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

[TheGiant]

I am curious what Icelake can bring to the table. Some reports on cannonlake show like 5% IPC increase. That is not that bad (not good either).
IMO Intel has heat transfer problems with much lesser area of the chips with 10nm and that is why they can not increase frequency to the optimized 14++more+ levels.
AMD should increase prices. To get the budget for R&D as fast as they can. Maybe Jim Keller will do something and in 2020 Intel does the Conroe jump and "there and back again, a CPU tale by Intel"

 

[tamz_msc]

There are a number of factors which hinder gaming performance on Ryzen, besides IPC deficit and clock speeds. These are memory latency, cross-CCX latency and L3$ being a victim cache. Unless these all are addressed in Zen2, Intel has no reason to worry losing the best gaming-CPU crown.

 

[maddie]

There are a number of factors which hinder gaming performance on Ryzen, besides IPC deficit and clock speeds. These are memory latency, cross-CCX latency and L3$ being a victim cache. Unless these all are addressed in Zen2, Intel has no reason to worry losing the best gaming-CPU crown.

Isn't IPC deficit at least partially caused by these, "memory latency, cross-CCX latency and L3$ being a victim cache", and not separate issues.

What was the word when Zen launched? Balance.

 

[Tech001]

My opinion as to why Zen 2 won't hit 5GHz:
1. Fresh TSMC 7nm process.
I suspect TSMC shipped more 7nm CPU's in the first day after the latest iPhone launch than Intel's 10nm in the 9 months or so, since they first "shipped" 10nm. By the time Zen launches, 7nm will be fairly mature in TSMC terms. In the last few years TSMC seem to have a pretty good record of delivering brand new processes and the ability to deliver large volumes of chips to the market.

2. Precedent. Has a TSMC process ever hit 5GHz before? Ever?
Oracle managed to ship a 32-core, >10 billion transistor SPARC M8 chip that runs at > 5Ghz well over a year ago, and that was running on TSMC's 20nm so I doubt whether TSMC process will be the limiting factor. On that same 20nm process Oracle also shipped the M7 32 Core 4GHz, and the S7 8 core 4.2Ghz so I don't think the high clock speed was a one hit wonder for TSMC.

3. Is AMD going to be able to successfully marry the zen 2 arch with this fresh TSMC process to produce a 5GHz-capable chip out of the gates?
Can't see why not - can't help thinking that the move from GF to TSMC may end up working out very well for AMD,

4. A process shrink means smaller cooling surface area. High frequency and voltage will mean concentration of heat in a reduced core area which will need to be moved as quickly as possible. This will pose its own challenges.

5. Is there any other process out there, besides Intel's, that's consistently hitting 5GHz on air cooling?
Yes, TSMC's 4 year old 20nm !

 

[tamz_msc]

Isn't IPC deficit at least partially caused by these, "memory latency, cross-CCX latency and L3$ being a victim cache", and not separate issues.

What was the word when Zen launched? Balance.

Not necessarily. It depends on what kind of IPC is being measured. Most applications aren't really affected by these parameters as they're not usually latency bound. Games are latency bound, and hence the difference is wider in games. For example, according to hardware.fr, the 1300X is able to match a 4670K in h.264 encoding. This is mostly a compute-bound scenario and Ryzen has similar IPC to Haswell. In games however, the 4670K is 7% ahead, a difference which can only be explained by the deficiencies in the parameters described earlier.

Without a radical departure from the Zen design in Zen2, those deficiencies would likely remain and prevent AMD from taking the gaming performance crown.

 

[french toast]

I think there is near to zero chance of AMD taking the outright performance crown from the 9900k..it clocks too high.
But they can bring comparable performance down the stack and beat intel on price ...making them the defacto choice for gamers, even if you are spending £2000 on a PC you are still going to be price concious, if you can get 9900k type performance for £100 less then most will take it, especially if that allows a better case, SSD or GPU.

I can't see how Intel can improve on 9900k performance for 2019, most probably they will just drop the prices down to what they should have been, I think AMD will be able to close the 12% gap for $350.

 

[Vattila]

Wow! After that IPC rumour, and perhaps after my passionate plea for setting some sound expectations for AMD, the poll has changed remarkably — now the results show the same 42% for and against the proposition, while 16% are undecided.

