Intel core i7 6950x

[Joost Buijs]

In percentage how much improvement i7-6950X vs i5-2500K, both at stock speed.

I estimate about 2.25x or 225% ?

Difficult to tell because I don't have a 2500k available, my feeling is that it will be more like 2.5x or even more.
The 2500k has a 10% higher clock but the 6950X has a higher IPC that will fully compensate for the lower clock frequency.

[Joost Buijs]

The average difference in n/s after a few runs was approximately 24% in favor of the 6950X.

Thanks for the info. I use "time to depth" figures in my own benchmarking rather than nodes per second.

Time to depth is meaningful when you want to benchmark the SMP capabilities of an engine, to benchmark a processor I think nps is more appropriate.
In fact a chess engine is not a good tool to benchmark a processor at all.

[Modern Times]

In fact a chess engine is not a good tool to benchmark a processor at all.

Well, a good tool if chess is what you are going to be using it for

[MikeB]

The average difference in n/s after a few runs was approximately 24% in favor of the 6950X.

Thanks for the info. I use "time to depth" figures in my own benchmarking rather than nodes per second.

Time to depth is meaningful when you want to benchmark the SMP capabilities of an engine, to benchmark a processor I think nps is more appropriate.
In fact a chess engine is not a good tool to benchmark a processor at all.

And the only benchmarks that matters to chess engines are the integer benchmarks. In fact the Standard Performance Evaluation Corporation (SPEC) includes a chess engine , sjeng , in its current benchmark testing. A Chess engine is a good tool to benchmark CPU's - but just one tool of many.

The SPEC membership is a who's who in tech - includes all the big tech names:

Code: Select all

SPEC Members:

Acer Inc. * Action S.A. * Advanced Micro Devices * Amazon Web Services, Inc. * Apple Inc. * ARM * Avere Systems * Cavium Inc. * Ciara Technologies Inc. * Cisco Systems, Inc. * Dell, Inc. * Digital Ocean * E4 Computer Engineering SPA * EMC * Fujitsu * Gartner, Inc. * Google, Inc. * Hitachi Data Systems * Hitachi Ltd. * Hewlett Packard Enterprise * HP Inc. * Huawei Technologies Co. Ltd. * IBM * Inspur Corporation * Intel * Lenovo * M Computers s.r.o. * Microsoft * NEC - Japan * NetApp * NVIDIA * Oracle * Principled Technologies * Pure Storage * Qualcomm Technologies Inc. * Quanta Computer Inc. * Red Hat * Samsung * SAP AG * Seagate * SGI * Sugon * Super Micro Computer, Inc. * SUSE * Symantec Corporation * Taobao (China) Software Co. Ltd. * Unisys * Via Technologies * VMware *

The prior standard included crafty.

https://www.spec.org/benchmarks.html

[bob]

The reason Crafty is not in the new Spec version is somewhat funny. When the new version was first being planned, one of my contacts at Intel wanted to move crafty to the threaded/parallel part of SPEC. I explained to him very carefully exactly WHY this was a bad idea (chess is too non-deterministic). But he forged ahead saying that wasn't a problem.

Here's the glitch. When you submit a code to SPEC, you have to submit several sets of test data, and one set of production data. You have to also submit output for each test set and the production set. The test data is usually used for PGO compiling, but the final data can only be run after the compiler tweaks have been completed (IE you can't PGO using the final data, only the test data) The various manufacturers are free to use whatever compiler options/versions they have to maximize performance. But the output has to match the output provided by the program author. And anyone that has used a parallel chess program knows this fails. Node counts never match, and it is quite common for even the scores or principal variations to not be perfect matches.

So, a few months later, I got a call from Intel, again, telling me that they were going to remove Crafty because of its non-deterministic behavior. They had apparently already decided to use Sjeng for the non-threaded SPECINT, so Crafty was left out. Didn't bother me at all as it was a lot of work to help manufacturers debug their compilers (I spent a ton of time on the phone with Sun, when Crafty was in, because they had a few quirks in their brand new library. I didn't miss losing that.

[jdart]

I have a dual Xeon 2670 I use for testing.

While this may look attractive, you have to realize the per-core performance is not really stellar by current standards and it has a relatively low clock rate (2.6 GHz).

A 16-core dual 2670 system has a Passmark score of 18702 (http://www.cpubenchmark.net/cpu.php?cpu ... cpuCount=2). This is only about 13% above the 8-core i7-5960x (the brand new i7-6950x is not benchmarked yet). For chess you also have to factor in that many programs can't use 16 cores as effectively as a smaller core count, and also the Xeon is a NUMA system so there are inefficiencies from that.

--Jon

[Joost Buijs]

I have a dual Xeon 2670 I use for testing.

While this may look attractive, you have to realize the per-core performance is not really stellar by current standards and it has a relatively low clock rate (2.6 GHz).

A 16-core dual 2670 system has a Passmark score of 18702 (http://www.cpubenchmark.net/cpu.php?cpu ... cpuCount=2). This is only about 13% above the 8-core i7-5960x (the brand new i7-6950x is not benchmarked yet). For chess you also have to factor in that many programs can't use 16 cores as effectively as a smaller core count, and also the Xeon is a NUMA system so there are inefficiencies from that.

--Jon

This is why I will wait for some time before I decide to buy new hardware.
At the moment a used 2670 costs about 120 euro's, with two of these it is possible to build a dual Xeon for less than 1000 euro, but I have the feeling that it won't be much faster than the 5960X clocked at 4GHz. and that it will draw a lot more current.
For testing purposes a cluster with cheap quad-cores will probably do better.

The reason that I don't like to use a chess engine for benchmarking a multicore CPU is that most chess engines have deficiencies when used on more than 4 (or maybe 8) cores.
For vary well parallelized workloads the 6950X is about 30% faster than the 5960X (using all cores), with my engine I saw a difference of 24%.

[Laskos]

I have a dual Xeon 2670 I use for testing.

While this may look attractive, you have to realize the per-core performance is not really stellar by current standards and it has a relatively low clock rate (2.6 GHz).

A 16-core dual 2670 system has a Passmark score of 18702 (http://www.cpubenchmark.net/cpu.php?cpu ... cpuCount=2). This is only about 13% above the 8-core i7-5960x (the brand new i7-6950x is not benchmarked yet). For chess you also have to factor in that many programs can't use 16 cores as effectively as a smaller core count, and also the Xeon is a NUMA system so there are inefficiencies from that.

--Jon

In fact as absolute maximum chess strength goes, your system is no much better than cheapo 6 core i7-3930K OC to 4.5 GHz. But it is a very good testing platform.

[jdart]

I am probably going to build a system with the i7-6900k (8 cores). It is too steep a price to go to the 10 core chip. What I'd really want, ideally, is the e5-2687w v4 (12 cores @ 3.1Ghz), but the bare chip alone is $2500.

--Jon

[Werewolf]

I am probably going to build a system with the i7-6900k (8 cores). It is too steep a price to go to the 10 core chip. What I'd really want, ideally, is the e5-2687w v4 (12 cores @ 3.1Ghz), but the bare chip alone is $2500.

--Jon

The E5-2687W v4 is actually 3.2 GHz all cores. But even so, a better option IMO would be the E5-2690 v4 which is slightly cheaper and offers 14 cores @ 3.2 GHz.