Many of us weren't expecting a 10+% performance bump in sky lake gaming, but we expected the tried-&-true 0-5% annual performance increase. It's good to hear the architecture might be able to provide that.
I'm a little disappointed with Skylake. Clock for clock performance seems worse than my 4790K. Even memory performance is worse, my DDR3 @2133Mhz if faster (in MB/s) than DDR4 @2133Mhz, and DDR4 has much worse latency. I didnt expect a huge improvement with Skylake but I dont want to go backwards in performance.
You're comparing an overclocked DRAM setting to a stock DRAM setting. The CAS Latency also governs the rate of data transfer, so you are probably comparing a CL of 10 (DDR3-2133 C10) to a CL of 15 (DDR4-2133 C15), so no wonder it comes out as slower. The more apt comparison would be DDR3-1600 C11 to DDR4-2133 C15, or some overclocked DDR4.
Check out our main 6700K/6600K Skylake review for more info, with IPC tests and DRAM comparison metrics: www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation
There are plans for a Skylake DRAM scaling article when a few other items are off my desk.
I'd love to read that. We don't often get new memory technologies, and it would be great if Skylake benefited meaningfully from the higher transfer speeds DDR4 will consistently reach once it matures.
I wouldn't expect anything dramatic, the P4 was the last Intel CPU that was RAM bottlenecked in normal use.
If anything saw more than a 1 or 2% difference it'd probably be the IGP, because graphics primarily uses streaming data access - making the throughput gains from DDR4 more important than the looser timings needed to support the faster data rate - and because the IGP is, unlike the rest of the CPU, badly memory starved on higher end models.
Sounds about right. Ever since they moved the memory controller onto the CPU die (Core series), Intel did a very good job hiding memory latency behind their three levels of cache. Save for some specific workloads, it doesn't really matter what memory you're running. Which is why I tend to hunt for power voltage instead of higher frequency when shopping for RAM.
"You're comparing an overclocked DRAM setting to a stock DRAM setting."
So? Few people use JEDEC spec RAM with overclocked processors.
DDR4 seems to offer something that primarily appeals to the server market: lower voltage (but with high latency). For desktop users it's definitely worthwhile to ask what's the point.
Of course a DDR3 at the same frequency then a DDR4 will have better performance, this is due to memory timing, it was the same thing for DDR to DDR2 and DDR2 to DDR3.
Yeah, I don't think it exists in earlier architectures, because you would only need something like this as a little buffer to cross the frequency boundaries between the core/bclk and pcie clocks, which will not always be in phase or even in sync now that they are generated from different clock sources.
Great article. This indicates that Skylake was a little rushed. I've been sitting on a VIII gene since they were available last month and can't find a 6700k. The good news appears to be that the longer it takes to get to my doorstep the more time there will be for intel to tie up these loose ends.
Course there is a dark side of that coin: the 30 day return window just came and went for everything i bought in preparation for Skylake, so if the parts don't work it's going to be a fun-filled warranty return circus around here.
"But at this time, the 10x ratio setting in Intel’s microcode (MRC) was not functioning as expected when motherboard manufacturers tried to initialise it during start-up. As a result, the ‘default’ value was used universally."
I'm assuming that we'll need a newer microcode before attempting to bump the FCLK up to 1GHz?
If your BIOS has an option to adjust the FCLK, then you have the new microcode. Otherwise you shouldn't be seeing it. As far as I can tell, no motherboard manufacturer made it a user-adjustable option until 1168 was made available. If by some chance the option to adjust is there and the MRC version isn't up to date, then it might show errors akin to an unstable overclocked system.
It may be that the processors already had the newer microcode when they initially started shipping, but that motherboard manufacturers hadn't had a chance to finish validating their motherboards w/ a 1Ghz FCLK.
Still remains the unexplored problem of poor PCI bandwidth.
Using non SATA SSD and many devices, risks detracting GPU performance due to the low PCI bandwidth to which skylake is capable, combined with the large number of dispositives which can use it on this processor.
