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

[PeterScott]

The speculation was done ages ago, because chip makers don't respond within a couple of months to competing products. Road maps are planned years in advance. AMD can essentially do nothing, or it can do marketing moves: Promotions and price changes.

The best promotion would be deals on 12 core Threadripper. Spend a bit more to jump to the HEDT platform with 50% more cores, quad channel memory.

I would try my damnedest to have a last clearance sale of TR 1920X for $499, to coincide with 9900K launch, and maybe some packages with the most value oriented TR MBs that might compare favorably with 9900K/Z390 combo.

IMO that would be the best way to rain on Intels 9900K launch.

 

[Markfw]

The speculation was done ages ago, because chip makers don't respond within a couple of months to competing products. Road maps are planned years in advance. AMD can essentially do nothing, or it can do marketing moves: Promotions and price changes.

The best promotion would be deals on 12 core Threadripper. Spend a bit more to jump to the HEDT platform with 50% more cores, quad channel memory.

I would try my damnedest to have a last clearance sale of TR 1920X for $499, to coincide with 9900K launch, and maybe some packages with the most value oriented TR MBs that might compare favorably with 9900K/Z390 combo.

IMO that would be the best way to rain on Intels 9900K launch.

Or put the 16 core on sale for $700, which it was for quite some time, and the 12 core was also on sale, I just forget for how much. The 16 core is still $780

Oh, and then launch a 32 core that blows away anything Intel has for anytime in the near future, which is supposed to go on pre-sale tomorrow morning I heard.

 

[PeterScott]

Or put the 16 core on sale for $700, which it was for quite some time, and the 12 core was also on sale, I just forget for how much. The 16 core is still $780

Oh, and then launch a 32 core that blows away anything Intel has for anytime in the near future, which is supposed to go on pre-sale tomorrow morning I heard.

An $1800 32 core CPU is not going to pull sales from $450 desktop parts.

OTOH, A $500 1920x is close enough to a $450 9900K, to sway people. The further above 9900K pricing you get, the less people will even entertain the idea.

 

[epsilon84]

An $1800 32 core CPU is not going to pull sales from $450 desktop parts.

OTOH, A $500 1920x is close enough to a $450 9900K, to sway people. The further above 9900K pricing you get, the less people will even entertain the idea.

I don't think people realise how close a 9900K will be to a 1920X in MT performance. Obviously it will be behind in Cinebench, but if we go by Stilts IPC calculations for Zen then an 8/16 9900K @ 4.7GHz won't be that far behind a 12/24 1920X @ 3.7GHz in MT performance and obviously destroy it in ST performance.

Of course the HEDT platform will have its own advantages beyond simply MT performance, such as extra PCI-E lanes, a better upgrade path to gen 2 TR etc... But mark my words in pure compute throughout a 9900K would be very close to a 1920X:

4.7 x 8 x 1.15 = 43.24 (1.15 is the relative IPC advantage)
3.7 X 12 = 44.4

If given the choice, unless I absolutely needed the HEDT platform features, I would choose a 9900K over a 1920X, especially since the 9900K should consume less power and cost siginificantly less once platform costs are taken into account, even if AMD discounts the 1920X as you suggested.

 

[Vattila]

With 16-core Threadripper 2950X hitting 4.4 GHz boost clock, AMD has finally realised the 10% performance increase expected from the 12LP process. Can they bin a 2800X to do the same at lower TDP, and would it be worth it? We have had no leaks or further hints about a 2800X at all, so I guess it is unlikely (although Jim Anderson would probably love to pull a rabbit out of the hat). Lowering the entry price to the Threadripper platform is likely the chosen tactic.

My wild speculation about a design respin with 12LP standard cell libraries was largely based on the need to do the work for the "Picasso" APU refresh. But that is probably done with the 7.5-track version of the libraries (versus 9T for "Pinnacle Ridge"), for better density and efficiency, and 7.5T does not clock as high as 9T (fewer tracks per cell and fewer fins per transistor, thus less drive current, as I understand it).