Thanks for all the feedback. It is interesting to read and get a feel for the reasons behind the scepticism. However, I would like more of you to address my point about AMD's planning, and provide some reasoning why so many expect AMD to fail to execute to those plans.

To reiterate — AMD's plans would have to be ambitious to make any sense. In particular, it is obvious that Zen 2 would have to target 10nm Ice lake performance. Otherwise, they would have had a very hard time getting any traction in the server space (which is glacially slow, even with a good product, such as Zen-based EPYC). Years back when decisions about business strategy and the roadmap were made, AMD could not foresee that Intel would stumble at 10nm. The Intel 10nm process was projected to have substantial performance, power and density advantages.

For Zen, as a first step, AMD targeted Haswell performance and as such set a 40% target for IPC uplift. They overshot that by 30%, reaching 52% and surpassing Broadwell performance. For Zen 2, they had bigger ambitions, aiming to compete in a bigger part of the server market, and targeted 10nm Ice Lake performance. To make that happen they went "all in" on 7nm to enable the performance and efficiency they needed (lookup AMD CTO Mark Papermaster's insightful comments on this).

Their big bet on 7nm included renegotiating the wafer-supply agreement with GlobalFoundries and working closely with both GF and TSMC on the needs and requirements they had for the process. In the end, TSMC won the contract, and GF threw the towel into the ring. So it makes no sense to assume that TSMC's 7nm HPC process should be the inferior of the two. As a reminder, GF targeted "5 GHz operation" as per CTO Gary Patton's statements. As for TSMC's 7nm HPC process and their entry into the server space, AMD will be their biggest and most important customer, as far as I understand. AMD is as such a key partner. Expect that they have been instrumental in formulating the process to suit their needs.

So — back to the thread topic — the low expectations for Zen 2 are surprising.

Also, note that i9-9900K only has to lose in a few key titles for its superior status as gaming champion to be called into question. If Zen 2 really gets a 13% IPC increase in scientific workloads, where Zen already does well, it may translate to bigger wins in gaming, in which Zen has had comparatively mediocre results. Ryzen 3000 may end up winning in more recent titles based on modern game engines, while i9-9900K wins in titles using legacy engines.

So, at best, I expect i9-9900K to be demoted to "disputed champion" when Ryzen 3000 arrives. At worst, which is more likely, provided AMD executes to plans, it will only win the extreme outliers, such as CS:GO.

If Ryzen 3000 underperforms my expectations, I will consider that a failure of execution on AMD's part and a sign that they fell well short of target. It may not matter much, now that Intel has stumbled so badly with 10nm and their roadmap — but had they not, AMD would be toast with a Zen 2 core failing to match Skylake performance levels.

Lisa Su's career won't depend on beating the i9-9900k in gaming benchmarks. […] AMD doesn't need to smash Intel into the dust, they need to offer a compelling product and so far they've done that.

If Intel had been on 10nm Ice Lake performance levels by now, AMD would need a Zen 2 core that could compete with that. Otherwise, AMD would have been in great peril, and Lisa Su would have been shown the door.

Will Zen2 really hit 5ghz though? That's really Intel's big advantage right now. I see a bunch of people in this thread claiming it's a foregone conclusion but is there data to support that conclusion?

We have statements from GF that they targeted 5 GHz operation. TSMC won the AMD contract, likely mostly to schedule. But I think it is unlikely that AMD chose to go all in with TSMC if they had to take a 10% hit on their performance targets.

Your calculations are seriously flawed because gaming performance isn't a linear extrapolation of clockspeed * IPC.

Ok. Educate me. Show me Zen vs Skylake/CFL IPC deficit at the moment in scientific workloads and how that translates to current IPC deficit in games. If you have links, I'd be grateful.

Anyway, since this is a prediction thread, I'll throw mine out there - unless Zen 2 can clock at 5GHz+ like a 9900K, I don't think it will match it, let alone beat it, in gaming.

As I pointed out above, Ryzen 3000 doesn't have to beat in every title for i9-9900K to lose its status as champion. It may not even have to beat the PT study average. Independent tests may set a lower bar. And there may be clear discrepancies between modern titles and legacy titles.

My opinion as to why Zen 2 won't hit 5GHz [is that TSMC will have teething issues and problems reaching 5 GHz].

