A SPARC CPU Overview: SPARC64-VII+ (M3), T3, T4, T5, M5, M10 - how do they compare?


Do you know how many SPARC CPU types does Oracle deliver currently? Not one, not two, but five. The SPARC64-VII+, the T4, T5, M5 and M10. It isn't easy to keep track of these, hence we created an overview table to help you in that - here you are: 

CPU Type

 SPARC64-VII+ (M3)  T3 T4 T5 M5  M10/SPARC64-X

Servers

M3000, M4000,
M5000, M8000,
M9000,
M9000-64

T3-1, T3-2
T3-4 

T4-1B, T4-1, T4-2
T4-4
T5-1B, T5-2, T5-4
T5-8
M5-32 M10-1, M10-4, 
M10-4S

CPU Core Codename

Jupiter++ S2 S3 S3 S3 Athena
CPU Clockfrequency 2.66-3 GHz 1.65 GHz 2.85-3GHz 3.6 GHZ 3.6 GHZ  2.8-3 GHz
Number of Cores per socket 4 16

 8 

16 6 16
Number of Threads per core 2 8

 8 

8 8 2
Minimum amount of RAM,
number of sockets,
and threads per Server

 8GB
1 socket
8 threads

8GB
1 socket
128 threads

 16GB
1 socket
64 threads

128GB
1 socket
128 threads
1TB
8 sockets
384 threads
32GB
1 socket
4 threads (2 cores activated from 16)
Maximum amount of RAM, number of sockets and threads per Server

4TB
64 sockets
512 threads 

128 GB
4 sockets
512 threads

2TB
4 sockets
256 threads 

4TB
8 sockets
1024 threads
32 TB
32 sockets
1536 threads
32TB
64 sockets
2048 threads
L3 Cache no L3$
(12MB L2$ though!)
no L3$

4MB

8MB 48MB no L3$
(24MB shared L2$ though!)

Virtualization supported

Dynamic Domains
(and Zones)
LDoms
(and Zones)

LDoms
(and Zones) 

LDoms
(and Zones) 
Hard Domains, 
LDoms
(and Zones) 
Physical Partitions (on 4S), LDoms
(and Zones) 

CPU architecture (ISA)

sun4u sun4v sun4v sun4v sun4v sun4v

 OS supported

Solaris 10, 11 Solaris 10, 11 Solaris 10, 11 Solaris 10, 11 Control Domain: S11 only.
Guest Domains: Solaris 10, 11
Solaris 10, 11

noteworthy features

electronically separated domains 512 threads in 2010! First S3 based, single thread AND throughput 1024 threads in 8 RU
Mainframe class,
Mission critical
Software on Chip


Allow me to add some thought-triggering facts too:

  • All the 6 of these CPUs have been released after Oracle taking over Sun.
    • (Oracle also has released Solaris 10 Update 10, Solaris 10 u11, Solaris 11 and Solaris 11.1 since then!) 
  • CPU cache is important: reading from RAM takes orders of magnitudes longer than from on-chip cache. 
  • It takes 5 years from CPU design start to deliver. Here's a quick thought-experiment: 
    • IF the T5 was already deep in the design phase before the acquisition in 2009, AND Oracle raised the investment into SPARC development THEN what can you assume about the next 1-3-5 years in terms of SPARC products? 

should you have questions, as usual, do not hesitate to ask in the comments! 

Comments:

Future SPARC processor will be optimized for Oracle DB, won't be ?
I expect future Exadata will run with SPARC Processor with new feature !

I suppose clock up to above 4GHz is very tough business technically
and from ECO point of view, you are going to change the SPARC processor
from general purpose processor to specific purpose processor.

Anyway, latest T5 processor is running at higher clock than the load map which had been available in 2009
so I see Oracle invests so much for SPARC processor development as Larry committed.

BTW, M10-4S has P-Partition feature which is almost same thing as Hard Domain in M5. Pls check it out.

Posted by guest on June 10, 2013 at 03:56 AM CEST #

Thanks for the M10 domaining feature hint, will check it out and update the post.
I am looking into these docs: http://jp.fujitsu.com/platform/server/sparc/manual/en/c120-e680-01en/index.html if you can provide any better source for information, do not hesitate to point me to it.

Posted by Karoly Vegh on June 10, 2013 at 02:17 PM CEST #

Regarding the install manual you refer, first Physical Partition is configured
using Modular style of M10-4S building Blocks, then Solaris is installed in PPar.
Finally configure guest domain with OVM 3.0.

I find Physical Partition is also described in white paper.

Manual Code: C120-E690-01EN

5.1.2 Partitioning Feature

My point of view M5's partitioning and M10-4S's Physical Partition is almost same.
Only difference is whether the partition is created in big enclosure or multiple enclosure.

However there may be some difference about redundancy, maintenance ability and anything else.
Of course CPU performance should be the big difference because of the clock and cache size !

Cheers ! Tomo-

Posted by guest on June 11, 2013 at 01:51 AM CEST #

I expect to see SPARC slide further into irrelevance with 3rd party app builders to the point of non certification/validation/support on SPARC/Solaris even for java or web app containers due to oracle not actively maintaining existing developer and 3rd party relationships.