I guess they could have done a respin with optimised 12LP 9T libraries ("Pinnacle Ridge" reused the 9T libraries from the 14LPP process, only replacing the transistor layers for the improved transistors, as I understand it), which according to GlobalFoundries' press material should give a 15% density gain (presumably from optimised cells and design rules rather than process dimensions, as the latter are exactly the same as 14LPP, according to Wikichip). A higher density 9T library should allow higher frequency, I presume (since higher density means less RC delay, which means higher frequency potential, as I understand it). All that said, it would probably be too costly for little gain, unless they had obvious architectural bottlenecks they could easily fix.

 

[Vattila]

With 9900K coming in October and the full 9th generation line-up coming in 2019-Q1, the launch schedule looks pretty much like a repeat of the 8700K launch last year. Do we expect higher volumes at launch this time around?

 

[epsilon84]

With 16-core Threadripper 2950X hitting 4.4 GHz boost clock, AMD has finally realised the 10% performance increase expected from the 12LP process. Can they bin a 2800X to do the same at lower TDP, and would it be worth it? We have had no leaks or further hints about a 2800X at all, so I guess it is unlikely (although Jim Anderson would probably love to pull a rabbit out of the hat). Lowering the entry price to the Threadripper platform is likely the chosen tactic.

My wild speculation about a design respin with 12LP standard cell libraries was largely based on the need to do the work for the "Picasso" APU refresh. But that is probably done with the 7.5-track version of the libraries (versus 9T for "Pinnacle Ridge"), for better density and efficiency, and 7.5T does not clock as high as 9T (fewer tracks per cell and fewer fins per transistor, thus less drive current, as I understand it).

I guess they could have done a respin with optimised 12LP 9T libraries ("Pinnacle Ridge" reused the 9T libraries from the 14LPP process, only replacing the transistor layers for the improved transistors, as I understand it), which according to GlobalFoundries' press material should give a 15% density gain (presumably from optimised cells and design rules rather than process dimensions, as the latter are exactly the same as 14LPP, according to Wikichip). A higher density 9T library should allow higher frequency, I presume (since higher density means less RC delay, which means higher frequency potential, as I understand it). All that said, it would probably be too costly for little gain, unless they had obvious architectural bottlenecks they could easily fix.

4.4GHz boost is only an extra 100MHz over the 2700X boost though. Unless a 2800X can sustain a 4.2GHz all core turbo I don't think it will be worth it, and even then we're barely trimming the performance deficit rather than making the 2800X truly competitive with a 9900K, so I have my doubts. It's possible we could have a 2800X though, just depends how AMD bins their best dies and how much liberty they take with their TDPs

 

[coercitiv]

I don't think people realise how close a 9900K will be to a 1920X in MT performance.

AFAIK, TR2 will launch before 9900K.

PS: why are you using overclocked numbers for 9900K and stock numbers for 1920X?

 

[VirtualLarry]

just depends how AMD bins their best dies and how much liberty they take with their TDPs

Well, they did give us the FX-9590 or whatever that crazy 5Ghz Vishera-derived CPU was. What was the TDP on that one? 140W? 220W?

 

[StinkyPinky]

I think it's likely the 9900k will get to 5GHz with a cooler while the 2920x will probably get to 4.3. Am expecting the 2920x to have about 20-25% better multicore but obviously at a price (cost and single core performance). The other issue is thunderbolt.

Of course the advantage of the thread ripper is you can probably slot in a 5Ghz 16 core once it hits 7nm

 

[epsilon84]

AFAIK, TR2 will launch before 9900K.

PS: why are you using overclocked numbers for 9900K and stock numbers for 1920X?

I'm not, have a look at the turbo speeds for the 9900K - 4.7GHz all core / 5.0 single core.

TR2 will obviously outperform a 9900K but I don't expect it to cost $500 either, do you?

