Coffeelake thread, benchmarks, reviews, input, everything.

[DrMrLordX]

You are missing a point. Practically all current game engines are more or less descendants of some ancient codes.

The hell they are. UE4 sure isn't. Nor is IDTech6. Or would you have us believed that IDTech6 is rehashed code from 2006?

It doesnt matter at all if its working.

Of course it matters! Go take a default install of Fallout 4 (yes, before using a settings tweaker to try to increase thread count, for what its worth). Compare it to a default install of BF1. Let's say on a 4c/8t processor, at 3 GHz. Now double the core count at the same uarch. Watch what happens to Fallout 4: practically nothing. BF1? In a 64-man multiplayer map, you probably get better performance.

Now throw Starcraft II into the mix, which struggles to tax more than two cores.

Hell there are some games out now that won't even run on old Phenom chips because they don't support the right SIMD instructions. Imagine that.

Nice example is AotS.

No it isn't. AotS results have never made any sense. There was some talk of the game making use of GPGPU at some point in its development, which I guess never happened. As a DX12 tech demo, it never really cut the mustard.

 

[epsilon84]

I am under the impression that once there is enough bandwidth, any more does not help. Better to reduce latency. 2-channel 2400 or 2666 MT/s appears to be the point of enough bandwidth, for i7-8700K.

I think you'll find the 8700K will scale well beyond DDR4-2666, especially in gaming. You can only reduce latency up to a certain point. Higher bandwith only makes no sense if you have to relax timings too much. However, using CL14 as a standard, we can use both DDR4-2666 CL 14 and DDR4-3466 CL 14, what do you think wins?

Of course at some point you will hit the point of diminishing returns. But regardless, the higher bandwith setup at the same timings will be faster.

 

[Bouowmx]

However, using CL14 as a standard, we can use both DDR4-2666 CL 14 and DDR4-3466 CL 14, what do you think wins?

That's not really a fair comparison though. The two kits have different bandwidth and latency in time: 10.5 vs 8.1 ns.

The comparison I'm interested in is 3200 MT/s C14 (8.75 ns), the basic low-latency kit, vs 2666 MT/s C12 (9.0 ns), the low-latency kit adjusted for Intel non-overclocking motherboards. This time, just testing bandwidth.

 

[Glo.]

I've got 2800 RAM on my ASRock Z370 ITX/ac with i3-8100, using XMP.

I know, but you should be able to get this on Z370 MoBo. What matters here is the possibility of getting 2666 MHz on H370 and B360 MoBo's combined with Core i3 8100 which has only 2400 MHz RAM controller.

If the word is true that we can get even higher on H370 and B360 - its done deal.

 

[TheLycan]

Memory bandwidth matters more than CL. This has been proven in the past years with ddr4. Also was proven that swwet spot for 6700k and 7700k is 3000 or 3200. Now its safe to asume that 8600k and 8700k need at least 3200, maybe 3600. Anything showing average fps is trash, one must to look at frametimes.

 

[eddman]

DDR4 3000 starts at 154$.

Well, yes, that's what I said. Also 3200 CL16 performs about the same as 3000 CL15.

My point was that very high frequency memories are quite expensive and not worth their price.

 

[IRobot23]

Well, yes, that's what I said. Also 3200 CL16 performs about the same as 3000 CL15.

My point was that very high frequency memories are quite expensive and not worth their price.

3000MHz @CL 15-16-16-35

vs 3200MHz @ CL 16 16 16 36

 

[CakeMonster]

Timings seem to go south above 3400ish. Last time I checked it seems CL16 was the sweet spot and that is around that speed.

 

[TheF34RChannel]

Timings seem to go south above 3400ish. Last time I checked it seems CL16 was the sweet spot and that is around that speed.

The sweet spot is the lowest timings you can get with a decent frequency of at least 3200MHz but not much, if any, higher - depending on your use (benches, gaming, content creation).

 

[IRobot23]

Memory bandwidth matters more than CL. This has been proven in the past years with ddr4. Also was proven that swwet spot for 6700k and 7700k is 3000 or 3200. Now its safe to asume that 8600k and 8700k need at least 3200, maybe 3600. Anything showing average fps is trash, one must to look at frametimes.

You do not understand.
CL is effecting bandwidth. Latency is also important. Less latency = less bandwidth needed... everything is related.

 

[DrMrLordX]

I don't know that latency reduction really moves bandwidth requirements in any particular direction.

In the simplest sense, bandwidth requirements are, at a minimum, based on the size of your dataset. That assumes that you are only going to pull data once, process it, and then spit out a result. Bandwidth requirements go up if you have to pull data multiple times to complete different calculations with the same data, assuming you can't cache it all.