It is a good point. Process technology progress is hard. They may stumble. But do not expect it. And be sure that TSMC is well aware of AMD's needs for their roadmap. And AMD has made their big bet on them delivering.

I agree, that extrapolating 13% IPC increase as a similar increase in gaming per-clock is absurdly optimistic.

It is not. In fact, as Abwx points out, games saw the bigger increase from Zen to Zen+. As I pointed out above, game code, and in particular legacy game code, has run uncharacteristically poorly on Zen compared to general workloads. Any improvement in the architecture that alleviates this issue (such as improving memory and inter-CCX latency, as you point out) may show disproportional gains in game code.

With GF's 7nm HP process, it might have been possible [to take the gaming crown]. On TSMC's 7nm, I expect 4.5-4.6 GHz.

Don't dismiss TSMC as a low-performance device manufacturer. They are aiming for leadership in HPC and to compete squarely with Intel. AMD chose them ahead of GF.

I'm just going to throw this out there that AMD is already winning, but not for the reasons in the poll. [Due to their SoC methodology and chiplet strategy,] AMD can make chips that are 90% of Intel perf and offer them from launch at 25-40% cheaper than Intel can and STILL make money hand over fist because their yields are so good. That's game.

It is true that performance is not the only factor. However, without Zen 2 performance competitive with Ice Lake (which would have been here by now, if Intel didn't stumble), AMD could just go home in the high-end x86 fight. Server customers wouldn't bat an eyelid, and AMD would be better off continuing working on their low-end Jaguar core, K12 and spearheading a niche in low-cost ARM servers.

 

[epsilon84]

Ok. Educate me. Show me Zen vs Skylake/CFL IPC deficit at the moment in scientific workloads and how that translates to current IPC deficit in games. If you have links, I'd be grateful.

Firstly, no, I have no idea how you can use some vague 13% increase in 'scientific' workloads and somehow extrapolate that into gaming performance.

So on to the point of why I think your clockspeed * IPC analysis is flawed when it comes to predicting the % uplift in gaming performance...

Let's say hypothetically there are 10000 frames rendered in a 100 second benchmark, for an average framerate of 100fps. At any given moment, the framerate has to either be CPU limited, or GPU limited, otherwise we would have an inifinite framerate, correct?

Let's assume we're using a low resolution to be more CPU bound, and the test is 50% CPU limited and 50% GPU limited. This means that a faster CPU will only improve the framerates in the portion of the benchmark that is CPU limited. The GPU limited portion of the benchmark maintains the same framerate regardless of CPU speed.

Therefore, any % increase in CPU gaming performance (through frequency and IPC gains) won't directly translate to the same % increase in actual framerates, because the GPU is just as influential in the final outcome of the benchmark results, even in something considered a 'CPU bound' scenario. I think this is where people may get confused a bit - 'CPU bound gaming' doesn't mean that every single frame rendered has the CPU as the limiting factor.

I'd say that in reality, most gaming benchmarks are actually more than 50% GPU limited, but it was just a hypothetical scenario that I tried to explain above, and 50% was an easy, simple figure to use.

Unless we use some absolutely absurd resolution like 640 x 480 on a 2080 Ti, it's basically impossible to make a gaming benchmark totally CPU bound (or close enough) to the point that it scales linearly with increased clockspeeds, or IPC.

Just to drive home my point, please consider the % improvements in actual framerates, compared to the % increase in clocks: https://static.techspot.com/articles-info/1613/bench/Average.png

8700K @ 5.2GHz 169 / 129
8700K @ 4.3GHz 163 / 125

So we're looking at a 3 - 4% gain in framerates from a 21% increase in frequency.

Why doesn't performance increase more from the substantially higher frequency? Because even at 1080P on a 1080 Ti, the framerates are still largely influenced by the GPU, despite it being known as a 'CPU bound' resolution.

As I said earlier, the only time you can truly get an (almost) 100% CPU bound scenario in gaming is to use totally unrealistic, archiac resolutions like 640 x 480. This has actually been done, for example: https://www.hardocp.com/article/2017/10/05/intel_coffee_lake_core_i58600k_vs_7600k_at_5ghz_review/4

This kind of data is useful from a purely scientific point of view, to analyse just exactly how fast a CPU can run gaming code, but in reality we will never see such big differences between CFL and Ryzen because of the points I listed earlier.

