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

[Vattila]

It's a big gulf to overcome, that ST performance deficit.

Exactly. It is a big problem for AMD. Bigger than you seem to realise. You are wrong to think that AMD can succeed without addressing it. It cannot be ignored. They didn't spend resources on overshooting the 40% IPC target for Zen just for fun. The performance deficit that is seen in game code is also seen in other important segments (in particular, enterprise server). There is nothing exclusive to game code.

Without a roadmap to address this, AMD would have fallen flat already. PC and server OEMs would not have given AMD the time of day. AMD product line? Partnering? Forget it. Intel will soon enough match core counts and threads — as seen today.

In fact, just today, the AMD share price fell a whopping 11%, in large part due to an analyst report from New Street Research, which initiated coverage on AMD with a "sell" recommendation. What was their investment thesis?

"Our detailed review of AMD's architecture and actual needs of the server and PC markets, combined with our analysis of competitive dynamics between Intel and AMD nevertheless suggest the party might be over soon," Ferragu wrote in a note to clients. […] "We don’t see Intel’s delay in the 10 nm node as a material moving part," Ferragu said. "Moving to the next node increases the performance of AMD chips, but still keeps them largely behind Intel on single-thread performance."

https://markets.businessinsider.com...ets-an-impossible-scenario-2018-10-1027634438

A lot of market analysts still see Intel as unbeatable because of their unassailable ST performance advantage. Incidentally, these are the guys that provide market analysis and advice to big businesses and investors.

Agreed — true — it is a big gulf. The size of the gap seems to be your main argument. Why do you think AMD is incapable of overcoming it? Not even on Zen-aware modern game code?

 

[epsilon84]

What exactly is 'zen aware' modern game code exactly? I'm not trying to be condescending, but I'm rather curious what constitutes 'zen aware' because I've seen this mentioned a few times in this thread.

https://www.gamersnexus.net/hwrevie...paste-delid-gaming-benchmarks-vs-2700x/page-4

Both F1 2018 and Far Cry 5 are 2018 games that show a massive gulf between CFL and Ryzen+, in fact that gulf isn't really any different to the gap in GTA 5, a much older game released in 2013!

What about Assasins Creed: Odyssey? That's a brand new game, and again showing a big gulf between the CFL chips and Ryzen+: https://www.techspot.com/review/1730-intel-core-i9-9900k-core-i7-9700k/page4.html

The game you highlighted in your OP, Forza Horizon 4, is a clear anomaly because *every single CPU* shows exactly the same framerate at 1080P, which means there is something else at hand limiting the performance here, whether that be drivers, game engine or some kind of frame rate cap - I don't know. But it's hardly a good example of a 'Ryzen optimised game' as you tried to portray.

In fact, looking at the techspot 9700K.9900K review, the only game where Ryzen+ draws close to the CFL chips without there being an obvious 'fps wall' for the CFL chips would appear to be Star Wars Battlefront II - but one game is hardly conclusive proof of your point. Do you perhaps have other examples that may show this 'zen aware' game code ?

Basically, yeah, I'm skeptical AMD can completely eradicate the ST deficit in one generation, especially when it comes to gaming. They can compete in many other ways while still having a small ST deficit, for example, by offering more cores/threads than Intel at the same price point. This has worked really well for Ryzen and I don't see how that advantage can't continue to be used in Zen 2. Unfortunately, gaming is one of the areas where ST performance is probably more important that MT throughput. Most game code really still only relies on a few fast threads, which is why even an i3 8350K basically matches a Ryzen 2700X in gaming: https://www.techpowerup.com/reviews/AMD/Ryzen_7_2700X/12.html

 

[PotatoWithEarsOnSide]

Memory latency badly affects Ryzen performance, particularly in those single-thread scenarios. A previous post highlighted how the latency reduction for Zen+ resulted in gaming performance beyond its relative latency improvement. This is surely the low hanging fruit that can bring the Zen 2 CPUs back to almost parity with Intel.
In reality, the next gen of Ryzen CPUs only need match Intel's IPC, but have a higher base frequency at a lower power envelop, for AMD to be the better option. Don't assume that everyone buying a K processor is a) overclocking, or b) has all the best other components to make the most of it (read: 9900k buyers won't all be buying water cooling).
Given the choice between a stock Ryzen 3700x at 4.2GHz base 4.6GHz ACT, and a 9900k at 3.6GHz base 4.7GHz ACT, which would you recommend bearing in mind the likelihood that the 3700x would ship with an appropriate cooler, and that the 9900k ships without a cooler but advertises a 95W TDP (that it doesn't adhere to except at base frequency)...?

