Speculation: AMD's response to Intel's 8-core i9-9900K

[coercitiv]

I have 8700K at home, but probably going to buy 9900K, because it has 16MB of L3 and I hope will provided 300-400Mhz more over my current 4.8Ghz clock. Yeah, that is how much i value desktop performance, and that is knowing that run compiles, JVMs and sometimes multiple linux/android VMs.

When I bought the 8700 instead of 8700K that's exactly what I was hoping for, a later 8c/16t as a long term upgrade for my desktop while the lower clocked 8700 would end up in the Plex server. I had to be cautiously optimistic about platform support though, as Intel managed to make a real mess with the Z170/270 -> Z370 transition (very poor communication). Fortunately that is no longer a problem, and the rumored solder is just the cherry on top.

That having been said, there's going to be a price to be paid in these forum discussions if we abandon comparison standards. I don't care about the actual value of power consumption, whether it's 65W, 95W, 150-250W - they're all valid use cases, what I care about is using standards in order to replicate benchmarks and be able to make apples to apples comparisons.

We're already starting to see the signs, having reached a point where CPU reviews are effectively Motherboard & CPU reviews, with the same CPU boosting to different clocks depending on the test motherboard (Intel had the MCE funtimes, latest AMD boards are exhibiting similar behavior ), and it's only gonna get worse - we'll have to argue for a couple of pages just to establish stock clocks.

We probably have to agree to disagree on this topic, so I'll give it a rest.

 

[JoeRambo]

We're already starting to see the signs, having reached a point where CPU reviews are effectively Motherboard & CPU reviews, with the same CPU boosting to different clocks depending on the test motherboard (Intel had the MCE funtimes, latest AMD boards are exhibiting similar behavior ), and it's only gonna get worse - we'll have to argue for a couple of pages just to establish stock clocks.

This is exactly why I trust Stilt's results more than random guy's from internet site who runs Stilt's 3466 memory profiles Good documented runs, using sensible settings and good selection of software AND what is more important investigating checking and reporting if hw actually sticks to those limits.

CPU variability, ambient temperature control, cooling all have obvious impact on those "turbo" CPUs from both vendors. Going deeper - it matters when running memory on XMP, if it happens to have an optimized profile for that kit timings or not. There can be a huge gap on exact same primaries if secondaries and tertiarys are very relaxed auto. Of course this is most important on "stock" settings, where bw/latency are in biggest demand.

 

[epsilon84]

Nope, it was also amazon no. 1. CPU sold because of that

Is it still that price? Pretty sure it's no longer available, at least it wasn't when I looked, unless I'm missing something. Otherwise I would buy a few and flip them on EBay for a handsome profit! I bet quite a few people did that, a 1920X is worth way more than $249.

 

[PeterScott]

Is it still that price? Pretty sure it's no longer available, at least it wasn't when I looked, unless I'm missing something. Otherwise I would buy a few and flip them on EBay for a handsome profit! I bet quite a few people did that, a 1920X is worth way more than $249.

It was definitely a mistake, taken down the moment they noticed it. It will be interesting to see if it gets honored as it will definitely be sold at a loss.

 

[PeterScott]

I'm just merely speaking from what I believe is technically feasible, 4.7GHz is definitetly in the upper ranges of that, but still plausible.

I do see where your scepticism is coming from though, even though I believe the soldered IHS would be a great idea especially on a chip like the 9900K. Let's hope the rumours are true.

Another leak/rumor on the 9700k/9900k clock speed. I am still skeptical, but all the leaks/rumors keep pointing in the same direction:
https://videocardz.com/77188/intel-core-i9-9900k-will-boost-up-to-4-7-ghz-with-all-eight-cores

 

[mikk]

This is a confirmation from the first coolaler leak. Insane clock speeds from those 8 core SKUs. 9900k will be the new mainstream king in Q4.

 

[maddie]

Are people still arguing the extent of the supremacy of the incoming i9 models?

The only thing capable of dethroning this will be if AMD goes for clocks on the GloFlo 7nm process in addition to the increased IPC with Zen2. We're in the pendulum days.

 

[epsilon84]

Are people still arguing the extent of the supremacy of the incoming i9 models?