Only if you're putting the SSD/etc on the CPU (Why? The southbridge has the same PCIe3 as the CPU.); or trying to run an x4 card on the southbridge forcing it fight with everything else for the 4xPCIe equivalent uplink to the CPU. If you're connecting things normally the GPU and SSD won't be sharing any bandwidth. IF you're trying to stuff a 3rd GPU onto the southbridge so your first two can have 8/8 bandwidth; then just as with previous generations you're pushing the mass market platform harder than its design was optimized for, and really should be using a higher end LGA 2011 system.
Nice work. I like to see Anandtech using it's influence to make products better!
On the other side: "Vut for discrete gaming it is pretty much par for the course. Arguably, the i7-6700K should be easier to get hold of over time compared to the i7-5775C as well. "
No one got the Vut? It jums to my face and I am not a native english speaker.
So does this mean the review batch of skylakes have their fclocks set at 800 but retail chips have it at 1000? If not and retail chips have it set at 800 as default, doesn't that mean you have to overclock the fclock just to reach parity of haswell?
No, it's a motherboard side issue due to Intel's MRC not working with 10x ratio at launch. The processors are still the same, and the updated MRC should be in the latest BIOSes soon (if not already).
Could this issue be a contributor to the processor shortages? As in, Intel waiting to make sure this is fixable via a firmware update rather than having to actually update the CPU? And since it's firmware fixable, no need to update 6700k and they'll open the floodgates sometime soon? They are NOT planning on updating the 6700k for this right - just motherboard firmwares will fix it?
I'm glad though that I got off the bleeding edge a few years ago, I've developed the patience to wait 6 months to move to new hardware much like the couple of weeks that I wait on MS updates. More time to get things done and less head pounding into desk. :)
EVGA tells me that their Z170 boards had FCLK set to 1000MHz (10 x BLCK, I assume) from the outset even though they haven't implemented IME 1168. Is this possible?
If you look at GT1 GT2 GT3 MFLOPS performance Haswell .. Broadwell .. Skylake, in wikipedia: https://en.wikipedia.org/wiki/Intel_HD_and_Iris_Gr... There is basically ZERO improvement in GFLOPS. Intel is playing its customers for fools.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
37 Comments
Back to Article
ImSpartacus - Tuesday, September 8, 2015 - link
Excellent follow up.Many of us weren't expecting a 10+% performance bump in sky lake gaming, but we expected the tried-&-true 0-5% annual performance increase. It's good to hear the architecture might be able to provide that.
MrSpadge - Tuesday, September 8, 2015 - link
+1This might also help GP-GPU apps which may be sensitive to the latency between CPU and GPU.
cactusdog - Tuesday, September 8, 2015 - link
I'm a little disappointed with Skylake. Clock for clock performance seems worse than my 4790K. Even memory performance is worse, my DDR3 @2133Mhz if faster (in MB/s) than DDR4 @2133Mhz, and DDR4 has much worse latency. I didnt expect a huge improvement with Skylake but I dont want to go backwards in performance.Ian Cutress - Tuesday, September 8, 2015 - link
You're comparing an overclocked DRAM setting to a stock DRAM setting. The CAS Latency also governs the rate of data transfer, so you are probably comparing a CL of 10 (DDR3-2133 C10) to a CL of 15 (DDR4-2133 C15), so no wonder it comes out as slower. The more apt comparison would be DDR3-1600 C11 to DDR4-2133 C15, or some overclocked DDR4.Check out our main 6700K/6600K Skylake review for more info, with IPC tests and DRAM comparison metrics:
www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation
There are plans for a Skylake DRAM scaling article when a few other items are off my desk.
