Next gen Zen 2/3 "Starship" and derivatives

[Insert_Nickname]

That's a straight-forward roadmap for the CPUs. But what will happen for the APUs?

2x CCX + IGP or 1x CCX + larger IGP. Perhaps both. That is the question. I'm leaning toward the first, since Intel already has 6 core CPUs in their mainstream lineup. But then AMDs 4C/8T RR holds up pretty well against the non-HT Intel 6 cores. So it could go both ways.

A hypothetical 8C/16T with an 768SP IGP APU would make a nice chip. Though it would be handicapped by memory bandwidth.

 

[Vattila]

Regarding frequencies, here is my reconstructed version of the Power vs Performance chart for GlobalFoundries' 7LP process (see this and this thread). I have also included an equivalent chart with absolute base frequency and power set to 3.0 GHz and 65 W, respectively. This is the base frequency and TDP of Ryzen 1700. It looks like this calibration gives sensible numbers for the chart. E.g. the penultimate data point for 14LPP lands at 3.6 GHz and 99 W, which is the base frequency and ~TDP for Ryzen 1800X. The last data point lands at 120 W power and 3.8 GHz base frequency, which sounds reasonable.

For 7nm SoC, a hypothetical 8-core Ryzen 3700 lands at 59 W power and 4.2 GHz base frequency. The ultimate data point for 7nm SoC lands at 98 W and 5.0 GHz, which sounds plausible for a hypothetical 8-core Ryzen 3800X.

The 7nm HPC curve is just crazy — perhaps only destined for IBM Power. However, we must take into account that my absolute power chart shows power for 8 cores running flat out. So, you could imagine extending the 7nm HPC curve downwards and make an 8-core Ryzen at 4.5 GHz base frequency and ~95 W power (whereas a 7nm SoC equivalent would consume ~75 W at 4.5 GHz). The HPC version would be able to boost frequency on up to 3 cores to 6.2 GHz while staying within a 95 W TDP, since 244 W (at 6.2 GHz on my 7nm HPC curve) divided by 8 cores is ~30 W per core.

In conclusion, the 7LP process provides lofty headroom for clock frequency. 7LP may hence introduce a much bigger difference between base and boost frequencies, especially with rising core counts, as the limiting factor will be power.

 

[whm1974]

2x CCX + IGP or 1x CCX + larger IGP. Perhaps both. That is the question. I'm leaning toward the first, since Intel already has 6 core CPUs in their mainstream lineup. But then AMDs 4C/8T RR holds up pretty well against the non-HT Intel 6 cores. So it could go both ways.

A hypothetical 8C/16T with an 768SP IGP APU would make a nice chip. Though it would be handicapped by memory bandwidth.

If AMD used some sort of on package memory that would get around these pesky memory issues.

 

[3DVagabond]

2x CCX + IGP or 1x CCX + larger IGP. Perhaps both. That is the question. I'm leaning toward the first, since Intel already has 6 core CPUs in their mainstream lineup. But then AMDs 4C/8T RR holds up pretty well against the non-HT Intel 6 cores. So it could go both ways.

A hypothetical 8C/16T with an 768SP IGP APU would make a nice chip. Though it would be handicapped by memory bandwidth.

We used to hear lots of complaining about Intel including iGPU on the chips people wanted to buy and use dGPU with. Paying for an iGPU that they didn't need. I wouldn't go any higher than 4c/8t on the APU's until they can put one with it (iGPU) with it that matches the performance of the CPU. Today's performance level would be ~RX 570'ish.

 

[Insert_Nickname]

If AMD used some sort of on package memory that would get around these pesky memory issues.

AMD could do something like Kaby-G with a Zeppelin die and separate GPU w/ HBM connected via the GMI links. Unfortunately, that'd likely require either a new socket, or BGA packaging. Thermals would be high, but I doubt higher then TR, which is relatively easy to handle with proper coolers.

But the market for something like that is properbly too small. Maybe we'll see something like it in the next gen consoles?

We used to hear lots of complaining about Intel including iGPU on the chips people wanted to buy and use dGPU with. Paying for an iGPU that they didn't need. I wouldn't go any higher than 4c/8t on the APU's until they can put one with it (iGPU) with it that matches the performance of the CPU. Today's performance level would be ~RX 570'ish.