 

[Vattila]

[Since there is a part of the frame time that is not affected by CPU speed], any % increase in CPU gaming performance (through frequency and IPC gains) won't directly translate to the same % increase in actual framerates

Understood — thanks for clarifying. This is how I was thinking about it as well, similar to Amdahl's Law about the limitations for multithreaded speedup (where the serial part of the workload is unaffected by parallelism, and hence sets a limit on the overall speedup).

Note however that my calculation of 24% pure CPU uplift well exceeds the 16% average lead in the PT study, which I used as an arbitrary bar to declare Ryzen 3000 a winner (although as explained before, I think the actual fair bar is far lower). Don't you think 24% is sufficient? What is your hunch?

Regardless, I am more interested in why you are sceptical that AMD will reach Skylake level of performance with Zen 2. Do you agree with me that Ice Lake levels of performance would have to be the target they set? If no, why? And if yes, why do you feel they will fall flat?

 

[mikk]

Unless we use some absolutely absurd resolution like 640 x 480 on a 2080 Ti, it's basically impossible to make a gaming benchmark totally CPU bound (or close enough) to the point that it scales linearly with increased clockspeeds, or IPC.

Someone has no clue. This is a 5:4 or 4:3 resolution which is nonsense because CPU load is way lower than a common widescreen resolution like 16:9. 720p noAA is the way to go if you want a CPU bound test, of course a fast GPU should be used. There is no need for a lower resolution.

 

[Zucker2k]

Understood — thanks for clarifying. This is how I was thinking about it as well, similar to Amdahl's Law about the limitations for multithreaded speedup (where the serial part of the workload is unaffected by parallelism, and hence sets a limiter on the overall speedup).

Note however that my calculation at 24% well exceeds the 16% average from the PT study, which I used as an arbitrary bar to declare Ryzen 3000 a winner (although as explained before, I think the actual fair bar is far lower). Don't you think 24% is sufficient? What is your hunch?

Regardless, I am more interested in why you are sceptical that AMD will reach Skylake level of performance with Zen 2. Do you agree with me that Ice Lake levels of performance would have to be the target they set? If no, why? And if yes, why do you feel they will fall flat?

I'm quoting @tamz_msc post below because it addresses most of what you're overlooking. Basically, it took Intel switching from an inclusive l3 cache to an exclusive one to expose the importance of this component to game code. The fundamental issue here, besides the cache structure (ring), is latency. So, this little discovery from a year and a half ago with the launch of Skylake-X points to the benefits of an inclusive l3 cache in gaming. This is why it's virtually impossible for AMD to close that small, yet significant gap, save for a fundamental change in game code.

There are a number of factors which hinder gaming performance on Ryzen, besides IPC deficit and clock speeds. These are memory latency, cross-CCX latency and L3$ being a victim cache. Unless these all are addressed in Zen2, Intel has no reason to worry losing the best gaming-CPU crown.

 

[Vattila]

Basically, it took Intel switching from an inclusive l3 cache to an exclusive one to expose the importance of this component to game code. The fundamental issue here, besides the cache structure (ring), is latency. So, this little discovery from a year and a half ago with the launch of Skylake-X points to the benefits of an inclusive l3 cache in gaming. This is why it's virtually impossible for AMD to close that small, yet significant gap, save for a fundamental change in game code.

Thanks for the point. Ok. So your thinking is that AMD will focus on other workloads where they do well and be content with being second-best in gaming? If so, fair enough. That said, I am not qualified to evaluate the claim here, but I am doubtful that the cache structure is such a fundamental barrier for AMD (seeing how great Ryzen performs in recent titles). Also, AMD is keen to state their focus on the gaming market — one of the few lucrative and growing PC segments — so I am inclined to think they will invest to be best.

 

[Thunder 57]

AMD should increase prices. To get the budget for R&D as fast as they can. Maybe Jim Keller will do something and in 2020 Intel does the Conroe jump and "there and back again, a CPU tale by Intel"

Not sure why people (not saying you) turn Jim Keller into a deity who is the only one who can design excellent CPU's. Do we even know what he's up to at Intel?

