Jump to content
Search In
  • More options...
Find results that contain...
Find results in...
Master Disaster

Dr Su will present Zen 2 and Navi at Computex 2019 (May 27th)

Recommended Posts

8 hours ago, Stefan Payne said:

With the same voltage it is more on the linear side but with higher clockrates we tend to increase the voltage, wich dramatically increases the Power Consumption.

The behaviour has been well understood by overclockers. If you run a CPU to its limit in a given condition, then generally more clock needs more voltage, and it isn't a linear relationship. In particular, there is the dreaded voltage wall, where for small gains in clock you have to use a lot more voltage. To try to overcome that people resort to extreme cooling like LN2. FWIW in my limited attempts with chilled water (close to 0C), it gets me about 100 MHz more than ambient water (on Intel), hardly worth the effort.

 

This is in part why we generally see lower clocks on higher core count CPUs. We can run them in a more efficient part of their clock range to keep the total power consumption down. You still gain from having more cores. That's not to say you can't hit similar clocks on higher core count parts, if you can throw enough power into it, and deal with the resulting heat. A stock 1700 might be rated 65W TDP, but in my attempts to push it to 4 GHz and beyond I hit a CPU reported 180W before reaching the thermal limits of a 280mm AIO and had to stop. I think stock ran at 3.2 GHz all cores and was super cool. So nearly 3x power consumption for 25% more clock isn't a great tradeoff.

 

8 hours ago, Stefan Payne said:

AMD has to increase the CPU Clockrates to really compete with Intel, so I doubt that they didn't adress this "Problem" of the Ryzen design.

The Question is how much did they do to increase the clockrates.

AMD don't necessarily have to increase clock. They've already gone way down the "more cores" road and are marketing it well to enthusiasts and power users who might be able to use it (as well as some enterprise uses of less interest here). That said, it doesn't mean I don't want more clock from Ryzen, but it isn't necessarily the first priority. They also have opportunity to improve performance in consumer use cases by other methods, for example, from cache and IMC improvements. Zen 2 will be different from what we know and it remains to be seen what happens there.

 

It wasn't recent, but Intel had previously mentioned they may design less peak performing CPUs in order to make greater gains elsewhere like in power efficiency. Now that I say it, I wonder what their mix of CPU sales are between laptops and desktops. The laptop market as a whole is supposed to be much bigger, but how does that translate into CPU sales? Point is, they may optimise more for power than performance if that's the biggest part of the market. Desktop CPUs do have the side benefit of not being power constrained so they can still push them into a high performance lower efficiency operating region.

 

8 hours ago, Stefan Payne said:

Remember, we aren't talking about high percentages, we are talking about 10-15% higher clockrates to get to the 5GHz and also not on all cores but "2 Core Boost".

And then overclockers will make all cores do that :) 


Main rig: Asus Maximus VIII Hero, i7-6700k stock, Noctua D14, G.Skill Ripjaws V 3200 2x8GB, Gigabyte Windforce 980Ti, Corsair HX750i, In Win 303 NVIDIA, Samsung SM951 512GB, WD Blue 1TB, HP LP2475W 1200p wide gamut

Gaming system: Asrock Z370 Pro4, i7-8086k stock, Noctua D15, G.Skill TridentZ 3000C14 2x8GB, Asus 1080 Ti Strix OC, Fractal Edison 550W PSU, Corsair 600C, Optane 900p 280GB, Crucial MX200 1TB, Sandisk 960GB, Acer Predator XB241YU 1440p 144Hz G-sync

Ryzen rig: Asrock B450 ITX, R5 2600, Noctua D9L, Corsair Vengeance LPX 3000 2x4GB, Vega 56, Corsair CX450M, NZXT Manta, Crucial MX300 525GB, Acer RT280K

VR rig: Asus Z170I Pro Gaming, i7-6600k stock, Silverstone TD03-E, Kingston Hyper-X 2666 2x8GB, Zotac 1070 FE, Corsair CX450M, Silverstone SG13, Samsung PM951 256GB, HTC Vive