Well, AMD will still have Pinnacle Ridge for the discrete card market.

Ironically, Raven Ridge and Intels mainstream lineup suffer from the exact situation. Both are essentially mobile first products shoehorned into desktop duty. I still think there is room for a true desktop class APU in the market. The question is whether that market is large enough* to bother with from a cost perspective.

*excluding consoles of course...

 

[whm1974]

AMD could do something like Kaby-G with a Zeppelin die and separate GPU w/ HBM connected via the GMI links. Unfortunately, that'd likely require either a new socket, or BGA packaging. Thermals would be high, but I doubt higher then TR, which is relatively easy to handle with proper coolers.

But the market for something like that is properbly too small. Maybe we'll see something like it in the next gen consoles?

I'm just talking abut the iGPU using on package memory. Like Intel did with Iris Pro.

 

[LightningZ71]

It does make you wonder what they could do at 7nm. Could they keep the general layout of the APU constant for a 7nm shrink, meaning that the die area used is around 30-35% of what it was at 14nm, then use the remaining 65-70% of the original 14nm die area to fill it with on chip memory and keep it all in the same exact package? That could give you an arrangement like Iris Pro or even bigger L4 caches. At 7nm, it is not out of the realm of possibility that they could be able to fit 512 MB or more into a die that can fit the same external package size. Just maintaining the same number of VEGA CUs in the iGPU section, but giving it that much local memory at high speed would make for a significant performance uplift. They could continue to use the HBCC tech to manage overflow memory to the system DRAM.

Obviously, compared to just shrinking the existing APU designs and performing any version based core and uncore updates, those chips would be much more expensive from a fabrication cost perspective (though, the total product cost difference shouldn't be a whole lot different). But, considering the performance uplift that could be expected, especially for highly integrated mobile products, it might be worth looking in to.

 

[Excessi0n]

Regarding frequencies, here is my reconstructed version of the Power vs Performance chart for GlobalFoundries' 7LP process (see this and this thread). I have also included an equivalent chart with absolute base frequency and power set to 3.0 GHz and 65 W, respectively. This is the base frequency and TDP of Ryzen 1700. It looks like this calibration gives sensible numbers for the chart. E.g. the penultimate data point for 14LPP lands at 3.6 GHz and 99 W, which is the base frequency and ~TDP for Ryzen 1800X. The last data point lands at 120 W power and 3.8 GHz base frequency, which sounds reasonable.

The problem with that mapping of relative values to absolute values is that it doesn't really match up with GloFo's marketing. If they actually thought they could get production chips past 6 GHz then they would be shouting it to the world because it would be a legitimately amazing achievement. Instead, they've only claimed that 7nm is good for operation past 5 GHz. As I've posted about before, I'm inclined to believe that the 14nm curve on that chart doesn't include the highest frequencies attainable with Ryzen since they are well past the point of diminishing returns.

Honestly though, that chart is pretty useless for this sort of technical speculation. The conclusions that one draws from it vary substantially depending on what frequency one thinks the 14nm curve stops at. And, since one can make good arguments for that point being anywhere in a range of values several hundred megahertz wide, that means that any conclusions being made are basically worthless.

 

[IntelUser2000]

If they actually thought they could get production chips past 6 GHz then they would be shouting it to the world because it would be a legitimately amazing achievement.

The highest overclock on air is at 5.7GHz. Above that, you need exotic cooling solutions, and the majority end up at 5.5GHz or less on air. That's usually using tricks as well. Like having only 1 core active, or Hyperthreading off, or running only the monitoring application.

The highest overclock ever was achieved by AMD FX and is at 8.79GHz frequency. Remember FX chips are based on the architecture with very long pipelines. Prescott Celeron chips are also among the top max overclock chips.

Interestingly, Sandy Bridge still seems to be the top core for highest air overclock. While Sandy Bridge can do it with 4 cores active, the similar clocks achieved by Broadwell-E is only done with 1 core.

Again, those are not stable overclocks. Stable overclocks end up at 5.2GHz or so. The amount of effort required above that point seems to indicate a physics limit, because it applies across ISAs, fabs, uarchs, and different process/lithography.