 

[arandomguy]

Personally I think there is a strong possibility a 8c/16t can "match" the 9900k in gaming. When I say match I mean for all intents and purposes performance between the two is not going to be a major differentiating point beyond hardened fanatics who want to debate it. If they want to be market disruptive they'll under cut pricing wise into the ~$400 range (or lower depending on street prices at the time).

However I don't understand this general insistence on comparing Zen 2 to CFL-R. Zen 2 is roughly 6 months out and in another 6 months roughly from that you'll see Intel's next release. Why is there such an insistence on focusing on those 2 as the competitors? Intel and AMD since Zen launched have staggered their releases and compete half a gen to half a gen.

It is very simple. For Zen2 to match the 9900K in that suite of games, at those settings with that hardware it needs to increase performance by 17% on average.

If you change ram, motherboard, drivers, GPU, games, resolution, settings, windows patch version, bios version etc then it can change the result and therefore the required uplift.

I think that it shows that Zen2 matching the 9900k core for core and thread for thread is entirely doable and that is all it was ever meant to show.

Can AMD match Intel? Yes, it looks possible. Will they? Let's wait and see.

Again this is an over simplification in terms of what translates to real performance via gaming benchmarks. Let's just look at clock speeds, would a hypothetical 30% increase in clock speeds equal 30% more performance in the test suite being discussed? The OC 9900k 8.5% increase in clock speed did not result in 8.5% increase in benchmark numbers in this test suite, not even a 1:2 ratio improvement. Some games had 0 improvement.

I'm just going to circle back to something I mentioned in this thread earlier which I didn't feel like arguing. I think the high side target is likely is roughly 15% IPC (note that how IPC is being used in these cases is subjective) and 15% clock speed and that will match the 9900k in these types of benchmarks. That doesn't mean that the 9900k has a universal 30%+ existing benchmark advantage over the 2700x in gaming.

 

[french toast]

Personally I think there is a strong possibility a 8c/16t can "match" the 9900k in gaming. When I say match I mean for all intents and purposes performance between the two is not going to be a major differentiating point beyond hardened fanatics who want to debate it. If they want to be market disruptive they'll under cut pricing wise into the ~$400 range (or lower depending on street prices at the time).

However I don't understand this general insistence on comparing Zen 2 to CFL-R. Zen 2 is roughly 6 months out and in another 6 months roughly from that you'll see Intel's next release. Why is there such an insistence on focusing on those 2 as the competitors? Intel and AMD since Zen launched have staggered their releases and compete half a gen to half a gen.



Again this is an over simplification in terms of what translates to real performance via gaming benchmarks. Let's just look at clock speeds, would a hypothetical 30% increase in clock speeds equal 30% more performance in the test suite being discussed? The OC 9900k 8.5% increase in clock speed did not result in 8.5% increase in benchmark numbers in this test suite, not even a 1:2 ratio improvement. Some games had 0 improvement.

I'm just going to circle back to something I mentioned in this thread earlier which I didn't feel like arguing. I think the high side target is likely is roughly 15% IPC (note that how IPC is being used in these cases is subjective) and 15% clock speed and that will match the 9900k in these types of benchmarks. That doesn't mean that the 9900k has a universal 30%+ existing benchmark advantage over the 2700x in gaming.

Well you have to compare to something? That is how this all works...

I think zen2 matching 9900k would depend on the IPC/PPC gain is composed of latency improvements that would benefit gaming more than general app performance...ie...'gaming ipc' is greater than 13% due to this reason + frequency benefits which has diminishing returns.

 

[TheELF]

So looking at TPUs summary page (https://www.techpowerup.com/reviews/Intel/Core_i9_9900K/19.html) the 2700X at stock has 84.9% of the gaming performance as the 9900K at stock at 720p across their gaming suite.