Gaming laptop: Asus FX503VD, i5-7300HQ, 2x8GB DDR4, GTX 1050, Sandisk 256GB SSD

Total CPU heating: i7-7800X, 2x i7-6700k, i7-6700HQ, i5-6600k, i5-5675C, i5-4570S, i3-8350k, i3-6100, i3-4360, 2x i3-4150T, E5-2683v3, 2x E5-2650, R7 1700, 1600

Link to post
Share on other sites
9 hours ago, r2724r16 said:

I'm assuming you mixed Kabylake and Skylake...

 

The 4770K was released in mid-2013. The 7770K was released in early-2017. That's a gap of over 3 years, in which the boost clock increased by 0.6 GHz.

 

How do you expect AMD to increase their boost clock from 4.2/4.3 GHz to 5.1 GHz in just one year?

Fair.

Even if AMD is able to manufacture a 16 core CPU that runs up to 5.1 GHz, do you actually think they'll sell it for ~U$500?

we dont actually have to imagine how big a jump 7nm is in terms of base and boost clock, we just have to look at what amd has achieved with radeon VII, stock performance (similar to all core boost) went from 1500 to 1800 (a little bit lower than that on both cases mostly because of power throttling, though that is not important here), max frequency went from 1750, to 2050+ (radeon VII being a low end bin has very high variability in the silicon so people have got everything from 1900 ish mhz to 2200 mhz),

this is going from 14nm to 7nm, applying the same improvement to zen it would be 4.3ghz base (how interesting the exact same as the leak's base) and 4.8ghz max clocks, with the better binning of chiplets going even further is possible, but 4.8 is a good baseline for what to expect

Link to post
Share on other sites
37 minutes ago, porina said:

The laptop market as a whole is supposed to be much bigger, but how does that translate into CPU sales? Point is, they may optimise more for power than performance if that's the biggest part of the market. Desktop CPUs do have the side benefit of not being power constrained so they can still push them into a high performance lower efficiency operating region.

In a way laptop usage aligns closer with the server market than desktop does as well, performance efficiency comes first. If Intel biases away their architecture and fab design from performance to power we might see a fair reduction in overclocking, very much doubt any performance reduction though because that would be far too hard to sell.

Link to post
Share on other sites
7 hours ago, leadeater said:

You have to increase the voltage to increase the clocks, that's how (current) transistors work and a characteristic of them.

Yes but the power consumption rises/falls linear with the frequency.

Resistive loads increase the power with the power of 2 -> Double Voltage = double the Current -> 4 times the power consumption.

 

With Transistors it might be even worse than that for various reasons.

 

 

So with an increase in frequency of 20% at the same voltage, the Power Consumption also rises by around 20%.

With Voltage it is different, obviously.

7 hours ago, leadeater said:

The increase in voltage allows quicker transitions from off state to on state and also makes the state change easier to distinguish as there is a greater voltage difference between the states.

Yeah also there is a minimal voltage for semiconductors that they need to work. ANd we are darn close to those voltages.

 

7 hours ago, leadeater said:

As transistors get smaller you have to decrease the gate oxide thickness and that starts to bring in problems like leakage/tunneling which makes the gate less efficient.

Yeah and you can't necessarily shrink the transistors these days, as seen with the PHY or other stuff that needs to drive signals over the Motherboard.

In that case you have a current requirement for that circuit that prevents it from beeing shrunk with the processes wich is why AMD threw out the Northbridge/Chipset part of the CPU and put it into a new Die because the size difference between 14nm and 7nm is probably rather negligable (maybe 10%, maybe 15% smaller)

 

7 hours ago, leadeater said:

This also increases power consumption along with another issue of clearly knowing if the transistor is off or on, one of the ways to combat that is to increase the voltage but that also increases the leakage hence non linear power increase to increase frequency at the upper end of the transistor capabilities.

