Intel i9 7900x CPUs

[Voxilla]

Regarding AVX 512, I really would like to get some more information about the crippled performance on the 6 and 8 cores.
I have been developing a state of the art volume rendering engine using AVX 512 using the Intel emulator and compiler.
I was all set on buying the 8-core. Now I'm very upset about the crippled AVX 512 on the 8 core.
The 8-core seemed a reasonalbe deal but the 10-core price is more than I (and most people) want to spend.

 

[Malo]

Regarding AVX 512, I really would like to get some more information about the crippled performance on the 6 and 8 cores.
I have been developing a state of the art volume rendering engine using AVX 512 using the Intel emulator and compiler.
I was all set on buying the 8-core. Now I'm very upset about the crippled AVX 512 on the 8 core.
The 8-core seemed a reasonalbe deal but the 10-core price is more than I (and most people) want to spend.

Does this page at Anandtech help?

 

[Voxilla]

Does this page at Anandtech help?

That's indeed where I first read about this:
"The six-core and eight-core Skylake-X parts support one fused FMA for AVX-512-F, although the 10-core will support dual 512-bit AVX-512-F ports, which seems to be located on port 5. This means that the 10-core i9-7900X can support 64 SP or 32 DP calculations per cycle, whereas the 8-core/6-core parts can support 32 SP or 16 DP per cycle."

Quite vague really. It seems AVX 512 on the 6 and 8 core will mostly perform no better as AVX2 on those processors, unless even worse also AVX2 get's crippled.
The former would mean AVX 512 being mostly pointless, the latter a complete disaster, on the 6/8 core
All reviews and benchmarks seem to be based on the 10 core, which isn't very helpfull with respect to the 6 and 8 cores.

 

[Malo]

All reviews and benchmarks seem to be based on the 10 core, which isn't very helpfull with respect to the 6 and 8 cores.

Unfortunately after the NDA leak Intel restricted the amount of CPUs that went out so we'll need to wait awhile longer for reviewers to get other models and do some more in-depth testing.

 

[Gubbi]

Anandtech has the full 7800X and 7820X and 7900X benchmarks.

To me it looks like Intel has a steady aim at both their own feet. Wide SIMD is all about throughput computing. IMO, it makes more sense to enable it on lower core count SKUS, given the different SKUs have more or less the same bandwidth. On high core count SKUs, the memory interface is going to be completely saturated except for a few cache friendly apps.

7800X: 336GFLOPS / 76.8GB/s = 4.38 FLOP/byte
7820X: 461GFLOPS / 85.1GB/s = 5.4 FLOP/byte
7900X: 1056GFLOPS / 85.1GB/s = 12.4 FLOP/byte

This is assuming AVX512 won't bring the operating frequency below base frequency, - which it will.

Cheers

 

[Clukos]

With a custom watercooling loop the 7900x is a beast:

Looks like it's thermal throttling quite a lot even with an expensive AIO. Which makes the decision to not solder these even more questionable by Intel, assuming a 20-25 C drop the 7900x would perform even better without the need for exotic cooling.

 

[CarstenS]

That's indeed where I first read about this:
"The six-core and eight-core Skylake-X parts support one fused FMA for AVX-512-F, although the 10-core will support dual 512-bit AVX-512-F ports, which seems to be located on port 5. This means that the 10-core i9-7900X can support 64 SP or 32 DP calculations per cycle, whereas the 8-core/6-core parts can support 32 SP or 16 DP per cycle."

Quite vague really. It seems AVX 512 on the 6 and 8 core will mostly perform no better as AVX2 on those processors, unless even worse also AVX2 get's crippled.
The former would mean AVX 512 being mostly pointless, the latter a complete disaster, on the 6/8 core
All reviews and benchmarks seem to be based on the 10 core, which isn't very helpfull with respect to the 6 and 8 cores.

I wonder why they did not test it, esp. if they had all necessary processors in their hands. From what I've seen so far, it's 32 DP-FLOPS per clock per core for i9-7900X (first-hand) and I got results from an outside source running the same test on a i7-7820K, indicating that it's got the same throughput per clock per core. But that was done using non-final software, so it might not be correct after all.

 

[Clukos]

Board problems

 

[BRiT]

"Are they even testing at all?" was one of the funniest things I've heard yet in a review video.

 

[Malo]

Board problems

Summary for the anti-video among us:

• 8-pin power for the CPU is not enough for 10-core Skylake-X, pulls over 300w
• VRMs on the "entry-level" $300+ boards overheat and cause CPU to clock down to 1.2Ghz
• Would recommend water-cooling for the VRMs
• Doesn't recommend any X299 boards at the moment and to wait a few months

 

[entity279]

He also said that removing the vrm heatsinks significantly decreased their temperature for some of the boards (as if adding useless heatsinks to memory chips was not enoungh)

 

[Malo]

He also said that removing the vrm heatsinks significantly decreased their temperature for some of the boards (as if adding useless heatsinks to memory chips was not enoungh)

To be accurate, removing the heatsinks and placing a 120mm fan over them.

 

[entity279]

Oh.. thanks. Videos are evil, it's their fault i've misremembered, not mine

 

[Davros]

Is there an alternative to x299 ?

 

[Blazkowicz]

For anything past Core 2 Duo and AM2/AM2+//AM3/AM3+ sockets there's no alternative to anything?

 

[Davros]

Your wrong there, for example socket 1151 has at least 4 alternatives

 

[Blazkowicz]

I meant chipsets too. Ever since Intel kicked nvidia out we've had chipset dictatorship (phones do the same thing, only everything is in a single chip or one main chip + one radio chip)

On traditional consumer desktop you do have e.g. Z170 and H110, H170, B150 - this suspiciously looks like the same silicon with disabled features. On high end consumer there's been only one without even half updates : X58, X79, X99, X299. Why make a $250 motherboard with crippled chipset?, might be the question.

Although this time there might be a mid-gen update to go with "Coffee Lake-X" and "Cannon Lake-X".
Also, CPU with 28 lanes vs 40 or 44 lanes is used instead of low end vs high end chipset.

 

[Clukos]

• 400 Watt through the VRM (just CPU), 500+ watt combined system load
• CPU in the 100-110 C range under 100% load
• VRM in the 100-105 C range under 100% load
• No delid on that 7800x

 

[Malo]

To expand upon that summary, the issue only occurs when:

• The BIOS settings are changed to allow a lot more power to go to the VRMs
• Certain safety features are turned off to allow temps to go higher
• A very specific heat generation test in Prime95 is used

It couldn't be reproduced in folding, OCCT, rendering or normal prime95 tests.

Seems like adjusting the settings to fit the narrative.

 

[Clukos]

Just a small correction that wasn't the 7900x pulling 500 watts, that was the 7800x. The 6 core Skylake-X.

Seems like adjusting the settings to fit the narrative.

Not really if the CPU is thermal/power throttling, all these settings do is allow the CPU to remain at a stable frequency with stable power delivery, and even then it throttles because of thermal issues.