Posted by Ian McGinley on June 13, 2013 at 01:06 AM CEST #

We can safely assume the prices of this hardware will be even more expensive then they are now, further alienating what remaining prospects at the time of acquisition Sun had in gaining fresh system administrators, since people cannot afford this hardware to play with at home, and testing it at work is no longer allowed or possible.

We can also assume what few of us remain on Solaris will be relegated to watching Oracle price itself out of the general market, while redhat Linux on cheap intel servers supplants it, and killing Solaris in the process as collateral damage.

(It is a shame for good sparc hardware, but the biggest sin will be the damage these pricing policies are causing, and will cause to adoption of Solaris. Unforgivable - inexcusable.)

Posted by guest on June 13, 2013 at 10:14 PM CEST #

Yes I am concerned about SPARC turning into "database processor" only.
As stated on T5 launch, "The ultimate optimization is hardware", I believe my worry is justified.

If database can still be orders of magnitude accelerated by incorporating more and more logic into the processor, what's the point of a generalized processor?

Maybe Oracle could consider a kind of "database co-processor" and leave SPARC as a general processor.

In the same vein, I'd also like see "co-processors" for that kind of power GPUs provide.

Regards.

Posted by Carlos Azevedo on June 14, 2013 at 05:31 PM CEST #

You need to add single socket blade servers like T4-1B and T5-1B to complete table.

Posted by guest on June 15, 2013 at 10:33 PM CEST #

Hi,

Major usage of SPARC Server is DB, so that I assume Software On Chip(SOC) is expected feature. CPU clock is very tough to increase recently
so SOC is very reasonable method to get more performance.

Of course SOC is one of the way to make CPU faster,
I believe Oracle will implement another features to make it faster,
both database usage and others.

What I worry of is if Oracle makes SPARC Server to be just a part of
appliance products like ODA/Commet and won't release general purpose
SPARC servers in the market.
When I hear about ODA, I felt it is very interesting product from HW
point of view and I preferred running x86 Solaris on it !

Tomonori-

Posted by Tomonori on June 17, 2013 at 01:53 AM CEST #

Update:

Physical Partitions added to the M10-4S, T4-1B and T5-1B added to the
respective series. Thank you for the hints.

Posted by Karoly Vegh on June 17, 2013 at 10:20 AM CEST #

As for 3rd party relations being maintained, let me recommend another blogsite, the partnertech blog, where a fragment of supported Software on Solaris 11 is pointed out: https://blogs.oracle.com/partnertech/

Posted by Karoly Vegh on June 25, 2013 at 02:49 PM CEST #

For years SPARC is a platform almost exclusively for ORACLE RDBMS, at least in Spain SPARC server sales have been almost exclusively for running ORACLE RDBMS.

May be the future, will be ORACLE RDBMS on a chip.

Posted by Felix Serrano on July 02, 2013 at 04:38 PM CEST #

I want to know what different about result of "floating-point arithmetic" by CPU type.
If possible, please help me.

Posted by guest on July 31, 2013 at 07:56 AM CEST #

Do we have best practive guidelines and recommendation for Deploymnet of Solaris 10 and 11 on T5 and M5 server..
Like What will be the size of filesystem ( rootpool and var dataset)
Swap size and other general optimization guidelines..
PR

Posted by guest on September 05, 2013 at 12:42 PM CEST #

The rpool usually consists of two complete harddisks in mirror. How far the capacity of those gets utilized depends on a number of factors, like if you put your zones into the rpool, how many times do you update and how mand BootEnvironments do you plan to keep, what software set do you plan to install, etc. A default Solaris 11 installation doesn't consume much space itself, probably around 500MB+dump+swap. A Solaris 10 installation consumes around 4GB, since - according to the EIS checklist - the best is to install the Entire packageset+OEM.

For Solaris 11.1 I recommend this section of the documentation:

http://docs.oracle.com/cd/E26502_01/html/E29007/ggrln.html

especially the sections:

[...]
For information about the swap and dump volume sizes that are created by the installation programs, see ZFS Root Pool Requirements.

Both the swap volume size and the dump volume size can be adjusted after installation. For more information, see Adjusting the Sizes of Your ZFS Swap and Dump Devices.

Consider the following issues when working with ZFS swap and dump devices: ...
[...]

Posted by Karoly Vegh on September 05, 2013 at 01:54 PM CEST #

Thanks Karoly for quick response.

Do we have EIS check list link too?

Regards
PR

Posted by guest on September 05, 2013 at 02:14 PM CEST #

I'm afraid the EIS checklist isn't publicly available, but your HW partner should have access to it.

Posted by Karoly Vegh on September 05, 2013 at 03:57 PM CEST #

Post a Comment:
  • HTML Syntax: NOT allowed
About

This is the Technology Blog of the Oracle Hardware Presales Consultant Team in Vienna, Austria. We post about our technology fields: server- and storage hardware, operating system technologies, virtualization, clustering, datacenter management and performance tuning possibilities.

Search

Categories
Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today