. is sandy, third gen. is ivy, fourth gen. is haswell, 5th gen. is broadwell.
Link_to_the_past Wrote:They have been focusing since the second generation of i7 on reducing power consumption
Not really. More on increasing power efficiency (ratio of performance to power consumption). Load power consumption goes up or down slightly every year but hasn't really changed much. Higher power efficiency enables them to make both products that achieve the same performance with less power and products that consume the same power with better performance. Filling different product niches, which is what they're doing. The actual microarchitecture itself isn't really aimed at either side of the scale though.
Link_to_the_past Wrote:and increasing integrated gpu performance more than increasing cpu power.
I would debate this as well. The vast majority of their budget and design emphasis still goes into increasing cpu performance. With the possible exception of some haswell variants IGP performance growth has remained fairly consistent over the last 10 years including in recent years. There doesn't appear to be any sudden shift towards GPU performance. Just normal growth that happens every year like clockwork. The die is still almost entirely cpu side stuff as well. That may change soon though.
Link_to_the_past Wrote:They moved all their focus to more mobile offerings and i think that the lacking competition from AMD helped establish that.
I slightly disagree here. They sell more laptop cpus because there is more of a demand for laptops. Regardless of how much competition there is from AMD. Even if AMD only had marketshare in the mobile market laptop cpu sales for Intel would still greatly outnumber desktop cpu sales. Making them more inclined to focus on the mobile market.
While there does appear to be some increase in focus on mobile devices I think you might be exaggerated its effect on the microarchitecture. They do focus on improving power efficiency a lot but that's not just because of laptops. It's practically the only remaining means of improving performance regardless of platform. Virtually any optimizations they make to improve performance effect all platforms equally. The architectures design approach is likely fairly platform neutral as it will be used in everything from tablets all the way up to high end servers. They have made some changes that only benefit laptops but most of these were fairly simple changes and reallocating that time elsewhere to other optimizations likely wouldn't have boosted performance significantly. Hell even if they did focus on desktops you likely wouldn't notice any difference because like I said earlier both platforms need the same improvements for the most part to improve.
As far as actual products are concerned both laptop and desktop cpus have grown in performance equally. Your sentence implies reduction on desktops in favor of laptops.
As far as ultra mobile is concerned they are only just beginning to establish any sort of presence at all in that market.
Link_to_the_past Wrote:If AMD was competitive they wouldn't be so lax regarding their cpu performance. Having established their dominance in cpu performance then they shifted their focus to areas they lacked, integrated gpu performance and ultra mobile, tablet, etc. presence.
Yeah, no. This I strongly disagree with. Intel has spent 40 billion dollars annually every year for the last 5 years. The vast majority of that (excluding expenses that they can't spend freely like taxes) goes to R&D for design and manufacturing towards improving cpu performance. AMD could be big, small, or dead and this wouldn't change. They can't sell new cpus if they aren't better than the old ones.
If AMD was more competitive they wouldn't be spending more on improving their architecture because there isn't any additional money they could be spending! Their profits are surprisingly thin due to their super high operating expenses. In fact they would have worse cpu performance because increased market share for AMD would mean less money for them to spend on R&D.
Your second sentence makes it sound like they just waited around until their cpu performance was faster than the competition then stopped improving it in order to focus on other things. Which is not true at all. All three of the things you listed have continued to improve at a fairly constant rate.
delroth Wrote:AMD is far too busy swimming in the PS4 and XBone money to make good CPUs.
They're barely going to make any money off of that. Since they're not handling the manufacturing they're likely only going to make a couple hundred million dollars over the next few years off of the royalties according to what the financial industry analysts are saying. Which is a drop in the bucket to them.
DatKid20 Wrote:AMD doesn't just integrate graphics and call it a apu.
Actually that's exactly what they did with llano. They did exactly what Intel did before them. They just decided to give it a fancy marketing name for no good reason.
In the future they may begin to become more than a typical IGP as HSA begins to take shape. But for now that's all they are.
And that's all I have time for today. I'm not waiting until tomorrow to finish my post. I'll get to you later DarkeoX.