In a best-case scenario your dataset fits somewhere in cache, so you just do a cache fill and roll with that. I don't think there is an absolute worst case (there are so many ways for things to go wrong), but lots of cache misses with pipeline stalls and other nonsense can get pretty bad, forcing multiple attempts to pull the same data from main memory due to repeated cache fills. With modern x86 CPUs, you often suffer badly from constantly having to go to main memory since the CPU's behavior can get pretty bad when it can't find what it needs in cache, repeatedly. Or at least that's my understanding of it. Someone better-informed could clean up my statements a bit, I think. I remember cache misses and thrashing being HORRIBLE on Con cores.

Regardless improving latency is only going to reduce memory bandwidth requirements if it somehow helps you to keep up on cache fills to reduce misses and commensurate thrashing. I'm not really sure that the RAM tweaking we do really affects real-world or even synthetic applications much in that manner.

It might be testable. I'm just not sure how one would record total bandwidth usage completing a given task in a high-level programming language. If all you're doing is repeating a simple operation on a large data set (like a[x] + b[y] or whatever) all you can really tell is the data size of each element, and how often you make the request. Recording the actual amount of data read from or written to memory during that operation is another matter entirely. There may be libraries for that? Clearly applications like Aida do this in some fashion, but I've never worked on a project like that. And Aida64 only tells you how many bytes/second the CPU CAN read to/write from memory, it doesn't tell you how many it actually reads from memory or writes to memory while performing a specific calculation or set of calculations.

 

[IRobot23]

I don't know that latency reduction really moves bandwidth requirements in any particular direction.

In the simplest sense, bandwidth requirements are, at a minimum, based on the size of your dataset. That assumes that you are only going to pull data once, process it, and then spit out a result. Bandwidth requirements go up if you have to pull data multiple times to complete different calculations with the same data, assuming you can't cache it all.

In a best-case scenario your dataset fits somewhere in cache, so you just do a cache fill and roll with that. I don't think there is an absolute worst case (there are so many ways for things to go wrong), but lots of cache misses with pipeline stalls and other nonsense can get pretty bad, forcing multiple attempts to pull the same data from main memory due to repeated cache fills. With modern x86 CPUs, you often suffer badly from constantly having to go to main memory since the CPU's behavior can get pretty bad when it can't find what it needs in cache, repeatedly. Or at least that's my understanding of it. Someone better-informed could clean up my statements a bit, I think. I remember cache misses and thrashing being HORRIBLE on Con cores.

Regardless improving latency is only going to reduce memory bandwidth requirements if it somehow helps you to keep up on cache fills to reduce misses and commensurate thrashing. I'm not really sure that the RAM tweaking we do really affects real-world or even synthetic applications much in that manner.

It might be testable. I'm just not sure how one would record total bandwidth usage completing a given task in a high-level programming language. If all you're doing is repeating a simple operation on a large data set (like a[x] + b[y] or whatever) all you can really tell is the data size of each element, and how often you make the request. Recording the actual amount of data read from or written to memory during that operation is another matter entirely. There may be libraries for that? Clearly applications like Aida do this in some fashion, but I've never worked on a project like that. And Aida64 only tells you how many bytes/second the CPU CAN read to/write from memory, it doesn't tell you how many it actually reads from memory or writes to memory while performing a specific calculation or set of calculations.

I was just an example.
Latency matters, bandwidth matter.

DDR4 4200MTs 19 26 26 46, might/will score lower than 3600 16 16 16 38 @ AIDA64.

 

[TheLycan]

You do not understand.
CL is effecting bandwidth. Latency is also important. Less latency = less bandwidth needed... everything is related.

You are missing a point. CL is timing. For example how many buses pass through a gate in 1 min. Bandwith is how many passengers a bus have. Games seems to scale more with bandwith, this is why you dont see 2133 in reviews.

Yes, CL has an impact in performance but its contribution is far less than a high bandwith.

 

[IRobot23]

You are missing a point. CL is timing. For example how many buses pass through a gate in 1 min. Bandwith is how many passengers a bus have. Games seems to scale more with bandwith, this is why you dont see 2133 in reviews.

Yes, CL has an impact in performance but its contribution is far less than a high bandwith.

We get that. MT/s gives you maximal theoretical bandwidth, CL gives you latency. Bad latency = bad bandwidth, you will not achieve same bandwidth with slow worst CL, you wont be close.

Since IMC can be inefficient at higher clocks, you will hardly reach same bandwidth and latency.
Lets say DDR4 3200MT/s CL 14 14 14 34 or DDR4 3600 CL 18 19 19 39. Try it out yourself, you will see.

 

[DrMrLordX]

Bad latency = bad bandwidth

Untrue, especially for non-JEDEC memory standards. RDRAM comes to mind.

 

[slashy16]

What do you guys think the timeline will be for the 8-Core Coffee lake CPU? Does Intel launch it before Christmas or in the first months of next year?

http://www.pcgamer.com/coffee-lakes-motherboard-compatibility-hints-at-intel-releasing-8-core-cpus/

 

[TheF34RChannel]

What do you guys think the timeline will be for the 8-Core Coffee lake CPU? Does Intel launch it before Christmas or in the first months of next year?

http://www.pcgamer.com/coffee-lakes-motherboard-compatibility-hints-at-intel-releasing-8-core-cpus/

9 months from now, this is not news and has been known for a while and I am now convinced there'll be an 8C CFL-S, coinciding with Z390.