Leakage was one of the things that killed the Pentium 4 back in the day and that was with 90nm.

Its also one of the main problems with modern chipdesign.

 

7 hours ago, leadeater said:

What's changing as we develop new fabrication technology is the metal oxides and the accuracy in which we can print the transistors, EUV is a big change to the latter.

Still silicon on Insulator is dead for whatever reasons.

AMD Used that the last time with the Original Bulldozer in 32nm.

After that the development of SOI regressed.

 

Its said that that helps with some of the Problems of modern CPU Design.

7 hours ago, leadeater said:

You can see that by comparing GloFo 14nm to Intel 14nm, how similar are those?

Ähm, how do we compare the two?
We need a design on both fabs - wich we don't have.


But what we have is a Design on Samsung 14nm and TSMC 16nm. More or less the same Chip. And that Chip showed that TSMC 16nm had a lower power consumption as the Samsung process.

 

Yes, it was an Apple CPU. The A9 to be precise wich was dual sourced on Samsung and TSMC. Though the question that remains:
How well was it optimized on either process??

 

Anyway, that's the only example for us to compare foundrys.

 

7 hours ago, leadeater said:

With the introduction of EUV and the increased accuracy/resolution that comes with it we could start to see more differences between products and SKUs. Like higher base clocks because of the increased transistor efficiency, or lower for more compelling low power options.

Yes and that is the unknown of the Zen2 die.

Is it designed/optimized for highest power efficiency at lower clocks or is it optimized for high frequency?
Or somewhere in between?

 

That is the big unknown right now. But its possible that AMD's design goal for Zen2 was to increase the frequency by 15%. That's not too much increase with an optimized design and a shrink with changing the fab as well.


GF had licensed the Samsung Process, and we already have a comparisation between Samsung and TSMC.

And in this case there was a ~5°C Difference in a mobile device outer shell according to the Article I've found....

 

7 hours ago, leadeater said:

Chiplets is also another major factor because we are now talking about isolation of power delivery of the cores to other parts of the CPU and their transistors. We gain a lot by doing this, heat reduction, heat source size reduction, true isolation. There's other potential gains like being able to increase the core voltage with less potential to effect other areas and transistors in the CPU, we have some control of that today already but there is the possibility for that to be further improved. This is also carried across to multi-chiplet packages where you (should?) have individual voltage regulators for each chiplet meaning greater control and less voltage drop under high power conditions.

I agree with you.

But the Chiplets is a necessity du to physical limitations of lower manufacturing processes. And also the financial ones.

It just makes more sense to throw anything out of the 7nm die for various reasons.

The things you mentioned here is one of the reasons to do it, another is cost. Its just cheaper to do it because according to bits and chips the yield of the 7nm process with AMD is at only around 70%.

 

7 hours ago, leadeater said:

The most interesting and important area for Zen2 isn't actually Ryzen it's Threadripper and EPYC. Either way I'd still keep a fair amount of caution in frequency expectation for Zen2/Ryzen 3000 because it's 7nm, EUV which is very new with unknown challenges and characteristics relating to the first part of this post. 

Yeah for EPYC the frequency isn't important as for the consoles.

So that is two out of 4 applications where the frequency is of no interest of the design. Even a 4GHz max. clock would be more than enough for those situations.

 

 

However it is different on the Ryzen products, wich include Threadripper.

To compete with Intel they need high frequencys of at least close to 5GHz especially if they want to have some headroom for the upcoming 10nm Intel CPUs (though AMD could at that time switch to 6nm)

 


"Hell is full of good meanings, but Heaven is full of good works"

Link to post
Share on other sites
54 minutes ago, Stefan Payne said:

Yes but the power consumption rises/falls linear with the frequency.

No it doesn't, but it's hard to generalize like this at all because transistor type and makeup changes this too.