Timbrelaine - Tuesday, September 8, 2015 - link
I'd love to read that. We don't often get new memory technologies, and it would be great if Skylake benefited meaningfully from the higher transfer speeds DDR4 will consistently reach once it matures.DanNeely - Tuesday, September 8, 2015 - link
I wouldn't expect anything dramatic, the P4 was the last Intel CPU that was RAM bottlenecked in normal use.If anything saw more than a 1 or 2% difference it'd probably be the IGP, because graphics primarily uses streaming data access - making the throughput gains from DDR4 more important than the looser timings needed to support the faster data rate - and because the IGP is, unlike the rest of the CPU, badly memory starved on higher end models.
bug77 - Wednesday, September 9, 2015 - link
Sounds about right. Ever since they moved the memory controller onto the CPU die (Core series), Intel did a very good job hiding memory latency behind their three levels of cache. Save for some specific workloads, it doesn't really matter what memory you're running. Which is why I tend to hunt for power voltage instead of higher frequency when shopping for RAM.Oxford Guy - Tuesday, September 8, 2015 - link
"You're comparing an overclocked DRAM setting to a stock DRAM setting."So? Few people use JEDEC spec RAM with overclocked processors.
DDR4 seems to offer something that primarily appeals to the server market: lower voltage (but with high latency). For desktop users it's definitely worthwhile to ask what's the point.
Klimax - Wednesday, September 9, 2015 - link
You might want to compute actual latency of various options in ns then to compare them just on raw CL number. You'll be surprised...Oxford Guy - Wednesday, September 9, 2015 - link
How? CAS 11 or 12 DDR3 2400 vs. CAS 15 or 16 DDR4 2400...Hurk - Sunday, September 13, 2015 - link
now we divide 12x3 and get 4CAS per instruction, and we divide 16x4 and get 4CAS per instruction. They are the same.iniudan - Tuesday, September 8, 2015 - link
Of course a DDR3 at the same frequency then a DDR4 will have better performance, this is due to memory timing, it was the same thing for DDR to DDR2 and DDR2 to DDR3.ikjadoon - Tuesday, September 8, 2015 - link
So, in Haswell, what is the default FCLK and the "recommended" FCLK?I think that's incredibly important to drawing any sort of conclusion...
Ian Cutress - Tuesday, September 8, 2015 - link
As far as we know, if such an option exists for Haswell, the option in Haswell is not user-adjustable.extide - Tuesday, September 8, 2015 - link
Yeah, I don't think it exists in earlier architectures, because you would only need something like this as a little buffer to cross the frequency boundaries between the core/bclk and pcie clocks, which will not always be in phase or even in sync now that they are generated from different clock sources.alacard - Tuesday, September 8, 2015 - link
Great article. This indicates that Skylake was a little rushed. I've been sitting on a VIII gene since they were available last month and can't find a 6700k. The good news appears to be that the longer it takes to get to my doorstep the more time there will be for intel to tie up these loose ends.Course there is a dark side of that coin: the 30 day return window just came and went for everything i bought in preparation for Skylake, so if the parts don't work it's going to be a fun-filled warranty return circus around here.
Gigaplex - Tuesday, September 8, 2015 - link
"But at this time, the 10x ratio setting in Intel’s microcode (MRC) was not functioning as expected when motherboard manufacturers tried to initialise it during start-up. As a result, the ‘default’ value was used universally."I'm assuming that we'll need a newer microcode before attempting to bump the FCLK up to 1GHz?
Ian Cutress - Tuesday, September 8, 2015 - link
If your BIOS has an option to adjust the FCLK, then you have the new microcode. Otherwise you shouldn't be seeing it. As far as I can tell, no motherboard manufacturer made it a user-adjustable option until 1168 was made available. If by some chance the option to adjust is there and the MRC version isn't up to date, then it might show errors akin to an unstable overclocked system.jagilbertvt - Tuesday, September 8, 2015 - link
It may be that the processors already had the newer microcode when they initially started shipping, but that motherboard manufacturers hadn't had a chance to finish validating their motherboards w/ a 1Ghz FCLK.extide - Tuesday, September 8, 2015 - link
That will come with the newer BIOS versions which allows you to set the option (or forces the new setting under the hood)Shiva_nf - Tuesday, September 8, 2015 - link
Excellent article and follow up. This is where reading Anandtech makes it worthwhile for me compared to other sites. Again great work Ianmarraco - Tuesday, September 8, 2015 - link
Still remains the unexplored problem of poor PCI bandwidth.Using non SATA SSD and many devices, risks detracting GPU performance due to the low PCI bandwidth to which skylake is capable, combined with the large number of dispositives which can use it on this processor.