The only thing capable of dethroning this will be if AMD goes for clocks on the GloFlo 7nm process in addition to the increased IPC with Zen2. We're in the pendulum days.

In ST or gaming or even typical desktop workloads, you're right.
You forgot about moar cores though, AMD is rumoured to go wider as well as faster, pretty sure a 12c or 16c(!!) Zen2 would own this in a heavy MT workload, though as I said previously where do you draw the line on how many cores is 'too many' on a desktop CPU?

 

[IEC]

Damn $249 for a 1920x?! Has to be a mistake.

I thought so and ordered one anyways before they stopped backorders. It might take them a while to ship but they appear to be honoring it.

 

[maddie]

In ST or gaming or even typical desktop workloads, you're right.
You forgot about moar cores though, AMD is rumoured to go wider as well as faster, pretty sure a 12c or 16c(!!) Zen2 would own this in a heavy MT workload, though as I said previously where do you draw the line on how many cores is 'too many' on a desktop CPU?

I've come to the belief, and stated in another thread, that AMD will stay with an 8C basic unit for the next 7nm iteration of Zen. Simply put, the high early cost of the process and relatively low yields. I do however, believe a speed demon is coming with higher IPC to take the absolute desktop crown, both ST and MT with a GloFlo die. We will have to remember that this will probably be the lead design for 7nm in that fab. This is worst than TSMC, who have already claimed HVP for early7nm.

The entire "how many cores is desktop, HEDT, etc" is a marketing and sales fabrication. To put simply.

How many cores for you? As many as you want or need.

 

[PeterScott]

In ST or gaming or even typical desktop workloads, you're right.
You forgot about moar cores though, AMD is rumoured to go wider as well as faster, pretty sure a 12c or 16c(!!) Zen2 would own this in a heavy MT workload, though as I said previously where do you draw the line on how many cores is 'too many' on a desktop CPU?

You made a pretty reasonable case why 12C TR (1920x) wouldn't necessarily own a 9900K, so I am not so sure that 12C Zen2 would, unless it is gets 12c and higher clock speed together.

Definitely interesting times for CPU watchers in the next 2-9 months.

If 9900K really is soldered with an 8 core Turbo of 4.7GHz, it will be a beast of a processor at nearly any workload thrown at it.

Then we will be watching to see which Zen2 rumors come true, as it's another potential beast of a processor.

I do think 8C will be the sweet spot on premium desktop for many years to come, much like 4C was for nearly a decade.

 

[eek2121]

Now that it is quite clear Intel is dragging their feet, I'm sure AMD will let Intel have their brief moment in the spotlight, and then unleash 7nm fury (no not THAT fury...though that'd be an interesting thing to do) in April of next year. I guess there is still the possibility of a 2800X.

It's too bad AMD had to split their platform up into 3 different sockets. 1 socket for desktop/HEDT and 1 for server would have helped them in this scenario. Maybe they will rectify this when the AM4 socket gets refreshed. The thought of being able to select a processor from 8-64 cores at X clock speed is an appealing one, and giving AM5 users access to quad channel memory would put another nail in Intel's coffin*.

* Note: Not an AMD fanboy, but Intel has been screwing up consistently as of late. I actually only have 1 AMD system and 3 Intel systems.

 

[PeterScott]

I've come to the belief, and stated in another thread, that AMD will stay with an 8C basic unit for the next 7nm iteration of Zen. Simply put, the high early cost of the process and relatively low yields.

A reasonable assessment.

I also think 8C is more (Twice as) likely than 12C on Zen2 for the same reasons you state. Maybe 1% chance of 16C, because it isn't impossible, but VERY unlikely.

I do however, believe a speed demon is coming with higher IPC to take the absolute desktop crown, both ST and MT with a GloFlo die. We will have to remember that this will probably be the lead design for 7nm in that fab. This is worst than TSMC, who have already claimed HVP for early7nm.

Not sure I agree on the ST performance. I am betting 5GHz will be unattainable on any next gen process in 2019/2020. Not 7nm(TSMC/GF), and not Intel 10nm.