 

[slashy16]

There is stock of the 8400/8600 at Frys for local pickup across the US. I just had a buddy in Concord California snag a 8600K for me.

https://www.frys.com/product/9339730?site=sr:SEARCH:MAIN_RSLT_PG

 

[moonbogg]

9 months from now, this is not news and has been known for a while and I am now convinced there'll be an 8C CFL-S, coinciding with Z390.

I think a lot of people buying 8700K believe it represents Intel's big step forward. Some of us realize Intel must release an 8 core in order to stay relevant in the near future, but I'd argue that most people buying 8700K actually think they are getting the next big step, but they are mistaken. The 8 core will come soon and if it comes with a new Z390 board, then oh man that will be so sexy.
8700K is amazing and I was temped to ditch my stuff and get one. Had they been in stock my will power may have lost that fight and my sig might look different right now. I had time to cool down and realize that an 8 core must be coming shortly. I think at least a handful of 8700K buyers might wish they had waited. They likely didn't know about the 8 core chip though or people told them it wasn't true or something.

 

[Arachnotronic]

I think a lot of people buying 8700K believe it represents Intel's big step forward. Some of us realize Intel must release an 8 core in order to stay relevant in the near future, but I'd argue that most people buying 8700K actually think they are getting the next big step, but they are mistaken. The 8 core will come soon and if it comes with a new Z390 board, then oh man that will be so sexy.
8700K is amazing and I was temped to ditch my stuff and get one. Had they been in stock my will power may have lost that fight and my sig might look different right now. I had time to cool down and realize that an 8 core must be coming shortly. I think at least a handful of 8700K buyers might wish they had waited. They likely didn't know about the 8 core chip though or people told them it wasn't true or something.

It's a year off. If you can wait a year, great, but if you want/need a big jump in performance from an older system or are a first time system builder/buyer, 8700K is a great choice.

 

[moonbogg]

It's a year off. If you can wait a year, great, but if you want/need a big jump in performance from an older system or are a first time system builder/buyer, 8700K is a great choice.

Its actually the only choice if you ask me. What else is there to buy? Not a single thing IMO. The only down side IMO is that soon an 8 core will come. That's not a typical situation for us to find ourselves in. Usually we know a new CPU is coming, but its always been just another boring quad core, so it was never a big deal. But now we have an 8 core coming hardly a year (maybe sooner?) right after Intel's first 6 core drops? That's very strange indeed and I fully expect some people will wish they had waited, particularly if their rig is working fine now.

 

[DrMrLordX]

We should probably be glad that the past 7 months have given us so many improvements in desktop PC processing power, regardless of the manufacturer. The 8700k is just another step in that direction. It ain't over yet, either.

9 months from now, this is not news and has been known for a while and I am now convinced there'll be an 8C CFL-S, coinciding with Z390.

Yup, June/July 2018.

 

[DrMrLordX]

double post

 

[TheF34RChannel]

I think a lot of people buying 8700K believe it represents Intel's big step forward. Some of us realize Intel must release an 8 core in order to stay relevant in the near future, but I'd argue that most people buying 8700K actually think they are getting the next big step, but they are mistaken. The 8 core will come soon and if it comes with a new Z390 board, then oh man that will be so sexy.
8700K is amazing and I was temped to ditch my stuff and get one. Had they been in stock my will power may have lost that fight and my sig might look different right now. I had time to cool down and realize that an 8 core must be coming shortly. I think at least a handful of 8700K buyers might wish they had waited. They likely didn't know about the 8 core chip though or people told them it wasn't true or something.

I purchased mine whilst being fully aware of the 8C; I need the performance now and not down the road - truth be told, as a gaming enthusiast I don't need 8 cores anytime soon and am better off with 6 higher frequency cores. Which, of course, does not mean that I won't drop an 8C in the board simply because I enjoy new advancing technology

Those buying a 6 core right now shouldn't regret it; 9 months is a long time to go.

Although mine's been running less than a day I can genuinely say it's the best CPU I've ever had, thus validating my purchase.

 

[moonbogg]

I purchased mine whilst being fully aware of the 8C; I need the performance now and not down the road - truth be told, as a gaming enthusiast I don't need 8 cores anytime soon and am better off with 6 higher frequency cores. Which, of course, does not mean that I won't drop an 8C in the board simply because I enjoy new advancing technology

Those buying a 6 core right now shouldn't regret it; 9 months is a long time to go.

Although mine's been running less than a day I can genuinely say it's the best CPU I've ever had, thus validating my purchase.

I wouldn't personally regret an 8700K either. Its still going to be sick as hell in 9 months, lol. Gaming benefits of 8 cores vs 6, with all else being equal, are rapidly diminishing at the moment. There will still be quite the handful of people who feel some remorse though. Just the way it goes I guess.