 

Rough, not CPU transistor example:

Figure-1.png

 

Also:

Quote

As switching frequencies increase, it becomes of paramount importance to reduce the switching losses in the converter. These are the losses associated with the transition of the switch from its on-state to its off-state, and back. The higher the switching frequency, the greater the number of times the switch changes state per second. Therefore, these losses are proportional to the switching frequency. Further, of these frequency-dependent loss terms, the most significant are usually those that take place within the switch itself. Therefore, understanding the underlying sequence of events in the switch during each transition, and thereby quantifying the losses associated with each of these events, has become a key expectation of any power supply designer.

And that's before bringing in to the mix electron tunneling which for CPUs is a major factor and is one of the big challenges in node shrinks.

Link to post
Share on other sites
23 minutes ago, Stefan Payne said:

I agree with you.

But the Chiplets is a necessity du to physical limitations of lower manufacturing processes. And also the financial ones.

It just makes more sense to throw anything out of the 7nm die for various reasons.

The things you mentioned here is one of the reasons to do it, another is cost. Its just cheaper to do it because according to bits and chips the yield of the 7nm process with AMD is at only around 70%.

I'm not sure why but I think you're taking what I said as counter points or reasons not to do it? It should have been obvious those were all positives, reasons to do, but I was just more talking about the situation itself.

Link to post
Share on other sites
52 minutes ago, leadeater said:

No it doesn't, but it's hard to generalize like this at all because transistor type and makeup changes this too.

 

Rough, not CPU transistor example:

Figure-1.png

Ähm, isn't that a logarythmic diagram??

 

 

52 minutes ago, leadeater said:

Just like to correct this because it's something you shouldn't carry forward, it's not correct. Doubling the voltage doesn't double the current, you can change the voltage and the current doesn't have to change at all but power must, or current has to drop or resistance increase.

It is correct as I was talking about Resistive Loads.

Aka a Resistor.

 

IF you double the Voltage on the Resistor, you double the current as well.


You can also calculate it.

Just take a 100 Ohm Resistor.

 

Voltage = Resistance * Current

 

Voltage

------------ =  Current

Resistance 

 

So for a 100 Ohm Resistor, at 12V we are talking about 0,12A or 1.44W

So for a 100 Ohm Resistor at 6V we are talking about 0,06A or 0,36W

 

I know that with silicon semiconductors its not as easy and that there are also other effects that influence it for example the thickness of the insulation layer.

But in General, those aren't Ohm'sch Resistors, they are mostly capacitors with a bit of inductors so its not as cut and dry as with a Resistor.


"Hell is full of good meanings, but Heaven is full of good works"

Link to post
Share on other sites
13 minutes ago, Stefan Payne said:

It is correct as I was talking about Resistive Loads.

Aka a Resistor.

Yea sorry I removed it just before you posted this as the situations I was going to reference honestly didn't matter at all, nor that related to what you were saying.

 

13 minutes ago, Stefan Payne said:

Ähm, isn't that a logarythmic diagram??

No, exponential. Point is you said it's linear and it's not, and that's only one factor for overall power for a CPU with billions of transistors. If switching loss is not linear then it's impossible for power draw for a CPU in relation to frequency to be linear.

Link to post
Share on other sites
40 minutes ago, leadeater said:

I'm not sure why but I think you're taking what I said as counter points or reasons not to do it? It should have been obvious those were all positives, reasons to do, but I was just more talking about the situation itself.

No, absoluitely not. I just added some additional arguments to go with the chiplet/MCM Approach instead of the Monolithic die.

 

I was just talking about a different side of the chiplet design, that allows other positive effects. Though I was more arguing on the financial aspect and the size of the I/O Die.

We both agree that the chiplet design makes the most sense with 7nm, just for different reasons.

 

I didn't  really think about the Power Consumption and design aspect.

That are very interesting points you made there.

 

 

 


"Hell is full of good meanings, but Heaven is full of good works"

Link to post
Share on other sites
8 minutes ago, Stefan Payne said:

I didn't  really think about the Power Consumption and design aspect.

That are very interesting points you made there.