So it looks like AMD need around a 17% uplift between IPC and raw clockspeed to match the 9900K on average in gaming. I think that is doable if the reported 13% IPC in scientific workloads is both true and applicable to gaming. I also expect they can do it with a much lower power draw.

Bah humbug,all they need to do is to push for a PS4 pro pro, that will push more bottleneck to the GPU and make them even,hey they might even go for nvidias raytracing to achieve this.
Never forget,as long as it gets the same FPS it's also as fast as.

 

[inf64]

I think it's too early to bring Zen2 for desktop in this conversation. There will surely be IPC uplifts but how much and what clocks will TSMC's process achieve are big unknowns. AMD is fine as it is right now. They have the sweet spot lineup, they are perf./$ kings for all but low resolution gaming (where 8700K is the best CPU to have).

 

[Abwx]

What about Assasins Creed: Odyssey? That's a brand new game, and again showing a big gulf between the CFL chips and Ryzen+: https://www.techspot.com/review/1730-intel-core-i9-9900k-core-i7-9700k/page4.html

This game was tested with a 2080ti by Computerbase...

https://www.computerbase.de/2018-10/assassins-creed-odyssey-benchmark-test/2/

The fact that that the 2700X score increase only by 6% between 1440p and 1080p while the 8700K has 20% progression should have ringed a bell that something is fishy in TS numbers...

 

[moinmoin]

However I don't understand this general insistence on comparing Zen 2 to CFL-R. Zen 2 is roughly 6 months out and in another 6 months roughly from that you'll see Intel's next release. Why is there such an insistence on focusing on those 2 as the competitors? Intel and AMD since Zen launched have staggered their releases and compete half a gen to half a gen.

Lately Intel takes time until their latest greatest chips are readily available to buy so it makes more sense to compare it to the then closer upcoming AMD chips than vice versa.

 

[Vattila]

What exactly is 'zen aware' modern game code exactly?

I've used the term "Zen-aware code" loosely to mean code that doesn't make the Zen architecture choke. For example, Valve's old "Source" engine (e.g. used by CS:GO) is not Zen-aware. On the other end of the spectrum, I presume Id Software's "id Tech" engine behaves well on Zen, considering AMD has a partnership with Bethesda Softworks (Id Software's publisher).

"This is a multi-title agreement where AMD and Bethesda will work together to enable support for the Vulkan graphics API and tune Bethesda’s games for optimal performance on AMD’s Radeon and Ryzen hardware."

https://semiaccurate.com/2017/02/28/amd-partners-bethesda-gdc/

Basically, yeah, I'm sceptical AMD can completely eradicate the ST deficit in one generation, especially when it comes to gaming.

Fair enough. As you pointed out earlier in this thread we are not far apart in our expectations in percentage terms, and you concede that Zen 3 may get where I expect Zen 2 should be, which is basically down to roadmap planning. I guess you look at any IPC advance as a tide that lifts all boats, while I think they will target the weak spots in the architecture — the particular issues that lead to poor single-threaded performance in gaming, database and business applications.

Still, I hope I've given you some food for thought when it comes to AMD's business plans, the roadmap they need to succeed, and why I expect so much.

Memory latency badly affects Ryzen performance, particularly in those single-thread scenarios. A previous post highlighted how the latency reduction for Zen+ resulted in gaming performance beyond its relative latency improvement. This is surely the low hanging fruit that can bring the Zen 2 CPUs back to almost parity with Intel.

I think zen2 matching 9900k would depend on the IPC/PPC gain is composed of latency improvements that would benefit gaming more than general app performance...ie...'gaming ipc' is greater than 13% due to this reason

Good points. I expect AMD's efforts on IPC to be very targeted. We should see bigger than average improvement on game code, not less.

 

[Vattila]

the only game where Ryzen+ draws close to the CFL chips without there being an obvious 'fps wall' for the CFL chips would appear to be Star Wars Battlefront II

That makes sense. The game is based on EA DICE's "Frostbite 3" engine, and EA DICE and AMD have a history of cooperation. In particular, they developed the Mantle API together, which was later donated to the Khronos Group and became the foundation for Vulcan.

