Rethinking optimal RAM Speed for Skylake builds...

All about them.

Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee

Post Reply
CA_Steve
Moderator
Posts: 7650
Joined: Thu Oct 06, 2005 4:36 am
Location: St. Louis, MO

Rethinking optimal RAM Speed for Skylake builds...

Post by CA_Steve » Sun May 29, 2016 9:43 am

When Skylake first came out, it appeared Intel had another generation that gained very little benefit in real world uses from DDR4 RAM speed increase of 2400 over 2133 except for some fringe cases (extreme overclocking, etc).
Larry compared 2133 vs 3000 on a stock i7-6700K, ADATA XPG SX910 128GB, GT 640 and found the performance difference to be minimal (except for iGP use).
Legit Reviews used an i7-6700K OC'ed to 4.6GHz, Corsair Neutron GTX 240GB SSD, GTX 980 Ti and found a tiny bump in Handbrake performance going up to 2400 or 2666 and then flat. Gaming with a 980 Ti was also flat (for the few games tested).
[H]ardocp used an i7-6700K at 4.5MHZ, Samsung 840 Pro 128GB, GTX Titan and found media apps like Handbrake to be flat across the RAM speed range. Video games run at low resolution with the Titan (so the gfx card didn't impact fps results) showed some scaling...

...and then Techspot used an i7-6700K at 4.5GHz, Samsung 950 Pro 512GB, and SLI GTX 980 Ti's and got completely different results. Handbrake and Adobe Photoshop CC, and games with significant scaling. Ok, it's not apples to apples versus the above due to stock vs OC CPU, low end GPU vs high end vs high end SLI, and also different SSDs. But there is a trend for the higher end systems. Compare the Witcher III SLI vs the single 980 Ti results. Very little benefit of faster RAM for the single card, but significant bump for SLI.

So, what are the differentiators? Is there some mix of overclocked CPU and high end GPU that taxes the memory and storage systems? Will this increase or decrease with the lower overhead and better CPU utilization of DX12 and Vulcan? Will this start to impact the single card gfx systems as Pascal and Polaris hit?

I'm leaning toward these recommendations as RAM prices have dropped:
- General purpose build, media center: DDR4 2133 is fine. Chances are DDR4 2400 cost the same.
- Gaming build 1080p: DDR4 2400-2666
- Gaming build 1440p+: DDR4 3000
- Heavy media creator, video work: home use DDR4 3000, workplace - faster (time is money :))

quest_for_silence
Posts: 5275
Joined: Wed Jun 13, 2007 10:12 am
Location: ITALY

Re: Rethinking optimal RAM Speed for Skylake builds...

Post by quest_for_silence » Sun May 29, 2016 10:28 am

CA_Steve wrote:...and then Techspot

Do you trust Techspot that much? What you say there about a swallow?

Though it's interesting, I think SPCR and [H] are two really serious sources to suddenly ditch them in favour of what Techspot says: I'm very dubious.
I think that would be interesting if your "colleague" Lawrence wanted to break his usual silence here on the forum (and maybe plan some suitable re-test).

CA_Steve
Moderator
Posts: 7650
Joined: Thu Oct 06, 2005 4:36 am
Location: St. Louis, MO

Re: Rethinking optimal RAM Speed for Skylake builds...

Post by CA_Steve » Sun May 29, 2016 11:50 am

I think there are edge conditions where the combination of CPU speed, memory speed and storage speed can each be the bottleneck. These are the conclusions I can draw from the above references:
- If you run stock CPU speed, then unlikely to see benefit of faster RAM. (SPCR)
- If you run OC i7 CPU, then unlikely to see benefit of faster RAM (Legit and [H]), unless:
--- you've got a fast SSD, and then you'll see improvements in intensive memory use cases like media creation apps and file archiving. (Techspot)
--- you've got a faster PCIe lane based SSD, and then you can see game fps improvements with high end SLI rigs. (Techspot)

After that it's speculation. For gaming, does the 980 Ti SLI setup at Techspot create it's own special case where higher overhead of SLI is helped out by faster memory? I wonder how the relatively flat single card sensitivity test Steven did would compare if he replaced the GTX 980 Ti with a GTX 1080 or the upcoming 1080 Ti / Titan?

For intensive media creation and archiving tasks, we know that NVME/PCIe based storage adds benefit. Is the higher SSD throughput then causing the memory system to bottleneck?

Abula
Posts: 3662
Joined: Tue Nov 02, 2004 12:22 pm
Location: Guatemala

Re: Rethinking optimal RAM Speed for Skylake builds...

Post by Abula » Sun May 29, 2016 12:05 pm

In the past there were a ton of benefits from faster ram since OCing was dependant on it.

Now you could get some, but overall to me its not worth it. I have done 4 builds on skylake, and found that even with QVList the boards are more picky with higher clock memory, in two builds i coudlnt even run them at their rated speeds. Overall i think if you do some editing then it might be worth it, but for gaming ill stick with stocks until i see different results.

I might consider them if i end up doing a Brodwell E (i doubt it), the only kit that caught my attention is the G.SKILL TridentZ Series 64GB (4 x 16GB) 288-Pin DDR4 SDRAM DDR4 3400 (PC4 27200) Intel Z170 Platform Desktop Memory Model F4-3400C16Q-64GTZ, still running at 1.2V =)

Rebellious
Posts: 322
Joined: Sun Dec 09, 2007 12:53 pm
Location: EU, USA

Re: Rethinking optimal RAM Speed for Skylake builds...

Post by Rebellious » Sat Jun 04, 2016 4:33 am

I built 3 Skylakes recently and found no real advantage in DDR4 speed. Here's one of them using DDR3 Samsung 'magic' ram:

http://www.passmark.com/baselines/V8/di ... 0297416810

I loaded PassMark test software and for comparison I searched for submitted results with the fastest memory scores. My sample outperformed 90% of them, and I assume that they are mostly overclocking geeks. Note I'm using a low-end mobo, DDR3-1866 9-9-9-28 which topped the list in latency score.

lb_felipe
Posts: 270
Joined: Wed Dec 24, 2008 12:21 pm
Location: São Paulo, SP, Brasil

Re: Rethinking optimal RAM Speed for Skylake builds...

Post by lb_felipe » Sun Jul 24, 2016 12:37 am

Abula wrote:In the past there were a ton of benefits from faster ram since OCing was dependant on it.

Now you could get some, but overall to me its not worth it. I have done 4 builds on skylake, and found that even with QVList the boards are more picky with higher clock memory, in two builds i coudlnt even run them at their rated speeds. Overall i think if you do some editing then it might be worth it, but for gaming ill stick with stocks until i see different results.

I might consider them if i end up doing a Brodwell E (i doubt it), the only kit that caught my attention is the G.SKILL TridentZ Series 64GB (4 x 16GB) 288-Pin DDR4 SDRAM DDR4 3400 (PC4 27200) Intel Z170 Platform Desktop Memory Model F4-3400C16Q-64GTZ, still running at 1.2V =)
No! That's a typo.

http://www.gskill.com/en/product/f4-3400c16q-64gtz

Post Reply