It's because they are used in a different place on the chip.
K (or Kappa κ) is a material property, that determines how much capacitance you get when it is used as the dielectric (insulator) between two metal plates (or wires).
For wires (what this article is talking about), you want minimum capacitance, because higher capacitance create more load on the transistor driving it and making it slower. Low-K is getting important here, as wires move closer and closer together, the capacitance between them rises, while the narrower wires also have more resistance (and resistance times capacitance is basically what determines the speed of the wire).
High-K dielectrics are used for the Gates of the transistors (actually the dielectric below the gate, often called the gate oxide). For a transistor, we actually want a high gate capacitance, as the driving strength (maximum current the transistor lets through) depends on that. Previously that was achieved by just making the gate dielectric thinner, but if we can't make it too thin, because then quantum effects become a problem and increase the leakage current significantly. So high-k dielectrics as a gate dielectric allow a higher driving strength without making it thinner.
Side Note: another way to increase the driving strength of the transistor is by increasing the gate area. This is basically how FinFet (and soon GAA) works, it increases the gate area by using a 3D-fin without making the transistor larger.
"The combination of the two can lead to increased power consumption while also limiting performance scaling, which is particularly problematic for datacenter grade processors that are looking to have it all."
At this point it is blindingly obvious that nobody cares about power consumption anymore. With CPUs approaching 600w power consumption, and GPUs approaching 1-1.5kW of consumption, not a single company is actually focused on efficiency.
New installations routinely top 150MW of power draw, with around 120kW per rack of power draw. This is the 70's & 80's gas guzzler generation of compute power. Power at any cost, we don't care if it isn't efficient. That's not what sells.
we have many more cpu cores, with decent generational ipc gains (especially amd), with low power models or simply toggling a tdp in uefi, gpus with tdp control, options for everyone's needs... it's not a gas guzzler if the performance and throttle control is there
If they didn't care about power consumption, then the current chips would self fry. They'd commit silicon Seppuku. AMD/Intel work to reduce power usage and then reinvest the savings into higher performance. Granted, the extremely high power targets of today's offerings are overly aggressive, but that doesn't change the fact that, at lower clocks, these CPUs can be very efficient.
Take for example, AMDs offerings running at 65W. They were found to have very high perf/watt by many benchmarkers -- at least last I read.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
7 Comments
Back to Article
ballsystemlord - Friday, July 12, 2024 - link
It would be helpful if the above article explained why they'd use low-K dielectrics after the industry worked so hard to develop high-K dielectrics.AlexDaum - Monday, July 15, 2024 - link
It's because they are used in a different place on the chip.K (or Kappa κ) is a material property, that determines how much capacitance you get when it is used as the dielectric (insulator) between two metal plates (or wires).
For wires (what this article is talking about), you want minimum capacitance, because higher capacitance create more load on the transistor driving it and making it slower.
Low-K is getting important here, as wires move closer and closer together, the capacitance between them rises, while the narrower wires also have more resistance (and resistance times capacitance is basically what determines the speed of the wire).
High-K dielectrics are used for the Gates of the transistors (actually the dielectric below the gate, often called the gate oxide). For a transistor, we actually want a high gate capacitance, as the driving strength (maximum current the transistor lets through) depends on that.
Previously that was achieved by just making the gate dielectric thinner, but if we can't make it too thin, because then quantum effects become a problem and increase the leakage current significantly.
So high-k dielectrics as a gate dielectric allow a higher driving strength without making it thinner.
Side Note: another way to increase the driving strength of the transistor is by increasing the gate area. This is basically how FinFet (and soon GAA) works, it increases the gate area by using a 3D-fin without making the transistor larger.
James5mith - Saturday, July 13, 2024 - link
"The combination of the two can lead to increased power consumption while also limiting performance scaling, which is particularly problematic for datacenter grade processors that are looking to have it all."At this point it is blindingly obvious that nobody cares about power consumption anymore. With CPUs approaching 600w power consumption, and GPUs approaching 1-1.5kW of consumption, not a single company is actually focused on efficiency.
New installations routinely top 150MW of power draw, with around 120kW per rack of power draw. This is the 70's & 80's gas guzzler generation of compute power. Power at any cost, we don't care if it isn't efficient. That's not what sells.
kn00tcn - Saturday, July 13, 2024 - link
we have many more cpu cores, with decent generational ipc gains (especially amd), with low power models or simply toggling a tdp in uefi, gpus with tdp control, options for everyone's needs... it's not a gas guzzler if the performance and throttle control is thereballsystemlord - Sunday, July 14, 2024 - link
If they didn't care about power consumption, then the current chips would self fry. They'd commit silicon Seppuku.AMD/Intel work to reduce power usage and then reinvest the savings into higher performance. Granted, the extremely high power targets of today's offerings are overly aggressive, but that doesn't change the fact that, at lower clocks, these CPUs can be very efficient.
Take for example, AMDs offerings running at 65W. They were found to have very high perf/watt by many benchmarkers -- at least last I read.
GeoffreyA - Sunday, July 14, 2024 - link
Compared to the Athlon and P4 days, today's CPUs are certainly more frugal in the low- and mid-range.AlexDaum - Monday, July 15, 2024 - link
Found a small mistake in the article:"Volta Ruthenium Copper Vapor Deposition" should be "Volta Ruthenium Chemical Vapor Deposition".