"Vulkan is derived from and built upon components of AMD's Mantle API, which was donated by AMD to Khronos with the intent of giving Khronos a foundation on which to begin developing a low-level API that they could standardize across the industry, much like OpenGL."

https://en.wikipedia.org/wiki/Vulkan_(API)

 

[PotatoWithEarsOnSide]

My guess is that if you're a 8700k owner that you'll be sitting on it until DDR5 and PCIe5 are available, unless the Ryzen 3700x is an absolute beast, which I doubt.
A 2700x owner is probably in the same boat except the 3700x would be a drop in replacement. But even then you'd want the difference in performance of the 3700x to be in excess of the difference in performance between a 8700k and a 9900k.
I just don't think that the 9900k does enough to consider buying, especially at its price point. If anything, I expect 8700k prices to go up as a result of the 9th gen products.

 

[TheELF]

I've used the term "Zen-aware code" loosely to mean code that doesn't make the Zen architecture choke. For example, Valve's old "Source" engine (e.g. used by CS:GO) is not Zen-aware. On the other end of the spectrum, I presume Id Software's "id Tech" engine behaves well on Zen, considering AMD has a partnership with Bethesda Softworks (Id Software's publisher).

CSGO has a fixed number of cores meaning that it will show differences in IPC PER CORE.
ID tech spreads out it's work to as many cores (threads really) as they need to,meaning it will show IPC PER SKU.
It's not any more or less zen-aware it just uses more of the CPU resulting in less of a FPS difference.

 

[Vattila]

The ["id Tech" engine is] not any more or less zen-aware [than the "Source" engine,] it just uses more of the CPU resulting in less of a FPS difference.

You are right that using more threads makes a big difference. However, "Source" shows ~32% deficit for the Zen architecture in CS:GO at 1080p (48% win for i9-9900K) according to the PT study. I doubt you would see the same deficit on modern code, like the "Frostbite 3" engine, even if you restricted it to the same number of threads as "Source" is able to use. Something is stalling badly for Zen when running "Source".

It is not that surprising, considering the age of the "Source" engine. It was developed long before Zen was conceived. Hopefully, CS:GO will be ported to "Source 2" as planned, and hopefully we will see better results.

 

[Vattila]

I would really like to see i9-9900K vs Ryzen 2700X benchmarked at the same fixed frequency, and at 720p, across games ranging from old titles like CS:GO to modern titles like Star Wars Battlefront II, to get a better feel for the IPC difference between the Core and Zen architectures in gaming.

Ideally, benchmarks would be repeated at multiple fixed CPU frequencies, so that we could deduce the GPU-bound part of the frame time, and thus estimate the true IPC difference (which, as discussed earlier, would be bigger than the difference seen in the benchmarks).

And, I would love to see in-depth exploration and analysis of the results to identify the spots in the Zen architecture that are to blame for the shortfall.

With that data and insight, we could then have an informed discussion about what AMD needs to do to overcome the gap in gaming IPC, and whether it is realistic without completely redoing the entire architecture.

With the IPC question settled, we could then argue whether Zen 2 on TSMC's 7nm HPC process can reach the necessary frequency for Ryzen 3000 to achieve gaming champion status.

PS. Please forward to independent testers.

 

[TheELF]

I doubt you would see the same deficit on modern code, like the "Frostbite 3" engine, even if you restricted it to the same number of threads as "Source" is able to use.

Yeah I would love to see this tested and not just benches between vastly different core numbers or core numbers that high that no GPU can keep pace with them no matter how low the resolution.
The closest you can come to this is comparing the i3-8350k to the 2200g but then everybody only screams cache...would love if some people here would actually test some games and share CPU utilization.

 

[Arkaign]

I would really like to see i9-9900K vs Ryzen 2700X benchmarked at the same fixed frequency, and at 720p, across games ranging from old titles like CS:GO to modern titles like Star Wars Battlefront II, to get a better feel for the IPC difference between the Core and Zen architectures in gaming.

Ideally, benchmarks would be repeated at multiple fixed CPU frequencies, so that we could deduce the GPU-bound part of the frame time, and thus estimate the true IPC difference (which, as discussed earlier, would be bigger than the difference seen in the benchmarks).