IBM's ZEC12 is the highest clocked commercial chip at 5.5GHz, but requires 300W to do so. Certain configurations on that server calls for liquid cooling, and IBM has very impressive packaging technologies for their POWER line, so for average folks, it may be just as exotic as using LN2. Intel chips like Kabylake and Coffeelake merely reduces the gap between max OC clock and stock clocks.

In short: 6GHz was never doable.

 

[Phynaz]

Honestly though, that chart is pretty useless for this sort of technical speculation.

It may be useless because he made it himself.

 

[itsmydamnation]

In short: 6GHz was never doable.

Your right and your probably wrong, a power 8/zen style core could probably reach 6ghz but neither of those are going to be on 7HPC. Power 9 (i assume onwards as well) has a significantly shorter pipeline length so even its clock will probably be in the mid to high 4 ghz.

So no one said anything about air, only what would be packaged, you also dont see GF going on about 14nm HP yet Z14 is still in the 5.2ghz range.

 

[Vattila]

Honestly though, that chart is pretty useless for this sort of technical speculation. The conclusions that one draws from it vary substantially depending on what frequency one thinks the 14nm curve stops at.

You are welcome to try out my spreadsheet yourself and experiment with the calibration of the chart. If you move much above or below 3.0 GHz — just a few hundred MHz, as you say — the chart stops making sense. With a base frequency above 3.0 GHz, the HPC curve quickly extends towards 7 GHz, and with a base frequency reduced to just 2.7 GHz, it gets difficult to extend the 14LPP curve to even 4.0 GHz without getting crazy power — it goes vertical pretty fast.

I think a 3.0 GHz base makes a lot of sense. With this calibration the SoC and HPC curves fall nicely around 5 GHz, which is the target frequency for 7LP as advertised by GlobalFoundries' 7LP product brief. Incidentally, the 14LPP product brief stated ">3 GHz" performance, so a calibration point set at 3.0 GHz makes sense in that regard too, as the relative chart is a comparison to 14LPP.

In any case, a little adjustment of the chart calibration doesn't change my conclusion; 7LP is going to be a pretty big advance, provided the process lives up to the chart.

 

[raghu78]

You are welcome to try out my spreadsheet yourself and experiment with the calibration of the chart. If you move much above or below 3.0 GHz — just a few hundred MHz, as you say — the chart stops making sense. With a base frequency above 3.0 GHz, the HPC curve quickly extends towards 7 GHz, and with a base frequency reduced to just 2.7 GHz, there is no way to extend the 14LPP curve to even 4.0 GHz without getting crazy power — it goes vertical pretty fast.

I think a 3.0 GHz base makes a lot of sense. With this calibration the SoC and HPC curves fall nicely around 5 GHz, which is the target frequency for 7LP as advertised by GlobalFoundries' 7LP product brief. Incidentally, the 14LPP product brief stated ">3 GHz" performance, so a calibration point set at 3.0 GHz makes sense in that regard too, as the relative chart is a comparison to 14LPP.

In any case, a little adjustment of the chart calibration doesn't change my conclusion; 7LP is going to be a pretty big advance, provided the process lives up to the chart.

I have done my own analysis on the GF 7HPC and GF 7SoC curves and also correlated with GF statements on target speeds for 7HPC of 5 Ghz and 7SoC of 3.5 Ghz. So working backwards

https://www.globalfoundries.com/sit...s/product-brief-7lp-7nm-finfet-technology.pdf

High Performance - 5GHz operation Server, Data Center, ASICs

1.4x = 3.5 Ghz (7SoC)
1x = 2.5 Ghz (14LPP)
2x = 5 Ghz (7HPC)

These speeds are very realistic and the above calculation also agrees with the sweet spot for 14LPP on voltage freq curve which is around 2.5 Ghz. Incidentally both the EPYC and Ryzen 2700u have an all core boost around 2.6 Ghz.

 

[Vattila]

1x = 2.5 Ghz (14LPP)

With a base frequency of 2.5 GHz, what base power do you suggest to calibrate the chart, assuming we are calibrating for a 8-core Ryzen chip? It seems we have to go very low, 35 W or less, to make the curve fit the Ryzen 1800X (3.6 GHz, 95 W).

 

[Hitman928]

2.5 GHz with 35-40 W power seems right on target to me.

 

[Vattila]

2.5 GHz with 35-40 W power seems right on target to me.

40 W base is too high to fit 1800X (3.6 GHz, 95 W) on the curve. 35 W could fit. With this calibration, AMD could make a 4.8 GHz 8-core Ryzen at 96 W, using the HPC version of the 7LP process.