DanNeely - Tuesday, September 8, 2015 - link
Only if you're putting the SSD/etc on the CPU (Why? The southbridge has the same PCIe3 as the CPU.); or trying to run an x4 card on the southbridge forcing it fight with everything else for the 4xPCIe equivalent uplink to the CPU. If you're connecting things normally the GPU and SSD won't be sharing any bandwidth. IF you're trying to stuff a 3rd GPU onto the southbridge so your first two can have 8/8 bandwidth; then just as with previous generations you're pushing the mass market platform harder than its design was optimized for, and really should be using a higher end LGA 2011 system.valinor89 - Tuesday, September 8, 2015 - link
Nice work. I like to see Anandtech using it's influence to make products better!On the other side: "Vut for discrete gaming it is pretty much par for the course. Arguably, the i7-6700K should be easier to get hold of over time compared to the i7-5775C as well. "
No one got the Vut? It jums to my face and I am not a native english speaker.
valinor89 - Tuesday, September 8, 2015 - link
Aaaand of course I had to make a typo... jums => jumpsByte - Tuesday, September 8, 2015 - link
So does this mean the review batch of skylakes have their fclocks set at 800 but retail chips have it at 1000? If not and retail chips have it set at 800 as default, doesn't that mean you have to overclock the fclock just to reach parity of haswell?Ian Cutress - Wednesday, September 9, 2015 - link
No, it's a motherboard side issue due to Intel's MRC not working with 10x ratio at launch. The processors are still the same, and the updated MRC should be in the latest BIOSes soon (if not already).LanceDiamond - Wednesday, September 9, 2015 - link
Could this issue be a contributor to the processor shortages? As in, Intel waiting to make sure this is fixable via a firmware update rather than having to actually update the CPU? And since it's firmware fixable, no need to update 6700k and they'll open the floodgates sometime soon? They are NOT planning on updating the 6700k for this right - just motherboard firmwares will fix it?Ian Cutress - Wednesday, September 9, 2015 - link
"just motherboard firmwares will fix it"Correct. Update your BIOS if you want the latest options, or wait a few weeks until functionality and compatibility is smoothed out.
RealBeast - Wednesday, September 9, 2015 - link
Thanks Ian, great work as usual.I'm glad though that I got off the bleeding edge a few years ago, I've developed the patience to wait 6 months to move to new hardware much like the couple of weeks that I wait on MS updates. More time to get things done and less head pounding into desk. :)
woofaki - Sunday, September 13, 2015 - link
Im about to upgrade my old rig...But this article confused me.I was about to get a skylake I5 one,should i still go for it or they are problematic?These are the parts i wanted to buy...
skylake I5 6600k
Asus Z170 Pro Gaming
Corsair Vengeance LPX 16GB DDR4-2400MHz
Ian Cutress - Sunday, September 13, 2015 - link
By the time you invest, it shouldn't be an issue.This was more addressing some of the results we saw with our launch-day review article that were not as expected.
svan1971 - Tuesday, September 15, 2015 - link
Skylake ? what's that ? Oh the cpu that can't be purchased 6 weeks after launch.JeremyReynolds - Wednesday, September 23, 2015 - link
EVGA tells me that their Z170 boards had FCLK set to 1000MHz (10 x BLCK, I assume) from the outset even though they haven't implemented IME 1168. Is this possible?JeremyReynolds - Friday, September 25, 2015 - link
I've since run HWinfo64 and this shows (under System Agent Status) that EVGA does have FCLK set to 1000MHz. Well done EVGA!systemBuilder - Friday, December 25, 2015 - link
If you look at GT1 GT2 GT3 MFLOPS performance Haswell .. Broadwell .. Skylake, in wikipedia:https://en.wikipedia.org/wiki/Intel_HD_and_Iris_Gr...
There is basically ZERO improvement in GFLOPS. Intel is playing its customers for fools.
systemBuilder - Friday, December 25, 2015 - link
vs. https://en.wikipedia.org/wiki/Intel_HD_and_Iris_Gr...