News TSMC Kicks Off 3nm Production: A Long Node to Power Leading Chips

[Admin]

TSMC produces its first batch of 3nm chips, says yields are good.

TSMC Kicks Off 3nm Production: A Long Node to Power Leading Chips : Read more

 

[zecoeco]

Just in time for AMD's Zen 5 Processors.
Can't wait to see what a big.LITTLE Ryzen can do

 

[bit_user]

Just in time for AMD's Zen 5 Processors.
Can't wait to see what a big.LITTLE Ryzen can do

The next Ryzen will be hybrid? Is there a good source on that?

Zen4 cores are more area-efficient than Golden Cove, so there wouldn't seem to be as much upside for them to do it. Still, I could believe it.

 

[Murissokah]

Meanwhile, I'm struggling to read a short article with 5 meanwhiles.

 

[bit_user]

Meanwhile, I'm struggling to read a short article with 5 meanwhiles.

There are about 0.5 per paragraph. Not ideal, but I'll happily put up with it for the free news coverage.

Anton has written a lot about the fabrication side of the semiconductor industry for both this site and Anandtech. It's an area he seems to know quite well.

 

[blacknemesist]

I can almost taste the "75% increase across the board with 25% less power consumption" marketing(not even going to touch on price) but still getting the plain old "throw more power at it until it can't hold anymore" architecture being implemented, 900W? No problem, we got power supplies and 16 pin connectors for days!

 

[Murissokah]

There are about 0.5 per paragraph. Not ideal, but I'll happily put up with it for the free news coverage.

Anton has written a lot about the fabrication side of the semiconductor industry for both this site and Anandtech. It's an area he seems to know quite well.

I was being annoying, but it seems to have payed off as the text is now reviewed.

 

[TerryLaze]

The next Ryzen will be hybrid? Is there a good source on that?

Zen4 cores are more area-efficient than Golden Cove, so there wouldn't seem to be as much upside for them to do it. Still, I could believe it.

AMD was forced to match intel stock power consumption and still barely matched 13th gen, they have to do something to cram more performance into the same power envelope if they want to keep making money.

All big cores just doesn't cut it, with too many of them they lose too much performance due to having to share the available power.

And no, there is no good source on that, just speculations because tsmc does a performance and a efficient version of all their nodes and it just makes sense that amd will have to do hybrid at some point.

 

[zecoeco]

AMD was forced to match intel stock power consumption and still barely matched 13th gen, they have to do something to cram more performance into the same power envelope if they want to keep making money.

All big cores just doesn't cut it, with too many of them they lose too much performance due to having to share the available power.

And no, there is no good source on that, just speculations because tsmc does a performance and a efficient version of all their nodes and it just makes sense that amd will have to do hybrid at some point.

They haven't really matched 13th in power consumption, they are still more efficient and less power hungry.

And Yes, it was rumored 6 months ago that Zen 5 will combine dense Zen 4c and Zen5 cores, a mix of advanced and older nodes in a hybrid arch (3nm+5nm).

 

[DavidC1]

Samsung isn't ahead of TSMC. How about hiring writers that know what they are writing about or do due diligence? In fact you probably need Samsung's GAA transistors to have a chance of catching up to TSMC's 4nm.

SS nodes are pure marketing nonsense. It would be if Intel called 14nm+ 13nm, and 14nm++ 12nm.

 

[TerryLaze]

They haven't really matched 13th in power consumption, they are still more efficient and less power hungry.

Oh, tsmc/amd is far behind both power consumption as well as efficiency.
AMD only pulls ahead in multi because they use more big cores and benchmarks focus more on software that utilizes full cores better.
https://www.techpowerup.com/review/intel-core-i9-13900k/22.html

All power measurements on this page are based on a physical measurement of the voltage, current and power flowing through the 12-pin CPU power connector(s), which makes them "CPU only", not "full system." We're not using the software sensors inside the processor, as these can be quite inaccurate and will vary between manufacturers. All measurements are collected and processed at a rate of 30 data points per second, on a separate machine, so the power measurement does not affect the tested system in any way. Our new data processing pipeline allows us to link recorded data precisely with benchmark runs, so we can easily create the charts below.