And, I would love to see in-depth exploration and analysis of the results to identify the spots in the Zen architecture that are to blame for the shortfall.

With that data and insight, we could then have an informed discussion about what AMD needs to do to overcome the gap in gaming IPC, and whether it is realistic without completely redoing the entire architecture.

With the IPC question settled, we could then argue whether Zen 2 on TSMC's 7nm HPC process can reach the necessary frequency for Ryzen 3000 to achieve gaming champion status.

PS. Please forward to independent testers.

This would be ideal. With 2080/ti dedicating so many transistors to non shaders it really didn't move performance all that much forward. GPU bottleneck in benches is severe. 720p minimum settings would really reveal what the performance potential and gap there is on CPU moving forward, and we could get a good idea of what type of leap we need from Zen2.

 

[CatMerc]

IPC doesn't have much weight in gaming. Zen's biggest deficiency in gaming is the high memory latency. Just look at how gaming gained disproportionally to most software with Zen+.

Larger caches, better pre-fetching, and getting memory latency as low as possible, are the defining factors.

You can profile games right now and you will notice how they tend to be heavily DRAM latency bound.

 

[happy medium]

IPC doesn't have much weight in gaming

Id rather have 20% more IPC than 20% less latency.
Where did you come up with this?
Link me.

 

[CatMerc]

Id rather have 20% more IPC than 20% less latency.
Where did you come up with this?
Link me.

IPC is directly impacted by latency. You can have 20% higher IPC from 20% lower latency if your program is 100% memory latency bound.

Games are on average around 35% bound on my Intel 8600K platform at tuned subtimings and 3400. Dropped to 2400 and default timings and the average is around 40%. Ryzen doesn't support VTune so it's harder to profile, but I'd estimate around 50-60% would be the average judging by the latency difference.

When I say IPC doesn't have much weight in gaming, I mean the computational grunt of the core has nearly nothing to do with it past a certain point. Skylake-X has similar gaming performance to Zen+, and its memory latency is also similar. But the core itself should be no less capable than client Skylake clock for clock.

 

[coercitiv]

Charlie from Semiaccurate says Intel decided to pull the plug on 10nm. Take it with a grain of salt for now, but Intel's reaction (or lack of) will soon be telling.

 

[DrMrLordX]

Charlie from Semiaccurate says Intel decided to pull the plug on 10nm. Take it with a grain of salt for now, but Intel's reaction (or lack of) will soon be telling.

Whaaaaaaat

 

[coercitiv]

Whaaaaaaat

We have a new thread on the subject, but if true the 9900K may become an interesting symbol even beyond it's gaming prodigy aura.

 

[LTC8K6]

Whaaaaaaat

That's been a very common theme here for a long time, that Intel could skip 10nm and go straight to 7nm.

 

[Vattila]

You can profile games right now and you will notice how they tend to be heavily DRAM latency bound. […] IPC doesn't have much weight in gaming.

As you corrected yourself in your next post, IPC is of course impacted by memory latency. Do you have any thoughts on where improvements can be made, e.g. cache hierarchy, load-store, DRAM controller?

Ryzen doesn't support VTune so it's harder to profile

Have you tried AMD μProf?

Charlie from Semiaccurate says Intel decided to pull the plug on 10nm.

Intel has denied this in a tweet and states that they are still "on track" to deliver 10nm product on the shelves for 2019-Q4.

Here is my take: 10nm has been relaxed according to SemiAccurate ("10nm--" or "12nm"), and I guess this is highly likely now. It seems that Intel is still pushing ahead to get Ice Lake-U/Y out on this process, in (somewhat) high-volume manufacture, for a notebook refresh before holiday 2019.

Further, I have a hunch SemiAccurate's cancellation claim has more to do with products than process. In particular, I suspect 10nm+ and the products planned for it, i.e. 10nm+ Ice Lake-S/X (as well as Ice Lake-SP for server), may have been cancelled, with full focus on the next-generation 7nm products instead (Granite Rapids etc.). This is just my speculation, though.

That's been a very common theme here for a long time, that Intel could skip 10nm and go straight to 7nm.

Yeah. If so, i9-9900K may be the best they can do for a long time — Intel's last hurrah in gaming for the foreseeable future, indeed.​