 

[french toast]

At what point will bandwidth play a part in all this?...9900k is going to need some fast memory to get to it's potential imo.

 

[epsilon84]

At what point will bandwidth play a part in all this?...9900k is going to need some fast memory to get to it's potential imo.

If you stick with stock standard DDR4 2666, there could be a slight bottleneck there. I'm sure most 9900K owners would run high end B Die RAM modules at 4000MHz+ though - who is going to gimp on memory on a $450 chip?

 

[coercitiv]

I'm not, have a look at the turbo speeds for the 9900K - 4.7GHz all core / 5.0 single core.

So we're throwing common sense out the window now?

The current 8700K 6c/12t does 95W at 4.3Ghz. A hypothetical 8900K 8c/12t at 4.7Ghz would result in a 10% increase in frequency, likely a 10% increase in voltage, 33% increase in core count. All these would total ~75% increase in power assuming same node, or in terms of node improvements to get 9900K within 95W would require a 40% power drop.

Do seriously expect Intel to drop 14nm power usage by another 40% at such high clocks?

TR2 will obviously outperform a 9900K but I don't expect it to cost $500 either, do you?

I don't know what to expect, but it certainly sounds weird to favorably compare 9900K over 2920X (or whatever bottom TR2 SKU) because of the price difference, that would mean 2700X would favorably compare over 9900K over a similar price difference.

Either we acknowledge what was said before about mainstream and enthusiast platforms - they're apples and oranges - or we're just gonna have a clown fiesta here, comparing different platforms based on preferred criteria while conveniently ignoring their shortcomings.

 

[JoeRambo]

Either we acknowledge what was said before about mainstream and enthusiast platforms - they're apples and oranges

This thread is speculation about AMD response to Intel desktop chip. Speculating about AMD responding to it with cheaper TR is perfectly fine. All merits of doing so should be evaluated, including ST/MT performance and total platform price (MB, quad ram kit).
And attempts to establish tangents about TDP were already discussed in this thread - neither vendor cares about sticking TDP, so we shouldn't either. If cooling allows and MB is properly built I expect 9900K to run 8C 4.7Ghz all core turbo all day long in non "prime" type loads.

 

[epsilon84]

So we're throwing common sense out the window now?

The current 8700K 6c/12t does 95W at 4.3Ghz. A hypothetical 8900K 8c/12t at 4.7Ghz would result in a 10% increase in frequency, likely a 10% increase in voltage, 33% increase in core count. All these would total ~75% increase in power assuming same node, or in terms of node improvements to get 9900K within 95W would require a 40% power drop.

Do seriously expect Intel to drop 14nm power usage by another 40% at such high clocks?


I don't know what to expect, but it certainly sounds weird to favorably compare 9900K over 2920X (or whatever bottom TR2 SKU) because of the price difference, that would mean 2700X would favorably compare over 9900K over a similar price difference.

Either we acknowledge what was said before about mainstream and enthusiast platforms - they're apples and oranges - or we're just gonna have a clown fiesta here, comparing different platforms based on preferred criteria while conveniently ignoring their shortcomings.

The best 14nm dies won't necessarily need any additional voltage at 4.7GHz compared to 4.3GHz on current CFL chips. My 8700K is stable at stock volts up until 4.6 - 4.7GHz and it is an average chip, does 5.0GHz at 1.37V so its not a golden sample by any stretch. The golden sample 8700Ks (and the 8086Ks) generally hit 5.0GHz between 1.25V to 1.30V and these are the dies Intel will use with a 9900K.

Of course the 9900K will consume more power than a 8700K, the same way a 8700K consumes more power than a 7700K despite having practically the same TDP (95W vs 91W).

I'm not expecting Intel to defy physics, but I'm expecting them to push the 9900K to the edge because this could possibly be their last big victory for a while on the desktop.

I actually think the 9900K somewhat blurs the line between HEDT and mainstream desktop - you're approaching HEDT levels of performance yet it's being marketed as a desktop chip.

