[numerobis]

I just noticed that the fastest systems in the Benchwell list (Xeon E5 2696 V3; 2x18x2.8GHz=100.8GHz) made a jump from ~107sec (BM2600) to 34sec (BM9800) by enabling NUMA, can this be true?!? (NUMA: http://en.wikipedia.org/wiki/Non-uniform_memory_access ) - assuming that the values and specs are correct, which is not assured these days when i look at all the wrong entries in the list. But this improvement is so huge, that the only explanation could be a quad socket system if these are real values and wrong system specs.

When i compare these values to the results of an i7 5960X they look even more phantastic... 5960X @stock (8x3.3GHz=26.4GHz) is listed with 364 seconds or BM766. So the dual CPU system with roughly 3 times more GHz is now more than 10 times as fast as the i7? How is this possible? My understanding of NUMA is that it only optimizes the RAM management for systems with more than one CPU, so this should not be necessary for single CPU systems and since this is the same CPU architecture the results should be comparable.


And btw. the Benchwell list is still completely messed up. This was almost 3 years ago http://www.maxwellrender.com/forum/view ... 5=#p362725
and now the new list is even more spammed with advertising from gpucomputer.pl (and strongboxtechnologie, treddi or RenderSolve) and many doubled or useless entries.

http://www.maxwellrender.com/benchwell

[eric nixon]

I dont think its that spammy, more entries are welcome, and more companies building these kind of boxes is welcome no?

Would like to see an 18core @3ghz+ score soon.... I guess dual cpu only due to cost, rather than 4 or 8!.. I cant see any boss buying a machine with more than 2 xeons at forseeable prices.

Re. Numa, I think my bios has that turned on, I'll check later.

[numerobis]

So something like this is welcome?

Code: Select all

6	Intel R Xeon R CPU E5-2696 v3 2.30GHz	1m40s	2779.03	72	62 GB	3.1.1.0	Win64	waldek	(info: gpucomputer.pl)
7	Intel R Xeon R CPU E5-2696 v3 2.30GHz	1m46s	2622.46	64	14 GB	3.0.1.0	Win64	waldek gpucomputer.pl	    (info: 2x Xeon e5-2696 v3 18 core 72 HT)
8	Intel R Xeon R CPU E5-2696 v3 2.30GHz	1m47s	2598.07	64	30 GB	3.0.1.0	Win64	gpucomputer.pl	 (info: waldek Windows server 2008 r2)

Code: Select all

9	Intel R Xeon R CPU E5-2697 v3 2.60GHz	1m52s	2482.6	56	63 GB	3.1.0.0	Win64	Matt87	(info: Matt87)
10	Intel R Xeon R CPU E5-2697 v3 2.60GHz	1m53s	2460.73	56	63 GB	3.1.0.0	Win64	(info: Matt87 Dual Xeon 2697 V3 3800 Mhz)
11	Intel R Xeon R CPU E5-2697 v3 2.60GHz	1m55s	2418.12	56	63 GB	3.1.0.0	Win64	(info: Matt87)	

These are obviously benchmarks of the same systems with slightly different BCLK or RAM speeds - BCLK 105MHz for the second system as the 3800MHz in the info shows, which is the single core clock btw., so this is also misleading. And you need to know this stuff to interpret the spartanic info. And what about the other 5 entries? They are pretty useless without the info what has been altered compared to the standard system.

And why would you need the score of a system with Hyper Threading turned of?!?

Code: Select all

15	Intel R Xeon R CPU E5-2696 v3 2.30GHz	2m15s	2061.2	36	14 GB	3.0.1.0	Win64	gpucomputer.pl	(info: 2x18 core e5-2696v3 HT OFF)

What's next? One core disabled? Two cores disabled? Who would disable HT for rendering? This makes no sense...
And there are many wrong, incomplete and therefore misleading entries in this list. As i said in the old thread i think maybe around 30%.

But this shouldn't be the topic of this thread. I would like to know if it is possible, that these entries on top of the list are real and if they are real, what is causing this massive speed gain. I mean, if this is true i would buy one dual socket system instead of maybe 4 8-core systems (combined 105,6GHz, BM ~3060). The dual socket sys with almost the same combined GHz would still be two times faster (BM9800)... Really?!?

[eric nixon]

Well if its true, that would explain the huge cost, so its probably true. Also, inventing a false render time would be a huge lie.

[numerobis]

No one?
NL?
Is there anyone with a similar system who has experienced the same or can run the benchmark?

I mean... i would have considered these fabulous results to be something that some more users are interested in. From BM2600 to BM9800 just by enabling NUMA? More than 200% plus?
This is much more than the increase of IPC of the processor architectures over the the last 10 years.

Edit:
I just noticed, that No. 4 in the list now is a single quadcore Xeon E3-1270 v3 that is listed with 1m02s and BM 5138.56 - so there is definitely something wrong.
And the same E5-2696 v3 is also listed as single CPU with 5% overclock at No. 59 with 3m37s, BM1284 ...

[T0M0]

Ok, so I tested NUMA off/on in Benchwell on following system
- 2x Intel Xeon E5 2680v2 (specification - Intel ARK) 40 logical cores totally
- ASUS Z9PE-D8 WS
- 32GB RAM Kingston HyperX 1600 MHz (KHX1600C9D3/8GX)
- Windows 7 SP1 x64
- Maxwell 3.1.1.0

In both cases the results were exactly similar:
- time: 3m16s
- bench: 1421.34

What I just noticed with NUMA off, Benchwell detected 31GB. (with NUMA on, it was 30GB).

Since Maxwell 3.0.99.6 there is added support for machines with more than 64 cores in Windows,
and that system you mentioned has 72 cores. From what I read/heard Windows use different process management for processes which use more than 64 logical cores.
So maybe the NUMA setting really affects the final result in that case.
But at least not here (only 40 cores).

[eric nixon]

I wish I had 20/40 cores ,

Thnx for checking, btw.

[numerobis]

TOMO, thank you very much for the results!
This is what i expected and i can't see, how this could be much different (maybe a bit) for a few more cores, magically increasing the overall performance by more than 200% (!). But maybe someone with an E5-2696 can clarify this...

[numerobis]

on year later...

(i think i should buy an E3-1270 v3 Quad Core - must be something special )

[eric nixon]

The 1270 score is for a full rack of networked blades (called Blazer). The top 3 scores remain unexplained.

[numerobis]

Good to know - even if i don't know how it is possible to run benchwell on a network and if it is possible somehow then i can't understand why this entry should be in the list. I think it makes no sense to post network scores in a single PC list especially if the setup is not specified! And if it is really true this must be ~13 Xeons to get this score.

But if it is really possible to run benchwell on a network i think this is also the explanation for the 3 top scores

[feynman]

mo' corez, mo' munny, mo' fun

[Tok_Tok]

There is a Max Turbo version of the E5 2699 V4, like this one: http://www.ebay.nl/itm/Intel-Xeon-E5-26 ... 51e4c16519

It's not the same CPU but there probably a Max Turbo version of the E5 2696. That might explain the fastest benchmarks.

[numerobis]

The "max turbo" of 3.6GHz refers to the single core turbo. The all core turbo of the 2699 v4 is 2.8GHz.
https://en.wikipedia.org/wiki/List_of_I ... 2814_nm.29

But even if it was 22 x 3.6GHz this wouldn't be enough to get a benchmark of ~9800 instead of ~2800.
I think the entries are clearly fake. There must be either a way to run benchwell on a network or it is possible to manipulate the data before it gets uploaded. I don't know.