Speculation: i9-9900K is Intel's last hurrah in gaming

Page 23 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Will Intel lose it's gaming CPU lead in 2019?


  • Total voters
    184
  • Poll closed .

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
AMD isn't ever going to take the gaming crown from Intel. IPC isn't the problem, it's the huge latency hit caused by the off-die memory controller. They've committed to this design and there is no going back.

Intel still doesn't believe in separating IO and memory controllers into their own die, and once they get their 10nm+/++ up and running at 14nm frequencies they'll comfortably extend the lead in gaming once again.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,482
14,434
136
AMD isn't ever going to take the gaming crown from Intel. IPC isn't the problem, it's the huge latency hit caused by the off-die memory controller. They've committed to this design and there is no going back.

Intel still doesn't believe in separating IO and memory controllers into their own die, and once they get their 10nm+/++ up and running at 14nm frequencies they'll comfortably extend the lead in gaming once again.
Dream on
 
  • Like
Reactions: Tlh97 and Drazick

CHADBOGA

Platinum Member
Mar 31, 2009
2,135
832
136
AMD isn't ever going to take the gaming crown from Intel. IPC isn't the problem, it's the huge latency hit caused by the off-die memory controller. They've committed to this design and there is no going back.

Intel still doesn't believe in separating IO and memory controllers into their own die, and once they get their 10nm+/++ up and running at 14nm frequencies they'll comfortably extend the lead in gaming once again.

Zen 3 may be out before Intel can respond with a 10nm product. :eek:
 

VirtualLarry

No Lifer
Aug 25, 2001
56,227
9,990
126
Hate to say it, but he does have a point. (*) Intel did something similar to Zen2 way back in Lynnfield days, the iGPU and the memory controller were a separate die from the CPU, I believe. It had latency "issues". Intel has gone pretty-much monolithic die since then.

Things might be a-changing for Intel too, though, simply due to yield economics. They might be forced to "use glue" themselves. (EMIB, multi-die, etc.) Just because of yield issues with more cores and smaller lithography, and costs going up. Right now, AMD is waaaay ahead of Intel in that dept. (in terms of multi-die and economies of scale and better yields)

(*) The second part, about multi-die adding latency. The first part, about Intel never losing the gaming crown, yeah, indeed, "Dream On".
 

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
Zen 3 may be out before Intel can respond with a 10nm product. :eek:
Not when it comes to servers at least, both Ice Lake and Zen 3 are 2020. Assuming another 15% IPC uplift from Zen 2 to Zen 3 it'll again be parity between Ice Lake and Zen 3 like it is now with Skylake and Zen 2. Intel will retain the advantage when it comes to latencies.
 

epsilon84

Golden Member
Aug 29, 2010
1,142
927
136
Uhh, some have the 3900K taking the gaming crown from the 9900k. Even the ones that don't AMD is just a few frames behind. Reel it in a little.

Sure, you can make the argument that the 3900X has the 'gaming crown' with an outlier review.

But there are 20 others out there with a differing view.

At the end of the day, I think both the 3900X and 9900K are overpriced as a strictly 'gaming only' CPU.

You can easily save $200 from either and still get comparable gaming performance.

But on the topic was of this thread, I think the results are definitive enough to say that the 9900K hasnt been 'dethroned' as the fastest gaming CPU, although in many cases the lead it has over the 3900X is small enough to be trivial.
 
Last edited:

PotatoWithEarsOnSide

Senior member
Feb 23, 2017
664
701
106
Yeah, I think a 3600 paired with a 5700/5700XT makes for a VERY potent upper mid range gaming machine - it's unprecedented value that we haven't seen in a while.

I know I'm sounding greedy, but oh how I wish those earlier rumours about Matisse clocking to 4.6/4.7GHz were true. Alas, it appears we still have that 4.3GHz wall for Zen 2, which is my main issue with the Ryzen 3000 CPUs, even though for the vast majority of gamers a 4GHz+ Ryzen 3000 chip would be more than enough still.
That is precisely what I'll be going for. 3600 B450 5700XT with 3200MHz RAM (not b-die either). £800 for the lot.
 

PotatoWithEarsOnSide

Senior member
Feb 23, 2017
664
701
106
AMD isn't ever going to take the gaming crown from Intel. IPC isn't the problem, it's the huge latency hit caused by the off-die memory controller. They've committed to this design and there is no going back.

Intel still doesn't believe in separating IO and memory controllers into their own die, and once they get their 10nm+/++ up and running at 14nm frequencies they'll comfortably extend the lead in gaming once again.
AMD stated it to be 2-3ns, which is an excellent trade-off if you're getting UMA instead of NUMA.
Not sure that's the huge latency hit that you make it out to be TBH.
 

