Intel Avoton C2750 Benchmarks - Supermicro A1SAi-2750F

dba

Moderator
Feb 20, 2012
1,478
183
63
San Francisco Bay Area, California, USA
I think $600 is going to end up high.

I agree on that cluster idea but 1U is expensive to cool since you have small fans.

Here is the lspci -vvv output for one of the NICs:




dba - I ran pts/stream triad

  • Single L5520: 25,720
  • E3-1230 V1: 13,227
  • Avoton C2750 12,063
  • AMD Opteron 3380: 10,868
  • Atom S1260: 1,759
  • Raspberry Pi: 293

The E3 and Avoton numbers seem to be half of what I would expect. Makes me a bit bummed that I standardized on the pts benchmark.
You have both memory channels populated, and with the fastest supported RAM, and yet it's showing only 12GB/s? I'm with you: that's a bit disappointing. I guess I shouldn't be disappointed given the wattage, but I would have hoped/expected to see at least 20GB/s given that Xeon E5s v2s are pushing 100GB/s with four memory channels. 25GB/s would represent 1/4 the memory bandwidth, nicely in proportion with the roughly 1/4 the CPU grunt of a Xeon E5 v2.
 

Patrick

Administrator
Staff member
Dec 21, 2010
12,184
5,265
113
You have both memory channels populated, and with the fastest supported RAM, and yet it's showing only 12GB/s? I'm with you: that's a bit disappointing. I guess I shouldn't be disappointed given the wattage, but I would have hoped/expected to see at least 20GB/s given that Xeon E5s v2s are pushing 100GB/s with four memory channels. 25GB/s would represent 1/4 the memory bandwidth, nicely in proportion with the roughly 1/4 the CPU grunt of a Xeon E5 v2.
The strange thing is that the Haswell chips should have the same spec'd 25.1GB/s memory bandwidth. That is why I am thinking it is a benchmark issue.
 

jazzbearz

New Member
Sep 20, 2013
2
0
0
Can you test it with the redis get/set?

redis-benchmark -n 1000000 -t set,get -P 32 -q -c 200

Results from servers I have:

L5630 :
SET: 595864.38 requests per second
GET: 900322.00 requests per second

L5520 :
SET: 562113.56 requests per second
GET: 855432.00 requests per second

AMD E-350:
SET: 204611.48 requests per second
GET: 244582.66 requests per second

Atom D525:
SET: 122910.52 requests per second
GET: 159540.52 requests per second
 
Last edited:

Patrick

Administrator
Staff member
Dec 21, 2010
12,184
5,265
113
Can you test it with the redis get/set?

redis-benchmark -n 1000000 -t set,get -P 32 -q -c 200

Results from servers I have:

L5630 :
SET: 595864.38 requests per second
GET: 900322.00 requests per second

L5520 :
SET: 562113.56 requests per second
GET: 855432.00 requests per second

AMD E-350:
SET: 204611.48 requests per second
GET: 244582.66 requests per second

Atom D525:
SET: 122910.52 requests per second
GET: 159540.52 requests per second
What configuration file/ install method are you using?

On the first run this is what I'm seeing (no config file added):
SET: 280820.00 requests per second
GET: 355366.03 requests per second

Using the default config:
SET: 344471.22 requests per second
GET: 426621.16 requests per second
 

jazzbearz

New Member
Sep 20, 2013
2
0
0
What configuration file/ install method are you using?

On the first run this is what I'm seeing (no config file added):
SET: 280820.00 requests per second
GET: 355366.03 requests per second

Using the default config:
SET: 344471.22 requests per second
GET: 426621.16 requests per second
Thanks Patrick

Those runs were configured with default configurations from EPEL/CentOS.
 

Aluminum

Active Member
Sep 7, 2012
431
45
28
lspci: Ethernet controller: Intel Corporation Device 1f41 (rev 03)
igb driver, awesome, and part of the i350 family it seems so all the VM feature goodness.

The linux kernel for these should be mainstream pretty quick depending how much your distro loves to do updates.

Since its intel driver based the BSDs might get it this year in stable releases instead of next year (or never) for marvell, which means pfsense will have it around 2015 instead of 2020. (only slightly sarcastic)
 

OBasel

Active Member
Dec 28, 2010
494
62
28
igb driver, awesome, and part of the i350 family it seems so all the VM feature goodness.

