4 x Intel Xeon good idea?

[Peperoni]

Hello,

What do you guys think about this configuration to run Stockfish games?
https://www.bargainhardware.co.uk/dell- ... ?c_b=21559

1 x Dell PowerEdge R820 8x 2.5" (SFF) - iDRAC Licence
4 x Intel Xeon E5-4657L V2 - 12-Core 2.40GHz (30MB Cache, 8.00GTs, 115W)
4 x Dell PowerEdge R820 Heatsink
12 x 16GB - DDR3L 1333MHz (PC3L-10600R, 2Rx4)

--> £811.20Inc. VAT

[Alayan]

The performance per core will be less than half that of a desktop Ryzen 5000 processor just from clockspeed and IPC, and it has much slower RAM and NUMA too. 12x16GB is overkill in total capacity but DD3-1333 is very slow.

48 cores makes it still very competitive at this price point from a pure throughput perspective if running parallel low-thread games, but if the goal is to frequently use all cores together on a single game it would probably lose to a modern 12 cores processor.

Another downside of the system is the huge power consumption. It makes the value proposition less attractive and can create annoyances.

If your use case involves running single-core games and the server dumping out a lot of heat (and noise) isn't an issue, then it could be a very attractive offer. Otherwise, I wouldn't recommend.

[AndrewGrant]

If you are trying to buy an existing server at a discount, Intel will probably have something good for you. If you want the best possible performance per dollar right now though, AMD wins the top 10 or so possible options. A good source for data is : http://ipmanchess.yolasite.com/amd---in ... -bench.php

I can say that I did some effort comparing NNUE's in Cfish between Intel machines and AMD machines for someone, and the result was that the AVX2 implementation in Intel and AMD were very similar -- IE the introduction of NNUE does not greatly change the rating list I posted.

[Raphexon]

Since it's a Xeon V2 it will not support AVX2 incase you want to run NNUE engines with it.

[Milos]

Since it's a Xeon V2 it will not support AVX2 incase you want to run NNUE engines with it.

That's correct. AVX2 came with Haswell. Xeon v2 is Ivy Bridge. Still difference between AVX (modern compile) and AVX2 shouldn't be more than 10-15% in nps.

[Milos]

The performance per core will be less than half that of a desktop Ryzen 5000 processor just from clockspeed and IPC, and it has much slower RAM and NUMA too. 12x16GB is overkill in total capacity but DD3-1333 is very slow.

48 cores makes it still very competitive at this price point from a pure throughput perspective if running parallel low-thread games, but if the goal is to frequently use all cores together on a single game it would probably lose to a modern 12 cores processor.

That is usually not true. IPC and clockspeed don't reflect same speedup from regular benchmarks (relying a lot on FP computations) to chess that runs mainly INT computations.
I still have 2xXeon v1 16 cores total config that runs at 17 Mnps SFNNUEdev from a starting position. That is comparable nps to 3800X, and my 2 CPUs cost around 80$ on Aliexpress atm, while 3800X is 350$.
With 48 Xeon v2 cores he'd probably get around (or even slightly over) 40 Mnps that is similar if not better than 3950X (and way faster than 3900X or even 5900X for example).

[Alayan]

That is usually not true. IPC and clockspeed don't reflect same speedup from regular benchmarks (relying a lot on FP computations) to chess that runs mainly INT computations.
I still have 2xXeon v1 16 cores total config that runs at 17 Mnps SFNNUEdev from a starting position. That is comparable nps to 3800X, and my 2 CPUs cost around 80$ on Aliexpress atm, while 3800X is 350$.
With 48 Xeon v2 cores he'd probably get around (or even slightly over) 40 Mnps that is similar if not better than 3950X (and way faster than 3900X or even 5900X for example).

But that's true. Clockspeed has a 100% correlation with performance for workloads that don't pressure caches or RAM. Chess engines do have a significant reliance on caches and RAM, but caches and memory performance are also significantly worse with those old Xeons, and higher clockspeed do translate in higher Stockfish NPS.

And IPC increases aren't just about FP, it's also about INT. The 5800X beats the 1700 by more than 50% on openbenchmarking and ipmanchess Stockfish benches, much more than the mere clockspeed difference. This also applies to Intel chips, and the lack of AVX2 gives a big handicap when running engines that do rely on it. We have direct benchmarking data on chess engines, there is no need to speculate.

And my point was:

48 cores makes it still very competitive at this price point from a pure throughput perspective if running parallel low-thread games, but if the goal is to frequently use all cores together on a single game it would probably lose to a modern 12 cores processor.

If running parallel single-core games, the total nps is a good indication of the performance, but multi-threading is lossy.

At equal overall nps, more threads plays weaker.

The scaling efficiency from 12 (24) to 48 (96) threads is really poor, so matching or moderately beating the modern 12 cores in total nps means losing in chess strength.

[AndrewGrant]

If running parallel single-core games, the total nps is a good indication of the performance, but multi-threading is lossy.

At equal overall nps, more threads plays weaker.

The scaling efficiency from 12 (24) to 48 (96) threads is really poor, so matching or moderately beating the modern 12 cores in total nps means losing in chess strength.

Indeed. I made this same argument recently. If you care about having the strongest machine for Engine vs Engine play using all threads at once per engine, then taking even a hefty NPS loss will still win.

Code: Select all

275.142.992	2x AMD EPYC 7742	256threads	
227.448.746	AMD Ryzen Threadripper 3 3990X @4.3Ghz	128threads	

I would bet my lunch and dinner that Stockfish on a 3990x, playing against Stockfish on 2xEPYC 7742, will win.

[jdart]

Note also this is a server box. It will be very loud with the fans going.

I'd recommend instead you look at workstations, for example:
https://www.ebay.com/itm/HP-Z640-Workst ... SwhHlemLQe

[Milos]

If running parallel single-core games, the total nps is a good indication of the performance, but multi-threading is lossy.

At equal overall nps, more threads plays weaker.

The scaling efficiency from 12 (24) to 48 (96) threads is really poor, so matching or moderately beating the modern 12 cores in total nps means losing in chess strength.

I don't agree on that in this particular case even with SF NNUE where LazySMP is practically broken beyond repair (I wonder how you SF developers didn't yet revert to old SFs SMP that scales much better with NNUE).
48 threads (no HT) with 40Mnps (4xXeonV2) is probably stronger (or at least on par) than 24 threads (with HT) and 28Mnps (3950X).
With classical SF difference would be even more pronounced.