PotatoWithEarsOnSide

Senior member
Feb 23, 2017
664
701
106
I've never seen anything from Intel at 30ns. 45ns sure, against 67ns for Ryzen.
However, headline numbers mean jack. The whole point of increased L3 is fewer trips to memory. Zen 2 has up to 2.5x the L3 of a 9900K. Effective latency is a combination of all requests, not just those that don't fit inside the cache system. Intel are still ahead, but the difference is much narrower than your headline figures imply.
The counter is that once Intel get their next node shrink they could also pick up easy performance from increased L3 themselves; low hanging fruit.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
I've never seen anything from Intel at 30ns. 45ns sure, against 67ns for Ryzen.
However, headline numbers mean jack. The whole point of increased L3 is fewer trips to memory. Zen 2 has up to 2.5x the L3 of a 9900K. Effective latency is a combination of all requests, not just those that don't fit inside the cache system. Intel are still ahead, but the difference is much narrower than your headline figures imply.
The counter is that once Intel get their next node shrink they could also pick up easy performance from increased L3 themselves; low hanging fruit.
You cannot compare numbers resulting from different test methodologies. Just look at Andrei's graphs. He goes over in detail on what is being measured, and I trust his numbers to be accurate.

It doesn't matter what size the L3 is if games request dataset sizes which exceed L3 capacity, which is what modern games do, which is why memory latency is so important, as implied in the graph from Intel.
 
  • Like
Reactions: Tlh97 and ozzy702

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
AMD isn't ever going to take the gaming crown from Intel. IPC isn't the problem, it's the huge latency hit caused by the off-die memory controller. They've committed to this design and there is no going back.
You remind of another great forum user who said something along the lines: "Solder is gone from mainstream and is never coming back."

Clock for clock Zen 2 is within 10% of Coffee Lake when it comes to lowres gaming, with a "huge latency hit caused by the off-die memory controller". So the latency is huge, meaning any improvement they make is bound to be significant in terms of performance, and they only need 10% to catch up in terms of perf per clock in gaming. Never say never.

Adding more L3 isn't going to help with only up to four cores utilizing it without performance penalty due to the CCX design.
This CCX design?

latency-jpg.8113
 

CHADBOGA

Platinum Member
Mar 31, 2009
2,135
832
136
Not when it comes to servers at least, both Ice Lake and Zen 3 are 2020. Assuming another 15% IPC uplift from Zen 2 to Zen 3 it'll again be parity between Ice Lake and Zen 3 like it is now with Skylake and Zen 2. Intel will retain the advantage when it comes to latencies.

Various sources are stating Intel won't have 10nm desktop till either 2021 or 2022 and as the discussion is about gaming, I wouldn't see servers as being that meaningful in this regard.

I don't see how waiting to 2022 is sustainable, so I imagine if Intel's process is still unviable, they will go to a foundry, but who knows then what date it will come out.
 
  • Like
Reactions: Tlh97 and guachi

Arkaign

Lifer
Oct 27, 2006
20,736
1,377
126
Various sources are stating Intel won't have 10nm desktop till either 2021 or 2022 and as the discussion is about gaming, I wouldn't see servers as being that meaningful in this regard.

I don't see how waiting to 2022 is sustainable, so I imagine if Intel's process is still unviable, they will go to a foundry, but who knows then what date it will come out.

I'm thinking based on process tech, the move towards splitting things up into various sub-bits, and frequency walls/thermal density, that we're in for some stagnation on per-core performance. Well .. MORE stagnation.

I wouldn't hesitate a bit to build any CL or Zen2 system with 6 or 8 cores depending on need. I don't think we're going anywhere for a WHILE. :/
 

PotatoWithEarsOnSide

Senior member
Feb 23, 2017
664
701
106
You cannot compare numbers resulting from different test methodologies. Just look at Andrei's graphs. He goes over in detail on what is being measured, and I trust his numbers to be accurate.

It doesn't matter what size the L3 is if games request dataset sizes which exceed L3 capacity, which is what modern games do, which is why memory latency is so important, as implied in the graph from Intel.
...from Intel.

Of course they are going to cherry pick their best case scenarios to highlight a point. It is well known that Ryzen suffers in FC5 and GTA V.
I'd like to see similar graphs relating to Zen+ and Zen 2. I've asked for such before.
If we look at CS:GO we can see that even Intel sees big uplifts in L3 hits, and that should suggest that Zen 2 L3 should hit even more. Guess what? Zen 2 now wins in CS:GO.
What matters how many fewer requests to memory there are as a result of increased L3, as each L3 hit there is a net 50ns+ improvement over Zen+. My own calculations indicate only 25% fewer trips to memory are needed as a result of larger L3 for Zen 2 to have the same real world latency as Intel.
You can't argue that larger L3 doesn't reduce the number of trips to memory required. The only question is by how much they are reduced. That's why headline figures mean diddly squat.
 
  • Like
Reactions: Abwx

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
Clock for clock Zen 2 is within 10% of Coffee Lake when it comes to lowres gaming, with a "huge latency hit caused by the off-die memory controller". So the latency is huge, meaning any improvement they make is bound to be significant in terms of performance, and they only need 10% to catch up in terms of perf per clock in gaming. Never say never.
The difference in memory latency right now is 2-2.5x between Zen 2 and Skylake. To bring down the 10% difference between these two in low res gaming to 0% or even exceed Skylake, AMD needs to halve the memory latency in Zen 3. That is going to be difficult with memory-controllers off-die.
This CCX design?
That's core-to-core latency, I'm talking about the latency when data size exceeds 16MB on both architectures.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
Various sources are stating Intel won't have 10nm desktop till either 2021 or 2022 and as the discussion is about gaming, I wouldn't see servers as being that meaningful in this regard.