 

[Gideon]

40 W base is too high to fit 1800X (3.6 GHz, 95 W) on the curve. 35 W could fit. With this calibration, AMD could make a 4.8 GHz 8-core Ryzen at 96 W, using the HPC version of the 7LP process.

Looking at this chart, I also really wish them to make a Raven Ridge successor with Zen2 cores on the 7nm SoC process

 

[Vattila]

Looking at this chart, I also really wish them to make a Raven Ridge successor with Zen2 cores on the 7nm SoC process

Sure. The lowest point on the SoC curve is 2.6 GHz at 14 W, for an 8-core Ryzen, with a boost frequency up to 4.2 GHz on this graph. Max boost for Ryzen Mobile 2500U is 3.6 GHz, with a 2.0 GHz base and 15 W TDP. An 8-core successor to Raven Ridge looks to be the obvious next step for the APUs, as AMD needs to offer something against Intel's 6-core. Design-wise it seems straight-forward; just use two quad-core CCXs and a Vega/Navi block. I would be surprised if it was not in the works already.

 

[PeterScott]

What is the source of that chart?

40 W base is too high to fit 1800X (3.6 GHz, 95 W) on the curve. 35 W could fit. With this calibration, AMD could make a 4.8 GHz 8-core Ryzen at 96 W, using the HPC version of the 7LP process.

 

[Vattila]

What is the source of that chart?

See: https://forums.anandtech.com/thread...-and-derivatives.2511914/page-2#post-39313608

 

[PeterScott]

See: https://forums.anandtech.com/thread...-and-derivatives.2511914/page-2#post-39313608

So made up by someone in the forum. I now see why it looks so optimistic.

 

[Hitman928]

So made up by someone in the forum. I now see why it looks so optimistic.

The scaling factor plot comes from Global Foundries. Vatilla is just trying to fit it to the (somewhat) known curve for Ryzen.

Vatilla, I think you're trying to be super precise based upon something that wasn't intended for that. I think you're graph looks good in a general sense, but I wouldn't put too much stock into it's accuracy, trying to get everything lined up precisely.

 

[PeterScott]

The scaling factor plot comes from Global Foundries. Vatilla is just trying to fit it to the (somewhat) known curve for Ryzen.

Vatilla, I think you're trying to be super precise based upon something that wasn't intended for that. I think you're graph looks good in a general sense, but I wouldn't put too much stock into it's accuracy, trying to get everything lined up precisely.

There is a lot more to achieving high clock speed than process alone.

Remember when AMD said they threw something like a Billion extra transistor at Vega to get clock speed up?

I seriously doubt we will see 5+GHz Ryzens on 7nm. That seems more like wishful thinking at this point.

 

[Vattila]

Let's play the pessimism game. Assuming Raghu78's calibration of the chart at 2.5 GHz and 35 W, let's say the process frequency gain is pretty much eaten up by the 7nm redesign, i.e. we shift the 7LP curves left so they almost overlap 14LPP. How does it look now? Still looks pretty good: A (poorly designed) 96 W 8-core 7LP Ryzen chip would run at 3.9 GHz base, while boosting to 4.3 GHz on ~6 cores.

 

[PeterScott]

Let's play the pessimism game. Assuming Raghu78's calibration of the chart at 2.5 GHz and 35 W, let's say the process frequency gain is pretty much eaten up by the 7nm redesign, i.e. we shift the 7LP curves left so they almost overlap 14LPP. How does it look now? Still looks pretty good: A (poorly designed) 96 W 8-core 7LP Ryzen chip would run at 3.9 GHz base, while boosting to 4.3 GHz on ~6 cores — and beyond on fewer cores, if we are allowed to extend the graph towards 5 GHz, as promised by GlobalFoundries.

We will see eventually.

Getting clock-speed up on tighter packed transistors tends to create issues, you end up with a lot more heat/area on the smaller die creating cooling issues, and it looks like Intel made some features larger on 14nm++ to improve higher clock speeds.

My bet is 7nm SoC will likely be less problematic in practice than 7nm HPC. I wouldn't be surprised if HPC ends up delayed.