You may have not made the connection but why I said TR and EPYC are the more interesting ones is because I think with the chiplet design and potential power/voltage controls I think we might see the frequency penalties of high core CPUs all but disappear. Think about a situation on a 64 core EPYC where only 8 cores are active and the software/scheduler is properly optimized so only 1 chiplet is active/in high power state, is there any reason why the clocks would have to be any lower than on say a single chiplet Ryzen 3000? I'd say no they should be able to be the same, which is awesome if possible.

Link to post
Share on other sites
21 minutes ago, wasab said:

AMD has a process lead over Intel in the first time ever 

Yeah and it also coincides with a period in which

-AMD has a superior more scalable architecture, 

-Intel cannot produce enough to meet demand which means some vendors have no choice but to turn to AMD.

If it had happened during bulldozer era it would not have mattered. This time it's as if AMD wrote the script for this LOL, they could not have asked for a better combination of circumstances.

 

This period will not last forever, so in the next two years AMD must make as much money as possible and also secure their future design wins with R&D and secure all their channel partnerships with industry partners so that when Intel does make a comeback they don't lost share. They have to avoid what happened when Intel came back after the Athlon64 era.

Link to post
Share on other sites
2 hours ago, Humbug said:

This period will not last forever, so in the next two years AMD must make as much money as possible and also secure their future

Its just a shame they don't have the same level of confidence from network admins as Intel.  If companies were confident they could upgrade fleet systems and servers to AMD with little risk of down time they'd be able to capitalise a lot more.


QuicK and DirtY. Read the CoC it's like a guide on how not to be moron.  Also I don't have an issue with the VS series.

Link to post
Share on other sites
9 hours ago, wasab said:

AMD has a process lead over Intel in the first time ever 

Not entirely true.

IIRC there was a small lead back in the 90s, for example the copper interconnect stuff.

 

This time its not a small advantage that Intel can counter tomorrow. Its a big one.

And Intel might not be able to keep up.


"Hell is full of good meanings, but Heaven is full of good works"

Link to post
Share on other sites
9 hours ago, Humbug said:

Yeah and it also coincides with a period in which

-AMD has a superior more scalable architecture, 

-Intel cannot produce enough to meet demand which means some vendors have no choice but to turn to AMD.

If it had happened during bulldozer era it would not have mattered. This time it's as if AMD wrote the script for this LOL, they could not have asked for a better combination of circumstances.

That sounds like AMD in the mid 90s...

 

Quote

This period will not last forever

I'm not sure about that.

It all depends on TSMC and how good their processes are and how much Intel can keep up.

Also how much Performance is lying dormant in the Zen Architecture.

There seems to be a lot of things you can do to improve the Performance.

For example the latency is rather huge, maybe work a bit on the Cache architecture, though that's gotten better with Zen. It was a desaster on Bulldozer.

Quote

 so in the next two years AMD must make as much money as possible and also secure their future design wins with R&D and secure all their channel partnerships with industry partners so that when Intel does make a comeback they don't lost share. They have to avoid what happened when Intel came back after the Athlon64 era.

AMD doesn't need to do the same bullshit they did when K8 was launched.

Basiically they didn't start improving on K8 and only slightly improved it. That was their mistake.

If they worked hard on the K8 design and dramatically improved it, including a 4 Core Version, it would have been better.

 

The mistake was the AM2 K8. That should never have happened. They should have done something better than that...

Then they could potentially have kept up with Core 2.

 

If you want to, you can call the Phenom an improved K8 (wich in turn you can call an improved K7)...

 

But right now it doesn't look like AMD does the same mistake as with the K8 back in the day, when they didn't do anything to improve the Chips Performance or as you call it today "IPC".


"Hell is full of good meanings, but Heaven is full of good works"

Link to post
Share on other sites
1 hour ago, Stefan Payne said:

I'm not sure about that.

It all depends on TSMC and how good their processes are and how much Intel can keep up.