There are a number of factors which hinder gaming performance on Ryzen, besides IPC deficit and clock speeds. These are memory latency, cross-CCX latency and L3$ being a victim cache. Unless these all are addressed in Zen2, Intel has no reason to worry losing the best gaming-CPU crown.

I think the first two can be improved upon. They may also increase the CCX size which would alleviate a lot of the issues at the cost of extra complexity. As for the victim cache, how much longer does Intel keep the same cache structure in their mainstream parts? Surely the small L2 is becoming a problem or else they would not have changed it with Skylake-X. Once Intel increases the L2 size, they will almost certainly have to go to a victim L3 as well. The 9000 series may be the last hurrah in that regard.

 

[Vattila]

Do we even know what he's up to at Intel?

I have a separate speculation thread on that actually:

Speculation: Keller's role at Intel

 

[Zucker2k]

Not sure why people (not saying you) turn Jim Keller into a deity who is the only one who can design excellent CPU's. Do we even know what he's up to at Intel?



I think the first two can be improved upon. They may also increase the CCX size which would alleviate a lot of the issues at the cost of extra complexity. As for the victim cache, how much longer does Intel keep the same cache structure in their mainstream parts? Surely the small L2 is becoming a problem or else they would not have changed it with Skylake-X. Once Intel increases the L2 size, they will almost certainly have to go to a victim L3 as well. The 9000 series may be the last hurrah in that regard.

Basically, the core wars meant Intel had optimize the cache structure to support far greater core counts. But Intel has room certainly for a 10-core, possibly 12-core on desktop. Remember the 6950x? That's a 10-core 20-thread chip with a 25MB inclusive cache.

 

[frozentundra123456]

Not sure why people (not saying you) turn Jim Keller into a deity who is the only one who can design excellent CPU's. Do we even know what he's up to at Intel?



I think the first two can be improved upon. They may also increase the CCX size which would alleviate a lot of the issues at the cost of extra complexity. As for the victim cache, how much longer does Intel keep the same cache structure in their mainstream parts? Surely the small L2 is becoming a problem or else they would not have changed it with Skylake-X. Once Intel increases the L2 size, they will almost certainly have to go to a victim L3 as well. The 9000 series may be the last hurrah in that regard.

The hype about Keller since he went to Intel has been very restrained compared to what it was from the AMD camp when he went there.

 

[Thunder 57]

Basically, the core wars meant Intel had optimize the cache structure to support far greater core counts. But Intel has room certainly for a 10-core, possibly 12-core on desktop. Remember the 6950x? That's a 10-core 20-thread chip with a 25MB inclusive cache.

I don't expect a mainstream 10 core chip anytime soon. They could probably do up to a 12 core with the current structure, but eventually they will need to shift to a non-inclusive L3 and probably drop the ring bus too.

The hype about Keller since he went to Intel has been very restrained compared to what it was from the AMD camp when he went there.

Very true, that hype when he was at AMD was unreal, and honestly, annoying.

 

[jpiniero]

The problem with Skylake-X with gaming appears not to be the switch to exclusive but rather that the L3 latency is slow due to being on the mesh.

Pretty confident Intel increased the L2 because of AVX-512 (and nothing else), and probably is increasing it in Icelake for the same reason.

 

[Thunder 57]

The problem with Skylake-X with gaming appears not to be the switch to exclusive but rather that the L3 latency is slow due to being on the mesh.

Pretty confident Intel increased the L2 because of AVX-512 (and nothing else), and probably is increasing it in Icelake for the same reason.

I keep hearing that but I haven't seen any proof. A mesh should offer the same or better performance at the cost of power and die size. Of course with it being their first go at it, that may not be the case.

 

[Kenmitch]

Kenstradomus prediction.

Out of the ashes a new king shall arise.
The mighty one will stand dumfounded.
Few will weep, many will rejoice.
The new king shall rule forever more.

I'm not sure on the last verse as the smoke was getting in my eyes....lol

Not even sure if this prediction is for this situation in the end. Eventually it'll come true.

 

[epsilon84]

Understood — thanks for clarifying. This is how I was thinking about it as well, similar to Amdahl's Law about the limitations for multithreaded speedup (where the serial part of the workload is unaffected by parallelism, and hence sets a limiter on the overall speedup).

