I had the opportunity to compare, all else being equal, my 5960x and a 6950x this weekend on my RVE.
EDIT1: Updated some numbers after a little more tuning.
EDIT2: Update with new info on DDR4 (128G working again)
EDIT3: Update with new info on DDR4 2T Overclocking head-room (2T vs 1T opens up headroom I have not seen on Haswell)
EDIT4: Seeing longer term stability issues with 128G
There is more to do before I fully understand what this chip can do and I have limited time to experiment, but here is what I've seen so far. I'm sure I am not alone in having seen a lot of things that are deeply concerning in the context of the price, but it helps to see concrete/direct comparisons to understand the value proposition such as it is or isn't with this chip.
The 5960x example to which the 6950x is being compared is not entirely a "fair" comparison, but rather a "best case" for the 59xx family. It sets a VERY high bar, but the huge premium charged is also a VERY high bar.
Cooling: full-board EK water block and large radiators - 420 + 280 that serve the CPU and 2 GTX980's (bitspower blocks) in a single loop with a D5 PWM pump.
BIOS: latest 3101
The good:
- RVE's default setup automatically had all cores turbo to 4GHz under load and was stable and ran cool (I saw high 30's and low 40's)
- (updated) Additional DDR4 overclocking head-room found (but fully stable numbers at the most extreme conditions not yet found)
- Idle temps are outstanding even at fixed voltages (19-22C for the cores and 25-29C for the package (which includes VRM and PCH owing to the full-board water block)
- Spin up from lower frequencies is MUCH more responsive - no observable difference in benchmark time in linux between "balanced" and "high performance" modes (even "power save" in linux). You go from 1200->4500 instantly and then back again based on load.
- As with Skylake, higher voltages than Haswell, but less heat so far until you hit the wall (a steeper slope to the temperature/frequency curve)
- No physical compatibility issues - drop in replacement for 59xx even with full-board waterblock (not too surprising, I hope?)
The bad:
- Cache OC appears to be relegated to lower speeds seen on other MBs (the 3.8GHz "wall")
- Had some initial problems with 128G kit and memory OC, but see above, these have been resolved to the benefit of the 6950x
- (NEW) longer term stability issues with 128G showing up - crashes after longer runtimes.
- Price.
Sequence of DDR4 issues:
1. 128G would not boot at all with 12-12-12-32-1T (1T is the key) which worked on 5960x
2. 64G would boot/run but eventually encountered corruption at 1T, but was stable at 2T
3. 128G now appears to be stable at 2800-12-12-12-32-2T SA @ 1.3v (auto voltage for SA)
4. 128G runs at 3200-14-14-14-34-2T (this is a 3000-14-14-14-34-2T kit)
The comparison (all values are "commanded" in BIOS settings, actually tend to be a little higher):
5960x:
- CPU 4.7GHz @ adaptive 1.25v
- Cache 4.2GHz @ offset +0.200
- BCLK 100MHz
- SA +0.220
- RAM - 128G 2800 12-12-12-32-1T (3000/14 kit, but faster at the application level at 2800/12)
- full chip load peaks at ~65C typically
6950x:
- CPU 4.5GHz @ 1.35v adaptive (best tstable so far without trying voltages beyond 1.37
- Cache 3.5GHz @ offset +0.250 (1.25v)
3.8GHz is the highest quasi-stable so far
Auto @ 4GHz produced 1.43v VCache in the BIOS, but even that eventually crashed
- BCLK 100MHz
- SA @ 1.3v (this is the "auto" setting)
- RAM:
--- 128G 2800-12-12-12-32-2T stable
--- 128G 3000-14-14-14-34-2T (sticker rating of memory kit) stable
--- 128G 3200-14-14-14-34-2T stable
---- 3333 BCLK 125 booted, but not stable - stopped at 1.37v
---- 3400 BCLK 100 booted, but not stable - stopped at 1.37v
Performance - rough comparison:
I limited to 8 cores with a custom, highly parallel and large memory application run (that scales linearly with time - that is, you can run it for 2 minutes or 200 minutes and the comparison is the same):
(using only 8 cores @ 2800CAS12)
6950x @ 4.5GHz/3.5GHz Cache - 1m 53s
6950x @ 4.5GHz/3.0GHz Cache - 2m
(using 9 cores - 10 broke my app - software issue @ 2800 CAS12)
6950x @ 4.5GHz/3.5GHz Cache - 1m 33s
(3200CAS14)
(8 cores)
6950x @ 4.5GHz/3.5GHz Cache - 1m 50s
(9 cores)
6950x @ 4.5GHz/3.5GHz Cache - 1m 30s
(using all 8 cores @2800CAS12)
5960x @ 4.7GHz - 1m 53s
5960x @ 4.4GHz - 2m
If you are wondering about how suspiciously similar those times are:
1. Keep in mind, there are 2 more cores waiting in favor of the 6950, IF your application can use them, you have potentially large speedup waiting.
2. This just confirms the reality that small changes to your setup can produce improvements as large as moving from one generation of chip to the next.
3. 1T vs 2T DRAM timings accounts for some of the lack of difference. The 5960x running at 1T has an advantage over the 6950x at 2T all else being equal.
4. As previously, I am guessing ASUS has some work to do in BIOS curves for the 6xxx family that you will see in cryptic messages like "improve system stability". This is where we stand today.
Temperatures:
(very overclocked)
6950x - peak so far observed under 10Core load is 67C, but average/typical package temp under load is 61C
5960x - peak 68-70C, average under load typically is 63-65C
(4GHz)
6950x - 41-45C
Idle (6950x Fixed voltage):
package = 29C
cores = 19-21C
Summary so far (again, this is a sample of one and precious little time devoted so far. So, I expect to learn more in the coming days):
Is it a big upgrade from 5960x? No. A big, unqualified snore of an "upgrade" from 5960x. The runaway majority of applications and use cases will see no noticeable improvement at all.
Obviously, the comparison here was against a very good 5960x, but you can scale your expectations accordingly and roughly linearly.
Gaming? No. Just No. Get a 5960x or a 5930k after their prices drop. Heck, get 3 5960x or 4 or 5 5930ks and use the extras for crafting projects.
Are there benefits to this generation of chip? Yes!
- Thermal production is improved and that means cooling should be easier, a 10 core is as easy or easier than an 8 core 5960, an 8 core is likely as easy as a 6 core is now and so on and so forth.
- Speedstep is improved - I'm seeing less penalty running in "efficiency" and "balanced" modes which means less heat in your room/house and less noise to cool it.
NOTE: I can't speak to whether windows does this as well yet, but if linux can do it, then the hard work is done, MSFT just has to not screw it up.
- AVX clocking separated also means some tweaking down the road as these chips mature along the lines of the Haswell line, hopefully that means fewer cooked chips for OC'ers. I didn't mention that as I haven't really played with it and none of my stuff uses AVX to any degree, but its nice to see it in the bios after hearing about all those prime95 chip bakings.