If you look at Intels ultra refined 14nm++, IIRC the last iteration that lead to higher clock speed, increased some features sizes. Shrinking doesn't necessarily get you to those super high clock speeds. Remember the GF/TSMC stall out at 20nm. That shrink lead to a mess. It was FinFet that saved the day, not a shrink. I doubt there is another jump like Finfet.

I also think Intel's super high 14nm++ clocks will haunt them for quite some time as 10nm may not get there. I am betting Intel consumer parts for Q4 2019, will be laptop parts, desktop will stick with 14nm. Time will tell.

 

[maddie]

A reasonable assessment.

I also think 8C is more (Twice as) likely than 12C on Zen2 for the same reasons you state. Maybe 1% chance of 16C, because it isn't impossible, but VERY unlikely.



Not sure I agree on the ST performance. I am betting 5GHz will be unattainable on any next gen process in 2019/2020. Not 7nm(TSMC/GF), and not Intel 10nm.

If you look at Intels ultra refined 14nm++, IIRC the last iteration that lead to higher clock speed, increased some features sizes. Shrinking doesn't necessarily get you to those super high clock speeds. Remember the GF/TSMC stall out at 20nm. That shrink lead to a mess. It was FinFet that saved the day, not a shrink. I doubt there is another jump like Finfet.

I also think Intel's super high 14nm++ clocks will haunt them for quite some time as 10nm may not get there. I am betting Intel consumer parts for Q4 2019, will be laptop parts, desktop will stick with 14nm. Time will tell.

In the past , I would have agreed fully with the speed issue, but it seems IBM personnel are now fully integrated into GloFlo and they certainly know GHz.

Interesting and good point on the maxed 14nm +++++++++++ complicating their future sales & marketing.


edit:
Just read the GloFlo marketing PDF on 7LP 7nm FinFET. They not only say ">40% performance improvement at iso power (vs. 14nm)", which can leave room for doubting the max clocks, but they also say outright "5GHz operation Server, Data Center, ASICs".

 

[IndyColtsFan]

I thought so and ordered one anyways before they stopped backorders. It might take them a while to ship but they appear to be honoring it.

Too bad I missed that sale or I would've bought one too.

 

[Dayman1225]

Just read the GloFlo marketing PDF

Exactly the problem, it's marketing.

They not only say ">40% performance improvement at iso power (vs. 14nm

FYI thats compared to 14nm LPE, not LPP so take that as you will.

 

[maddie]

Exactly the problem, it's marketing.

FYI thats compared to 14nm LPE, not LPP so take that as you will.

Marketing, yes, but they do also claim 5GHz outright. That cannot be weaseled out of.

 

[PeterScott]

Marketing, yes, but they do also claim 5GHz outright. That cannot be weaseled out of.

Still remains to be seen.

Also saying a process is capable of 5GHz, doesn't mean every design can hit that. You actually have to design parts that will clock that high. Remember when AMD was asked what all the extra transistors were doing on Vega. They said a huge number of transistors were to enable higher clock speeds.

I'll be surprised if ZEN2 hits 5GHz. Though it would be a pleasant surprise. Intel would definitely have a lot more to worry about if that happens.

 

[ub4ty]

Are people still arguing the extent of the supremacy of the incoming i9 models?

The only thing capable of dethroning this will be if AMD goes for clocks on the GloFlo 7nm process in addition to the increased IPC with Zen2. We're in the pendulum days.

There's no doubt that it's supreme or that Intel in general has had supreme performance. The question becomes, now that there is adequate competition, which there wasn't in the 2/4 core blood letting phase of intel's reign, at what price is this supreme performance. If the price differential weren't so big, it wouldn't be a question as to whether or not I went w/ Intel. I have gone with them for many years because they had best performance. However, DDR4 prices are insane and its a new paradigm where 1st class storage also matters. My NVME drive costs more than my processor and has a quad core processor on it. My GPU cost more than my processor and is 100s of times faster at particular workloads. The world no longer revolves around the processor per say. It is at such a point that I question paying double the price of an AMD processor for 20-30% performance improvement. I've got other components to think about. The funny thing is, I paid about $400/$500 for a 4 core years ago so these prices of the 8700k/9900k aren't foreign. I just don't see the value in it. And that's the deciding factor for me as an intelligent reasoned consumer... VAUE