The linux kernel for these should be mainstream pretty quick depending how much your distro loves to do updates.

Since its intel driver based the BSDs might get it this year in stable releases instead of next year (or never) for marvell, which means pfsense will have it around 2015 instead of 2020. (only slightly sarcastic)
Yea... total suckage. It's a big reason why pfsense isn't used more. BSD + own distro. Same reason FreeNAS is always behind. If you aren't Windows or Linux drivers can be a mess.
 

zer0sum

Active Member
Mar 8, 2013
520
216
43
igb driver, awesome, and part of the i350 family it seems so all the VM feature goodness.

The linux kernel for these should be mainstream pretty quick depending how much your distro loves to do updates.

Since its intel driver based the BSDs might get it this year in stable releases instead of next year (or never) for marvell, which means pfsense will have it around 2015 instead of 2020. (only slightly sarcastic)
Hmmm, I wonder if Intel SRIOV works with these?
 

Patrick

Administrator
Staff member
Dec 21, 2010
12,184
5,265
113
Just ran that redis server benchmark on an E3-1220 V3:

SET: 1392757.62 requests per second
GET: 1766784.50 requests per second
 

MiniKnight

Well-Known Member
Mar 30, 2012
3,014
922
113
NYC
Saw this pricing with a... get this... 26 week lead time on avnetexpress:

1-$478.1978
2-$466.7053
5-$458.7595
10-$454.1302
50+-$442.2000

Seems quite reasonable actually.
 

omniscence

New Member
Nov 30, 2012
27
0
0
One major tip is that although we see a Marvell quad Ethernet solution listed, the drivers used for the Supermicro A1SAi-2750F are Intel branded.
The Marvell part is only the (quad) PHY. The MAC is integrated. I think that makes sense, when a lot of these are integrated in a single unit they will be directly connected to a specialized switch IC.

Hmmm, I wonder if Intel SRIOV works with these?
Exactly my thought when I saw i350 compatible.
 
Last edited:

Marco

New Member
Sep 23, 2013
17
0
1
I've been waiting for so long this new Atom processor and a new board from Supermicro to upgrade my (Supermicro) home server, but now that A1SAi-2750 is out I'm a bit doubty about buying it (and anything with a c2750). An Atom c2550 would have been the way to go, for these reasons:

1) the board is too expensive, ~500 dollars is on pair with the expense for a low power xeon system (an X10S** motherboard in the range 150$-200$, the heatsink 30$, a xeon E3-1220Lv3 or E3-1230Lv3 is another 210$-260$); moreover the difference between the c2750 and a c2550 is about 85$ (171$ vs. 86$) which is a fairly relevant amount of money for a low cost system.

2) the performance of a c2750 chip looks considerably high, for sure great for microcloud, but useless for many appliance/nas/network/home systems where you have less chance to keep your cores busy. The main point is that the Atom's core is pretty simple and if you want to get out of it a lot of performance you need to go parallel and use as many cores as possible. Conversely, low power xeons, though plagued by the large L3 cache (= higher power consumption), the absence of high frequencies and many cores, come with tons of SIMD instructions (SSE, AVX, etc.) which are indeed more useful in many SOHO/NAS/home scenarios, in particular for packet processing and RAID calculations. For the same reason I'd like to know how the integrated Ethernet controller compares to the i210 (for example, how many rx/tx queues?).

3) with an E3-1220Lv3 equipped system you can go fanless and it's likely to be the same for a E3-1230Lv3 xeon, at least to some extent or when using power capping (c224 and c226 PCH only). In any case only moderate air flow is required, so it's easy to create a real low noise system. To achieve the same goal a c2550 or lower is to be choosen. Selling a noisy SOHO system makes no sense to me, I really don't understand the c2750 decision. BTW, I've seen a 35db fan, even though it won't run at full speed very often, I'm afraid it's still too loud for my likings.

Can you tell something more about the noise? Is it possible to disable some cores from the BIOS to lower the maximum power consumption and go fanless? It would also be very interesting to perform some tests that aim to expose the architectural differences between low power xeons (targeting complex tasks) and avoton atoms (targeting simple and higly parallel tasks), because I'm afraid that for many people xeon processors is be a better choice at roughly the same cost.
 