 

[Endymio]

Further confirmation that Moore's Law is dead.

 

[BGA DRAM]

Sapphire Rapids have lower all-core boost clocks and single-thread performance due to Golden Cove thermal design laxity.
Intel's design beats AMD only in situations where unlimited power is allowed.
In the server segment, which creates a lot of added value, Golden Cove is naturally inefficient.

 

[bit_user]

Oh, tsmc/amd is far behind both power consumption as well as efficiency.

Spoiler: TechPowerUp single-threaded benchies

https://www.techpowerup.com/review/intel-core-i9-13900k/22.html

It's a mistake to focus on single-threaded efficiency, because these Ryzens have the overhead of the I/O die. As they increase the number of active threads, Ryzens amortize that overhead, and eventually overcome it.

AMD only pulls ahead in multi because they use more big cores and benchmarks focus more on software that utilizes full cores better.

LOL, nice try. We all know little cores are more efficient. AMD is at a disadvantage for not having a hybrid CPU, but it's truly a testament to Zen 4 that they're able to overcome that and the IOD overhead, in order to pull out a win on multi-threaded efficiency.

 

[TerryLaze]

It's a mistake to focus on single-threaded efficiency, because these Ryzens have the overhead of the I/O die. As they increase the number of active threads, Ryzens amortize that overhead, and eventually overcome it.

That doesn't change the fact that ryzen does need the IO to operate and that they need that much power to operate a single core, also intel also needs IO to operate so the comparison is on a fair level.
Yeah AMD has a completely inefficient IO chip, whos fault is that?! Yet everybody thinks that AMD CPUs work without them somehow?

LOL, nice try. We all know little cores are more efficient.

More power efficient,yes. That's my whole point, that's why I said that it would make sense for AMD to adopt them as well.
But they don't produce the same amount of performance as a big core even when the big cores are power limited.

 

[brandonjclark]

Meanwhile, I'm struggling to read a short article with 5 meanwhiles.

You guys are counting my "know'm sayins?"

 

[Deleted member 14196]

You guys are counting my "know'm sayins?"

Ok J-Roc

 

[jackt]

Can we hope in a cpu gpu refresh ? like the same chips today made in 3nm ? ...soon ?

 

[bit_user]

That doesn't change the fact that ryzen does need the IO to operate

Not the APUs, at least so far. The energy-efficiency of those CPUs is much better than the chiplet versions.

and that they need that much power to operate a single core, also intel also needs IO to operate so the comparison is on a fair level.

The issue is what you're trying to infer from that data. If you're trying to infer some general truth about the Zen 4 core, then it's a poor test.

Yeah AMD has a completely inefficient IO chip, whos fault is that?!

It's not really that the I/O Die is inefficient, it's more that it's even there. The chip-to-chip interconnect, itself, burns power.

More power efficient,yes. That's my whole point, that's why I said that it would make sense for AMD to adopt them as well.
But they don't produce the same amount of performance as a big core even when the big cores are power limited.

When you start with a collection of P-cores and then add E-cores, the resulting efficiency only improves. So, your argument that Ryzen 7000 somehow wins at multithreaded efficiency because it has only P-cores is completely bonkers. That's a liability it has to overcome, and it does!

 

[bit_user]

Can we hope in a cpu gpu refresh ? like the same chips today made in 3nm ? ...soon ?

Not CPUs, at least. However, AMD is releasing Zen4-based low/mid-range laptop APUs made on the N4 node, in 2023.

I hope AMD will refresh its 7000-series GPUs, but I'm not holding my breath.

 

[TerryLaze]

Not the APUs, at least so far. The energy-efficiency of those CPUs is much better than the chiplet versions.

If you can find a single core test that shows actual power consumption like the one I linked then please share it.