I don't see how waiting to 2022 is sustainable, so I imagine if Intel's process is still unviable, they will go to a foundry, but who knows then what date it will come out.
You mean the leaked slides which puts Rocket Lake in 2022? Those are likely to be outdated or fake.
 
  • Like
Reactions: Tlh97 and Zucker2k

B-Riz

Golden Member
Feb 15, 2011
1,482
612
136
I think Zen2 succeeds the ways that it does due to design, it doesn't look like 7nm really helps much overall. By that I mean that a hypothetical Zen2 on 12nm/14nm would probably be roughly the same thing with higher power/heat (but perhaps less heat density).

We may well be in an era when process tech and new nodes are truly seeing edge thin improvements at extreme cost and engineering difficulties.

If that's the case, we may see nothing substantial for at least 2-3 years. Good time to get a Zen2 build (or if just gaming, keeping/building a CL K-series OC setup).

I don't remember where I saw it, it was a YouTube piece I think, but 7nm allowed for the increased cache size of Zen2 while keeping most of the other OG Zen stuffs, Zen2 very much benefits from 7nm.

Zen2 and 7nm has fixed a lot of the frustrations of Zen (memory, gaming perf, etc.) while keeping the same price points or better.
 

Arkaign

Lifer
Oct 27, 2006
20,736
1,377
126
I don't remember where I saw it, it was a YouTube piece I think, but 7nm allowed for the increased cache size of Zen2 while keeping most of the other OG Zen stuffs, Zen2 very much benefits from 7nm.

Zen2 and 7nm has fixed a lot of the frustrations of Zen (memory, gaming perf, etc.) while keeping the same price points or better.

Probably Adored, guy was constantly throwing BS against the wall.

Nothing would stop them from putting the Zen2 design on 12nm or 14nm, it would just be bigger dies, higher power consumption, lower power/heat density. Perhaps even higher clock headroom. Look at the Zen2 ICs under the IHS, there's loads of room.

7nm primary advantage seems to be how many dies per wafer, thus profit, as long as they keep the fab costs under control.

IOW, I think the good stuff on Zen2 is all design, not really anything special to do with 7nm.

We saw similar things with Vega to Vega 7. 7nm made nearly zero difference.
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
Sure, you can make the argument that the 3900X has the 'gaming crown' with an outlier review.

But there are 20 others out there with a differing view.

At the end of the day, I think both the 3900X and 9900K are overpriced as a strictly 'gaming only' CPU.

You can easily save $200 from either and still get comparable gaming performance.

But on the topic was of this thread, I think the results are definitive enough to say that the 9900K hasnt been 'dethroned' as the fastest gaming CPU, although in many cases the lead it has over the 3900X is small enough to be trivial.
this
the difference is not 1 or 3%, it is much more especially on the 1 and 0,1% percentile
when 9900K is oced it is even more
5GHz 9900K is a power hog in apps, but while gaming thats more spiky than constant load
I would like to see a review of gaming power consumption in CPU intensive scenes

the whole internet seems as incompetent, just running benches without much insides in them
I really wanna see the difference in power of the whole system with 3700x and 9900k at stock and oced and normalise it to performance as it is very non linear
one of the few articles with this comparison https://www.purepc.pl/procesory/tes..._7_3700x_premiera_architektury_zen_2?page=0,3
but just only performance, power only while cinepeenbenching
 

Qwertilot

Golden Member
Nov 28, 2013
1,604
257
126
IOW, I think the good stuff on Zen2 is all design, not really anything special to do with 7nm.

We saw similar things with Vega to Vega 7. 7nm made nearly zero difference.

Isn't that because they didn't redesign V7 for 7nm & then clocked it up quite high as well?

The benefits from new nodes are still there, but they do definitely (Intel's processes as well of course) seem to require huge investments to really realise, and to generally favour low clock/lowish power things. Physics is definitely fighting back.
 

Arkaign

Lifer
Oct 27, 2006
20,736
1,377
126
Isn't that because they didn't redesign V7 for 7nm & then clocked it up quite high as well?

The benefits from new nodes are still there, but they do definitely (Intel's processes as well of course) seem to require huge investments to really realise, and to generally favour low clock/lowish power things. Physics is definitely fighting back.

Vega 64 1.274Ghz, 4096 Cores, 14nm, 295W
Vega 7 1.400Ghz, 3840 Cores, 7nm, 300W

The 7 doubled the HBM bus from 2048 to 4096 bit and a little more than doubled memory bandwidth, which is basically all of the performance increase came from (eg; downclock and undervolt the Vega7 to Vega64 levels, and performance doesn't drop much from stock, but remains above the 64 due to the 200%+ memory speeds).