I think the discussion comparing it to the 1920X is perfectly valid - even the 2920X even though they are technically not in the same category. Like I said, the 9900K creates a bit of overlap between the current desktop flagships (8700K/2700X) and the 12/24 TR chips.

 

[Zucker2k]

So we're throwing common sense out the window now?

The current 8700K 6c/12t does 95W at 4.3Ghz. A hypothetical 8900K 8c/16t at 4.7Ghz would result in a 10% increase in frequency, likely a 10% increase in voltage, 33% increase in core count. All these would total ~75% increase in power assuming same node, or in terms of node improvements to get 9900K within 95W would require a 40% power drop.

Do seriously expect Intel to drop 14nm power usage by another 40% at such high clocks?


I don't know what to expect, but it certainly sounds weird to favorably compare 9900K over 2920X (or whatever bottom TR2 SKU) because of the price difference, that would mean 2700X would favorably compare over 9900K over a similar price difference.

Either we acknowledge what was said before about mainstream and enthusiast platforms - they're apples and oranges - or we're just gonna have a clown fiesta here, comparing different platforms based on preferred criteria while conveniently ignoring their shortcomings.

The only problem with this analysis is that 4700Mhz is turbo speed, so it's not constrained by TDP figures.

 

[epsilon84]

The only problem with this analysis is that 4700Mhz is turbo speed, so it's not constrained by TDP figures.

Pretty sure the TDP topic was discussed ad nausuem in the other 9900K thread... Some people will continue to use that as a hard limit or ceiling on potential clockspdeds when Intel clearly states TDP is derived from the base clock, not the maximum turbo clock.

Under load I would expect a 9900K to be around 130W, based on the power consumption of the 8700K +33%. The soldered IHS should help somewhat with power leakage, otherwise it would probably be around the 140W mark due to the higher clocks.

 

[LTC8K6]

The current 8700K 6c/12t does 95W at 4.3Ghz.

Intel calculates TDP at base clock, which is 3.7Ghz for the 8700K.

Processor Base Frequency
Processor Base Frequency describes the rate at which the processor's transistors open and close. The processor base frequency is the operating point where TDP is defined. Frequency is measured in gigahertz (GHz), or billion cycles per second.

TDP
Thermal Design Power (TDP) represents the average power, in watts, the processor dissipates when operating at Base Frequency with all cores active under an Intel-defined, high-complexity workload. Refer to Datasheet for thermal solution requirements.

 

[coercitiv]

Intel calculates TDP at base clock, which is 3.7Ghz for the 8700K.

Great, show me a CPU only load that uses 95W on 8700K while running at 3.7Ghz

 

[epsilon84]

Still going on about TDPs?

Some perspective here - it's a highly clocked 8/16 CPU - of course it's going to consume a lot of power under full load. Yes, far beyond the 95W TDP. No, it won't burn down the house, but you'll want to have a decent cooler for this chip, as any sensible person would.

 

[PeterScott]

I'm not, have a look at the turbo speeds for the 9900K - 4.7GHz all core / 5.0 single core.

TR2 will obviously outperform a 9900K but I don't expect it to cost $500 either, do you?

I have seen the rumored boost table, but I have a really hard time believing that the 9900K will have 8 core 4.7GHz core turbo that it shows. That's 400MHz higher than 8700K with two more cores on top. That would imply big process improvements, which at this point, I would think they are well into diminishing returns with 14nm++.

Also I don't think there is a 15% IPC advantage outside of gaming. Here is a clock for clock 8700K vs Ryzen gen 1/2 comparison:
https://www.techspot.com/article/1616-4ghz-ryzen-2nd-gen-vs-core-8th-gen/page3.html

It's mainly games where there is an IPC advantage. Otherwise Ryzen does quite well.

But otherwise, your point is noted. It isn't like this is 50% lead for 1920x. Clocks speed and IPC cuts into that lead, where it ends up remains to be seen.