On performance and supremacy, most definitely intel dominates 6 and now 8 core. It doesn't dominate performance on HEDT and it doesn't dominate enterprise any longer and its pricing only serves to dominate my wallet. So, i went w/ AMD for 8 core and for HEDT and when I go to enterprise hardware it will be with AMD as I have already invested in their platform and its performance characteristics. This is what's not being grasped... They have a cohesive simplified sensible transition along their scaling at an insane value. Today's performance is tomorrow's turtle given the ever increasing progress of technology. The ship sailed on 8 core. Intel wasn't around. The ship sailed on HEDT. Intel wasn't around. Now comes the late delivery (with better performance) at a significant price premium. Was I supposed to wait around for this? I bought a platform and I am settled for years until the next major YUGE bump. AMD captured me first, I'm not buying anything until 7nm this includes GPUs. Even then it has to make sense in terms of performance/value.

 

[ub4ty]

Now that it is quite clear Intel is dragging their feet, I'm sure AMD will let Intel have their brief moment in the spotlight, and then unleash 7nm fury (no not THAT fury...though that'd be an interesting thing to do) in April of next year. I guess there is still the possibility of a 2800X.

It's too bad AMD had to split their platform up into 3 different sockets. 1 socket for desktop/HEDT and 1 for server would have helped them in this scenario. Maybe they will rectify this when the AM4 socket gets refreshed. The thought of being able to select a processor from 8-64 cores at X clock speed is an appealing one, and giving AM5 users access to quad channel memory would put another nail in Intel's coffin*.

* Note: Not an AMD fanboy, but Intel has been screwing up consistently as of late. I actually only have 1 AMD system and 3 Intel systems.

Having 3 sockets is just fine. There's no sensible or sound way to converge such scaling onto a single socket.
It's really 2 sockets from a conversational standpoint. One die occupies a particular socket paradigm. MCM dies occupy a particular socket. The only reason the MCM platform has 2 sockets is because of the motherboard footprint and wiring. You can put an EPYC processor into the same socket as Threadripper. It just isn't wired to for it on the motherboard.

AMD has the most simplified and cohesive product line. This is a far cry from socket/chipset mania intel (ARTIFICIAL SEGMENTATION FOR PROFIT).
AMD has just about the minimum number of sockets/platforms you can have. You can't reduce this further.

I am given a clear and visible layout for scaling core count w/ AMD. I just explained it to someone not centered in tech on a phone call in 5min.
With Intel, my head would explode : A decision they made when they had the market cornered to increase margins through artificial segmentation.

 

[epsilon84]

You made a pretty reasonable case why 12C TR (1920x) wouldn't necessarily own a 9900K, so I am not so sure that 12C Zen2 would, unless it is gets 12c and higher clock speed together.

Definitely interesting times for CPU watchers in the next 2-9 months.

If 9900K really is soldered with an 8 core Turbo of 4.7GHz, it will be a beast of a processor at nearly any workload thrown at it.

Then we will be watching to see which Zen2 rumors come true, as it's another potential beast of a processor.

I do think 8C will be the sweet spot on premium desktop for many years to come, much like 4C was for nearly a decade.

1920X is Zen v1, lower clocks, lower IPC. Using my calculations the 1920X was still ahead in MT, albeit not by a lot.

Add 20% fmax and 10% IPC to a 1920X and the equation changes considerably.

Of course if Zen2 stays as 8C for mainstream desktop then I'll agree it may not necessarily dethrone the 9900K in anything except power consumption or performance/watt

 

[epsilon84]

Marketing, yes, but they do also claim 5GHz outright. That cannot be weaseled out of.

In these days of single core turbo boosts, its easy to market something as 5GHz - just look at the 8086K - exact same turbo boosts as a 8700K except a 1C boost to 5GHz for lulz. Voila! '5GHz CPU!!!'

The upcoming 9900K is also technically a 5GHz chip too since it is rumoured to boost that high for 1C/2C loads, though it's not nearly as ridiculous as the 8086K is, as its sustained all core turbo is at 4.7GHz also (yes I know many people are still sceptical of this, but it is the info we have thus far)

What is stopping AMD from 'doing a 8086K' just to get on the 5GHz bandwagon?