Note however that my calculation at 24% well exceeds the 16% average from the PT study, which I used as an arbitrary bar to declare Ryzen 3000 a winner (although as explained before, I think the actual fair bar is far lower). Don't you think 24% is sufficient? What is your hunch?

Regardless, I am more interested in why you are sceptical that AMD will reach Skylake level of performance with Zen 2. Do you agree with me that Ice Lake levels of performance would have to be the target they set? If no, why? And if yes, why do you feel they will fall flat?

I actually think Zen 2 would come pretty close IPC wise to Skylake when it comes to gaming. A lot also depends on how AMD much AMD can improve latencies as well. As it stands, I believe Zen+ is about 10% behind per clock, or thereabouts. I'm not sure if it can actually overtake Skylake at gaming per clock, I don't have a crystal ball. But I'm not even going to try to extrapolate an estimate from an unrelated and unknown scientific workload and try to guess how much effect that has on actual gaming performance.

The main thing I'm more skeptical about is the ability of Zen 2 to clock to 5GHz. I guess even 4.5GHz with Skylake level IPC would bring Zen 2 'close enough' (within 5% actual in game framerates) to a 9900K that it doesn't really matter what CPU you choose for gaming. But to actually exceed a 9900K like you confidently predicted in the OP and subsequent posts, I have serious doubts. As I said, its one thing to draw close, its another thing to beat a 9900K in the majority of the games to declare the 9900K 'dethroned' as the 'best gaming CPU'.

I don't think we are *that* far off in our expectations - we're only talking about a 10% swing. I think Zen 2 would probably be 90 - 95% 9900K level whereas you think it would be 100 - 105% 9900K level? Would that be a correct assumption on my part?

 

[Vattila]

I suspect TSMC shipped more 7nm CPU's in the first day after the latest iPhone launch than Intel's 10nm in the 9 months or so, […] TSMC seem to have a pretty good record of delivering brand new processes and the ability to deliver large volumes of chips to the market. […] Oracle managed to ship a 32-core, >10 billion transistor SPARC M8 chip that runs at > 5Ghz well over a year ago, and that was running on TSMC's 20nm so I doubt whether TSMC process will be the limiting factor. […] can't help thinking that the move from GF to TSMC may end up working out very well for AMD

Great comments — a very good maiden post, if I may say so!

Tip: Please edit your post. Your comments all ended up in the quote box, looking like you didn't provide any new content at all.

 

[Vattila]

I don't think we are *that* far off in our expectations - we're only talking about a 10% swing. I think Zen 2 would probably be 90 - 95% 9900K level whereas you think it would be 100 - 105% 9900K level? Would that be a correct assumption on my part?

Yeah, that's fair. Considering how much Intel has squeezed out of Skylake and 14nm, anything slightly beyond i9-9900K performance probably would have competed decently against the initial 10nm Ice Lake as originally envisioned.

It may be the case that AMD has priorities regarding workloads and architecture that mean that gaming will see less uplift, making your more modest expectations more realistic.

On the other hand, we have statements from AMD executives that they are aware of the server workloads in which Zen performs relatively poorly, such as database and business applications (this is what makes the enterprise segment of the server market the hardest one to crack so far). I suspect the architectural issues regarding memory and cache-coherence latency, the same issues that affect game code, are at play here as well. AMD has stated that they want to improve in this field with Zen successors, so that they can address a larger part of the server market. This may benefit game code.

The main thing I'm more sceptical about is the ability of Zen 2 to clock to 5GHz.

AMD would have had to plan for that. It clearly looked to go that way with Intel's roadmaps and the 10nm projections. As Tech001 points out though, TSMC has done 5 GHz before. So has AMD.

 

[french toast]

A couple of people have mentioned that tsmc N7 is not hitting it's frequency targets...has anyone got a link or info for that?...I can't find anything on that, we only have A12 and Kirin 980 to go on and that could just be their preferred optimisation.

 

[Trumpstyle]

No Amd won't beat intel, Intel has 18% IPC advantage when it comes to gaming and a decent clock advantage.

Compare intels i7-8700k (6/12 core cpu) with amds 2600x (6/12 core cpu) and you see 18% IPC advantage for intel in gaming. Those charts show 2,8ghz clocked with turbo-boost disabled.

Only reason why we see some games where AMD are close in benchmarking is because of gpu limits nothing else.