The issue is what you're trying to infer from that data. If you're trying to infer some general truth about the Zen 4 core, then it's a poor test.

There is nothing to infer here, to make one single zen or 13th gen core do that work you will need that much power.
The infering begins when you start arguing if you are going to have the CPU always work at 100% or not.

It's not really that the I/O Die is inefficient, it's more that it's even there. The chip-to-chip interconnect, itself, burns power.

Again intel also needs extra stuff to make things work.
Neither one can sell cores in a vacuum.

When you start with a collection of P-cores and then add E-cores, the resulting efficiency only improves. So, your argument that Ryzen 7000 somehow wins at multithreaded efficiency because it has only P-cores is completely bonkers. That's a liability it has to overcome, and it does!

At least for 12th gen 2 P-cores have a result of about 5000 which drops to about 4000 with more cores because of power.
Two e-cores are about 2000 and also reduce as more and more cores share power.
If you take a look here, the 8 p-cores provide 18820 which comes out to 2352 per group of core+HTT
While if you subtract the dual core+HTT score of the p-cores(5303) from the e-cores result then 12844-5303=7541 and divided by the 8 e-cores that is a score of 942.
So the tldr:
8 big cores while all of them at full load=each one( +HT) gets 2352,
all of 8 e-cores running at full load= each one gets 942
A 16 p-core CPU from intel that could get enough power (or if zen 4 could) , would blow the 13900k away.

But that is the issue here, you can't have many big cores running with all the power they need.
https://www.pugetsystems.com/labs/a...-e-cores-compare-2289/#Efficient-Core_Scaling

 

[bit_user]

At least for 12th gen 2 P-cores have a result of about 5000 which drops to about 4000 with more cores because of power.
Two e-cores are about 2000 and also reduce as more and more cores share power.
If you take a look here, the 8 p-cores provide 18820 which comes out to 2352 per group of core+HTT
While if you subtract the dual core+HTT score of the p-cores(5303) from the e-cores result then 12844-5303=7541 and divided by the 8 e-cores that is a score of 942.
So the tldr:
8 big cores while all of them at full load=each one( +HT) gets 2352,
all of 8 e-cores running at full load= each one gets 942
A 16 p-core CPU from intel that could get enough power (or if zen 4 could) , would blow the 13900k away.

But that is the issue here, you can't have many big cores running with all the power they need.
https://www.pugetsystems.com/labs/a...-e-cores-compare-2289/#Efficient-Core_Scaling

You lost the plot. We were talking about energy-efficiency. Running all 8 E-cores uses nowhere near the full package power budget. This shows them using only 48 W.

https://www.anandtech.com/show/1704...hybrid-performance-brings-hybrid-complexity/4

Even though they don't perform as well as P-cores, their energy-efficiency is much better. The more E-cores you add into the mix, the better your overall energy-efficiency gets.

 

[TerryLaze]

You lost the plot. We were talking about energy-efficiency. Running all 8 E-cores uses nowhere near the full package power budget. This shows them using only 48 W.
...
Even though they don't perform as well as P-cores, their energy-efficiency is much better. The more E-cores you add into the mix, the better your overall energy-efficiency gets.

Oh, tsmc/amd is far behind both power consumption as well as efficiency.
AMD only pulls ahead in multi because they use more big cores and benchmarks focus more on software that utilizes full cores better.

I don't know what plot you where following.

 

[bit_user]

I don't know what plot you where following.

You did, though. You just quoted it. You've even agreed with me that going hybrid would help AMD's efficiency. So, I'm having trouble believing you. I think you were trying to change the subject.

There's no way that this statement is true:

"AMD only pulls ahead in multi because they use more big cores and benchmarks focus more on software that utilizes full cores better."​

If you hold to this line, then I can only conclude that you're arguing in bad faith. E-cores are more efficient, period. Show me a single workload where this isn't true. That's the only way you can support this argument. Otherwise, just take the loss and move on.