Like Peterscott I'll be very surprised (but pleasantly so) if Zen2 will hold sustained 5GHz clocks... might give me an actual reason to go AMD again for the first time in over a decade. The things holding me back since Zen launched were the lacklustre ST and low clocks, if AMD can get both metrics up to (or exceed!) CFL then I'll be rapt!

 

[maddie]

Still remains to be seen.

Also saying a process is capable of 5GHz, doesn't mean every design can hit that. You actually have to design parts that will clock that high. Remember when AMD was asked what all the extra transistors were doing on Vega. They said a huge number of transistors were to enable higher clock speeds.

I'll be surprised if ZEN2 hits 5GHz. Though it would be a pleasant surprise. Intel would definitely have a lot more to worry about if that happens.

In these days of single core turbo boosts, its easy to market something as 5GHz - just look at the 8086K - exact same turbo boosts as a 8700K except a 1C boost to 5GHz for lulz. Voila! '5GHz CPU!!!'

The upcoming 9900K is also technically a 5GHz chip too since it is rumoured to boost that high for 1C/2C loads, though it's not nearly as ridiculous as the 8086K is, as its sustained all core turbo is at 4.7GHz also (yes I know many people are still sceptical of this, but it is the info we have thus far)

What is stopping AMD from 'doing a 8086K' just to get on the 5GHz bandwagon?

Like Peterscott I'll be very surprised (but pleasantly so) if Zen2 will hold sustained 5GHz clocks... might give me an actual reason to go AMD again for the first time in over a decade. The things holding me back since Zen launched were the lacklustre ST and low clocks, if AMD can get both metrics up to (or exceed!) CFL then I'll be rapt!

For sure, this is speculative, but quite within the realm of possibilities. To be certain, when Papermaster said that they had to double up their teams for the transition to 7nm, it might mean the increase in work for the TSMC process, but some more reasoning makes this unlikely.

If they doing 7nm CPUs at both foundries, for separate markets, then they might as well do it properly. In spite of the similar dimensions there are still a lot of differences that will need optimization tuning for each case, and we all know the needs of servers are different than those on the desktop. A sort of half way between fully shared dies and completely distinct designs. They can save a lot of money and still reap most of the benefits. It just seems the type of smart, lower cost decision making AMD has exhibited lately. A shared RTL design.

 

[AtenRa]

I did the following test on Z370:

Copied large disk image to external USB3 Sam T5 drive from SAM M.2 that is connected to chipset
And ran Aida64 linear test on a different Sam M.2 drive during the same time. I guess the results speak for themselves:

During:

After USB copy was done:

Could you please run your M2.0 alone, just to see how much bandwidth it can get. Because if your single M2.0 doesnt reach the maximum bandwidth of what the Intel DMI 3.0 can give, then you will not get any bottlenecks using a second device (USB/SATA etc)

 

[AtenRa]

While I admit I also hate the intel dmi 3.0 limits, in reality it will probably mostly work fine. The only time it is a limit is when you want to copy between 2 fast m2 drives. In most other cases the end-point will be the limit be it a sata ssd or something on usb.



See above. I'm no fan of dmi3 limits. But if you think logically, you will only saturate a fast m2 ssd if you copy to another fast ssd. And then intel and AMD fail. intel because of dmi limits, amd because the second one on the chipset is slow. In your case with a sata/usb target dmi3 won't matter because the end point, the sata or usb drive will be the bottleneck and not dmi.

If your single M2.0 drive can reach the Intel DMI 3.0 throughput , using a USB 3.0 and or SATA drives at the same time will limit your total bandwidth to the DMI 3.0. Thus your M2.0 drive will get a bottleneck and will not reach full speed.
If your single M2.0 drive doesnt reach the Intel DMI 3.0 bandwidth then you will not have a problem even copying from 2x M2.0 drives.

In AMDs case,
If your M2.0 drive can reach the 4x PCIe bandwidth, then adding a second or even third device to the system will not limit your first M2.0 drive bandwidth like in the Intel DMI 3.0.

The problem with AMDs chipset it that they only have PCI-e Gen 2.0, and every drive you install in the chipset will get limited by this bandwidth.
It is simple as that.