 

[Zucker2k]

Great, show me a CPU only load that uses 95W on 8700K while running at 3.7Ghz

That's easy, actually; it consumes less than TDP at stock clocks.

The Core i7-8700K has a TDP of 95W, but consumes 86.2W at full load, of which the cores account for 78.6W. The rest of the power is consumed mostly by the uncore and the memory controller.

The Core i5-8400 is rated at 65W, and consumes only 49.3W at full load, of which 41.7W is from the cores. That leaves 7.6W on the table for the uncore and memory controller, which is almost identical to that of the Core i7-8700K, showing the similarity in design.

Source: https://www.anandtech.com/show/1185...-lake-review-8700k-and-8400-initial-numbers/5

The bad rep Intel gets is people using AVXx stress testing figures to compare to AMD (which clearly does a poor job of running AVX code as 'efficiently' as Intel due to AMD playing catch up in that realm, and implementation. Like I opined in another thread, Coffeelake is very efficient at Ryzen clocks, but the opposite is far from the case, if not impossible altogether.

Edit: If Mr. ABWX slaps his usual -23% VR efficiency deficit on top of these figures, like he does all his Ryzen power consumption calculations, you can imagine how low these numbers can be.

 

[epsilon84]

I have seen the rumored boost table, but I have a really hard time believing that the 9900K will have 8 core 4.7GHz core turbo that it shows. That's 400MHz higher than 8700K with two more cores on top. That would imply big process improvements, which at this point, I would think they are well into diminishing returns with 14nm++.

Also I don't think there is a 15% IPC advantage outside of gaming. Here is a clock for clock 8700K vs Ryzen gen 1/2 comparison:
https://www.techspot.com/article/1616-4ghz-ryzen-2nd-gen-vs-core-8th-gen/page3.html

It's mainly games where there is an IPC advantage. Otherwise Ryzen does quite well.

But otherwise, your point is noted. It isn't like this is 50% lead for 1920x. Clocks speed and IPC cuts into that lead, where it ends up remains to be seen.

I believe 4.7GHz all core turbo is within the realms of possibility, though it would be pushing the 14nm process to its absolute limits. As I said I've done testing on my statistically average clocking 8700K and 4.6 - 4.7GHz is the threshold of stability at default voltages. Just to be clear, 4.6GHz is for AVX stability testing like Prime95, everything else non AVX is stable at 4.7GHz at stock volts.

A better binned CPU using the best dies, plus a soldered IHS, would likely extend this range an extra 200MHz at least, which leaves a bit (though not a lot) of headroom for a 9900K to hit such clocks. Of course, we can't defy physics so logically a 9900K would consume significantly more power than a 8700K, regardless of what the TDP rating is.

WR to IPC, as I stated earlier, I referenced The Stilts testing which put CFL 14.5% ahead of Zen refresh. Zen 1 (which the 1920X is based off) is another 3% slower per clock on top of that IIRC, but I wanted to use 15% to not over exaggerate the IPC difference. Also Zen SMT is slightly more efficient than Intel HT at MT scaling so I thought that would make the 15% figure fairly accurate when talking into account MT workloads.

As I said, I expect the 1920X to have a slight edge in MT performance but people may be surprised how close the 9900K would be.

Add to that the much higher ST performance plus lower platform costs would make the 9900K the better buy for a desktop system IMO.

 

[rbk123]

That really is no better. Neither anger at Intel, nor Love for AMD are valid reason to engage in FUD, and massive exaggeration of tiny issues.

As far as Irony, show me where I am engaged in FUD against AMD.

Not only did I not say it was better or even ok, on the contrary I implied it resulted in bad behavior. As for irony, your self-appointed roll of fighting for truth, justice and the Intel way has you trying to kill all the AMD FUD, yet oddly enough, you don't lift a finger trying to kill any Intel FUD. The fact you think the FUD is so heavily 1-sided, speaks volumes on your built-in bias.