Last edited:

vegaman

Member
Sep 12, 2013
60
3
8
Auckland, New Zealand
I'd expect prices to drop $100 or so once they become more available. It's going to have lower power consumption than a low power Xeon as well (which I don't really see the point of tbh).
I'm mostly interested in pfSense throughput, VPN and Snort benchmarks on Rangeley. I don't think the hardware support will quite be there yet though. So I'll have to wait for that.
For now I'm going to compare prices of the SuperMicro barebone options to buying the case and motherboard separately. Maybe I'm better off waiting for the barebones and using a case I already have for my firewall for now.

Edit: Just had a look on Intel ARK. So the Avoton has Turbo Boost but Rangeley doesn't, interesting. Yeah the price difference between the 4 and 8 core chips is pretty big :-( . But it's only a 5 (Rangeley) or 6 (Avoton) watt difference in the Max TDP, so I don't think the power saving vs. performance cut really weighs up. Especially when it's likely to be idle a lot anyway.
 
Last edited:

dealcorn

New Member
Oct 12, 2011
24
0
1
RAID 6 Performance?

For home NAS/Archival Server use, I am curious how many cores Avoton requires to support what RAID 6 parity calculation speed? Can 8 cores properly feed the 2.5GbE (or 4 x 2.5GbE)? What RAID 6 parity calculation speed can 2 or 4 cores support? This was a prior generation issue.
 

Marco

New Member
Sep 23, 2013
17
0
1
I'd expect prices to drop $100 or so once they become more available.
I'm a long-time Supermicro customer for my own systems, where I work we have tons of Supermicro servers, as far as I can tell I've never seen any significant drop in prices over time, unless a new product replaces it (but it might not be the case as well). IMHO it's not reasonable to expect prices below 450$/430$ (the CPU alone is 170$, the board is at least another 150/170$ + the seller + others). The only chance to offer a low priced but highly valuable board is to use the Atom c2550 (that's what I was expecting).

It's going to have lower power consumption than a low power Xeon as well (which I don't really see the point of tbh).
Well, it depends, but more or less I'd say no:
- RAM: same power consumption, maybe slighly less for regular UDIMM (either 1.5V and 1.35V)
- CPU (TDP/typical/idle):
CPU + PCH + NICs 13+4+1.5 ~ 19W TDP 2/5+1.5+0.5 ~ 5/7W typical 1/3+1+0.2 ~ 3/4W idle
SoC ~ 20W 20W TDP guess 4/7W typical guess 2/4W idle
- remaining parts (BMC, etc): all the same (more or less, no USB 3.0 chip for X10 boards though)

It would be really interesting to see some real numbers for both chips but also regarding the E3-1230Lv3 and the Atoms c2550/c2558. Moreover E3 xeons seem to have better power management techniques.
In any case, the real problem is not saving 1W or 2W more, it's not a big difference, it's all about costs (again not very different actually), and performance (we lack a detailed comparison).

I'm mostly interested in pfSense throughput, VPN and Snort benchmarks on Rangeley.
It really depends on the level of traffic you have, but I'm sure that a c2558 would just do, for many many people.

But it's only a 5 (Rangeley) or 6 (Avoton) watt difference in the Max TDP, so I don't think the power saving vs. performance cut really weighs up. Especially when it's likely to be idle a lot anyway.
Yeah, "just" 5/6W, enough to require a fan in a 1U or compact chassis, that's the problem (and not a +0.000...$ on your bill) for those who want a silent system (but also OEMs that build NAS devices).
As for me, I don't get it, the Atom c2750 means a higher price point, different thermal design, overkill performance in many not so demanding scenarios. IMHO, of course.
 

MiniKnight

Well-Known Member
Mar 30, 2012
3,014
922
113
NYC
I would also like to see 8-core benchmarks. I'm more interested to see if pfsense SUPPORTS the Avoton processor, maybe Rangeley.
 

vegaman

Member
Sep 12, 2013
60
3
8
Auckland, New Zealand
I was referring to the prices on Amazon ($550 or $600) when I said I'd expect the price to drop. I'm ignoring the place with a 26 week lead time.
I see your point about wanting a quiet server, so for that and lower cost they probably should provide 4 core options as well.
I'd have to see actual measurements to be convinced about the power usage vs the low power Xeons though.