I mean that right now AMD has both the process advantage via TSMC and they also have superior technology which allows them to make big powerful high core count CPUs using infinity fabric interconnects much cheaper and with better yields than Intel can. Right now Intel has no answer, they just have no way to make 16+ core CPUs as easily as AMD can because they don't have a technology to overcome poor yields.

 

But over the years Intel will develop that ability too, it may take some time but it will happen. And it terms of fabrication both companies will start approaching the limits of silicon.

 

I think what matter is that AMD should secure the right industry partnerships. They need to be very close with to all the big data centers in the world, to all the laptop vendors and various retail channels having good supply, contracts and understandings with all these guys so that everyone in the channel views them as a very reliable partner. So that when Intel comes back with better CPUs they can leverage that to keep Intel at bay, obviously in addition to launching better future products...

1 hour ago, Stefan Payne said:

The mistake was the AM2 K8. That should never have happened. They should have done something better than that...

Then they could potentially have kept up with Core 2.

 

If you want to, you can call the Phenom an improved K8 (wich in turn you can call an improved K7)...

 

But right now it doesn't look like AMD does the same mistake as with the K8 back in the day, when they didn't do anything to improve the Chips Performance or as you call it today "IPC".

Ya you are right- it seems to be a good thing that Lisa Su has done is that she has like 2 or 3 CPU architecture teams working at the same time and they leapfrog each other with Zen 2, Zen 3 etc... So far they are executing well on that.

 

I remember when AMD launched the bulldozer fx-8150. It was such a big disappointment. But then a year later they launched the improved piledriver fx-8350. It still didn't beat Intel but at the time I was thinking this is a good improvement in under a year- it had slightly better IPC, clocked slightly higher and was a bit more efficient. I thought that if AMD keeps up by releasing new stuff like this every year on the desktop they could close the gap with Intel.

 

But then they just stopped, after piledriver there was just nothing in terms of high performance efforts... It seems like they just gave up and said 'this is not gonna work, bulldozer uarch is a turd so lets put all our eggs  into project Zen R&D and just survive for a few years until then'.

Link to post
Share on other sites
9 hours ago, Humbug said:

But over the years Intel will develop that ability too, it may take some time but it will happen. And it terms of fabrication both companies will start approaching the limits of silicon.

 

I think what matter is that AMD should secure the right industry partnerships.

This is it right here, they have to make hay while the sun shines (as do all companies), And while the stars might be lining up for them, the biggest reason they have to capitalise fast is because Intel have shown several times now that they don't need a Die shrink to be competitive.   Even when they were caught out with the initial Ryzen release it didn't take them long too have an answer, then when the zen+ hit the shelves Intel answered again with a another 14++++nm  response.  It is mostly speculation that Intel won't be able to keep up with Ryzen 2 so if AMD do have a lead for any length of time they need to pound the market hard.

 

 


QuicK and DirtY. Read the CoC it's like a guide on how not to be moron.  Also I don't have an issue with the VS series.

Link to post
Share on other sites
1 hour ago, mr moose said:

This is it right here, they have to make hay while the sun shines (as do all companies), And while the stars might be lining up for them, the biggest reason they have to capitalise fast is because Intel have shown several times now that they don't need a Die shrink to be competitive.   Even when they were caught out with the initial Ryzen release it didn't take them long too have an answer, then when the zen+ hit the shelves Intel answered again with a another 14++++nm  response.  It is mostly speculation that Intel won't be able to keep up with Ryzen 2 so if AMD do have a lead for any length of time they need to pound the market hard.

 

 

Agreed amd needs to jump far ahead with ryzen 2/3000

So they can charge more per chip

They still need to pay the middle man though and under cut their silicon overhead just to compete

Hopefully they are dabbling in quantum which we already know Intel is

Amd needs focus on branding their name even on consoles

 

They have the potential and cult following but seems like they continuously miss the means on branding

Itd prolly make investors happy too

Which is a big deal

 

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now


×