FRONTPAGE Intel Skylake-X (i9 7900X) and Kaby Lake-X (i7 7740K) CPU Review

[deeppow]

Having delayed an update for quite awhile, I was planning to update to the x-cpus and mobos. Maybe I'll wait till the smoke clears from all this. New motherboards etc. are always an issue early on in their development but this seems a little more than normal.

 

[EarthDog]

Just get one with a robust vrm and a decent heatsink. Be sure to have good airflow. Really ismt much as you will be thermally limited by the cpu unless you are looking to delid.

 

[Woomack]

Delidding these chips looks like a pain considering all what is on the PCB. I also doubt that most users who paid for 8c+ will even think about delidding.
I'm still not sure what about motherboards. VRM cooling can be an issue but power connectors or power phases shouldn't. Most boards have 8 phases or more and even these cheapest boards have power design for 500W+. So hot power cables can be issue on the PSU side. Of course it's still better to get a board with 8+8 or 8+4 CPU power connectors but I'm just not sure if it's really required while overclocking ( at least on ambient temps ).

 

[Woomack]

What I see is that memory performance again depends more on the cache frequency than memory settings. On my board 3000 15-16-16, 3600 17-18-18 and 3600 15-15-15 have almost the same bandwidth in AIDA64 ( ~88/85/82GB/s read/write/copy ). Once I set 500MHz higher cache then bandwidth goes up by ~4GB/s. Next 500MHz and I already see 95GB/s.

So far I could set 3600 15-15-15 quite easy but 3733 is not booting on my memory regardless of settings. Maybe it's BIOS or something else as board could boot at 3600 even when by mistake I used settings which shouldn't work.

Here is result at 3600 15-15-15 and cache at 3300MHz.



One more weird thing. Results in 3DMark/PCMark10/VRMark at 3000 XMP/CR2 are about the same or better than at manual 3600/CR1.

I have ASUS TUF Mark 2 which was the cheapest board I found in local stores. When CPU is not overclocked ( still boosts up to 4.5GHz ) then VRM temp is below 50°C during all 3DMarks, PCMarks, Cine15 and some other tests. Max registered on stock CPU+GTX1080Ti in all tests was 322W (total whole platform, registered by TP-Link Smart Plug ). I had no time to play with CPU overclocking yet.

 

[rpjkw11]

My wife and I lived in Rio Rancho, NM for 15+ years. There's a big Intel plant there. We both were AMD fans, though. Eight years ago when I was planning a new PC, she asked if I was buying AMD and I told her, sadly, I wasn't because Intel sold what I wanted/needed. NOW, I have a real, legitimate choice! Intel's proposed $2k i9 7820X is just too ridiculous for my consideration, so when the Threadrippers hit the shelves, I'm buying a 16/32 1950X for HALF of Intel's extortion-like price and because I can then enjoy COMPETITION at long last.

 

[EarthDog]

The 7820X is $600... 7900X is $1K...

I assume you can use/need all those cores?

 

[Woomack]

Finally passed 100GB/s in AIDA64



[email protected] 1.27V / 64GB DDR4-3600 13-13-13 / ASUS TUF Mark 2

 

[Woomack]

I have to check my water block mounting or maybe change TIM as something isn't right with my temps ... or just different tests heat up my 7900X more.
When I used AIDA64 on CPU+FPU only then CPU temp was ~87°C but at 1.15V and after 30 mins full load. In the review in AIDA64 is 90°C but at 1.3V ( VID is 1.25V so maybe it works closer to 1.25V, software is not showing this right as I see ) while also memory and cache tests are enabled and it usually lowers temps.
Maximum wattage at the wall at 1.2V was about 440W so CPU had about 350W during AIDA64 CPU+FPU test.

OC of my 7900X looks like this:
4.5GHz 1.15V - seems stable
4.7GHz 1.20V - seems stable
4.8GHz 1.20V - benchable
4.9GHz 1.25V - benchable, overheats during full load on all cores
5.0GHz 1.32V - benchable in less demanding tests ( boots at 1.29V without issues )

 

[mackerel]

Finally passed 100GB/s in AIDA64

Missed this earlier... nice ram read bandwidth... but it is pretty much equal to L3! I've heard L3 was low and is something to look at also. This is still with your OC cache right? My 6700k @ stock is benching L3 read of 269 at 4000 NB clock. If you normalise for that it seems to agree with a comment elsewhere that L3 seems to be half the speed of Skylake-S. I wonder if this is related to the change from inclusive to victim cache.

On a side note, I guess I should take up their offer to upgrade my aida64 to current so it will support Skylake-X.

 

[Woomack]

Cache OC gave about 60% L3 bandwidth improvement. For some reason it's lower than it should be on this screenshot but wouldn't be much higher anyway. 150GB/s tops like write and copy.
Intel said something about their changes in cache and reordering its usage. I still can't find article where they said that but clearly they focused on L1 and L2 more.

 

[mackerel]

For highly optimised software, I wonder if they need to adjust any optimisations for the new cache structure? L2 is certainly a lot fatter, although I think it also picked up some extra latency on the way. The current way is more similar to how AMD have been doing it. There are still differences but I don't recall those right now.

 

[wingman99]

Sky lake X cache.

The big upswing for these processors is the rearranged cache arrangement, with Intel moving from 256KB of L2 cache on the previous generation to 1MB of L2 cache on Skylake-X (and changing the L3 cache from being a fully inclusive cache to being a non-inclusive cache). This significantly enhances software which is L2 cache size sensitive, although it remains to be seen how much of an effect it will have for consumers. http://www.anandtech.com/show/11542...by-lakex-time-line-preorders-and-availability

 

[Woomack]

This is what I had on my mind but I think it was in other article. Still the same info so thanks for posting that.