Sparc T4 from Oracle

[jhu]

Faster than Sparc T3, adds out of order execution finally. Still paying off my T1...

 

[Soulkeeper]

interesting, I always enjoy reading the arch comparisons on arstechnica when these things are refreshed.

 

[lol123]

TPC-H benchmarks are very impressive according to this post at RWT: http://realworldtech.com/forums/index.cfm?action=detail&id=122688&threadid=122683&roomid=2

Maybe the SPARC is back?

 

[Idontcare]

Its maximum speed is said to be at least 3GHz, nearly double that of the 1.67GHz T3.

:hmm: I could see 3GHz on 28nm, but if they managed to get 3GHz silicon at sellable yields (>1 3GHz chip per wafer) on TSMC's bulk-40nm then that is some darn good engineering and design

 

[Soulkeeper]

"While the T4 is based on 40nm lithography, Oracle has also announced that a 28nm version, the T5, is a year ahead of schedule." --ars

Global foundries 28nm maybe ?

 

[blckgrffn]

:hmm: I could see 3GHz on 28nm, but if they managed to get 3GHz silicon at sellable yields (>1 3GHz chip per wafer) on TSMC's bulk-40nm then that is some darn good engineering and design

Well, it took them awhile, right? When did they (Sun) first start talking up this CPU? About as long as I have been in Enterprise IT if I remember correctly (~2006).

 

[Idontcare]

Well, it took them awhile, right? When did they (Sun) first start talking up this CPU? About as long as I have been in Enterprise IT if I remember correctly (~2006).

Don't blame SUN, their foundry at the time (Texas Instruments) abandoned the pursuit of advanced CMOS nodes beyond 65nm. As such, SUN was forced to change foundries for 45nm and beyond mid-stream their designs for such nodes.

This change in plans completely disrupted SUN's CPU design timeline as they had to adapt designs at the time to the changes in planned design rules between TI and TSMC.

That said, SUN is notorious in the industry, perhaps more so than any other IC design house outside of Bitboys, for having had some of the most over-budgeted and delayed processor designs on record.

They once had a chip internally codenamed the "Millennium" chip. (was to be SUN's UltraSparcV chip) It was called this because when the design was started in 1995 it was supposed to come out in 2000, i.e. at the turn of the millennium.

After purportedly spending $5B developing it, SUN cancelled the chip some 2 weeks after it finally tapeout...in 2004

By that measure, Ellison is already working miracles within SUN's product development timeline.

 

[BlueBlazer]

Faster than Sparc T3, adds out of order execution finally. Still paying off my T1...

It does seem that there's a new re-emphasis on single threaded performance. Looks like they can only push the processor this far with balance between a need for performance and multi-threading. :hmm:

 

[Abwx]

It does seem that there's a new re-emphasis on single threaded performance. Looks like they can only push the processor this far with balance between a need for performance and multi-threading. :hmm:

Of course , if your perfs/thread are about the level of a 486,
although with a high frequency , surely that single threaded
performance deserve perhaps a little more attention....

 

[Idontcare]

It does seem that there's a new re-emphasis on single threaded performance. Looks like they can only push the processor this far with balance between a need for performance and multi-threading. :hmm:

It's one of the things that made a critical difference between Intel's Atom and AMD's Bobcat.

 

[BlueBlazer]

TPC-H benchmarks are very impressive according to this post at RWT: http://realworldtech.com/forums/index.cfm?action=detail&id=122688&threadid=122683&roomid=2

Maybe the SPARC is back?

Interesting. This part.....

Oracle SPARC T4-4 Server, 4S/32C/256T, Oracle SPARC T4, 3GHz, 512GB, Solaris 10 8/11, Oracle 11g R2 EE with Partitioning - 201,487 QphH@1000GB at 4.60 USD per QphH@1000GB

By comparison to other "traditional" database submissions that's damn good score.

HP ProLiant DL980 G7, 8S/80C/160T, Intel Xeon E7-4870, 2.40 GHz, 2048GB, Ws2008 R2 EE, SQL Server 2008 R2 EE - 219,888 QphH@1000GB at 1.86 USD per QphH@1000GB
SPARC Enterprise M8000 Server, 16S/64C/128T, Fujitsu SPARC64 VII+, 3GHz, 512GB, Solaris 10 8/11, Oracle 11g R2 EE with Partitioning - 209,534 QphH@1000GB at 9.53 USD per QphH@1000GB
IBM System x3850 X5 8P, 8S/80C/80T, Intel Xeon E7-8870, 2.40 GHz, 2048GB, Ws2008 R2 EE, SQL Server 2008 R2 EE - 173,962 QphH@1000GB at 1.37 USD per QphH@1000GB
IBM Power 780 Model 9179-MHB, 8S/32C/128T, IBM Power7, 4.14GHz, 512GB, RHEL 6, Sybase IQ Single Application
Server Edition v.15.2 ESD #2 - 164,747.2 QphH@1000GB at $6.85 USD per QphH@1000GB
HP Integrity Superdome 2, 16S/64C/64T, Intel Itanium 9350, 1.73GHz, 512 GB, HP-UX 11i v3, Oracle 11g R2 EE with Partitioning - 140,181 QphH@1000GB at 12.15 USD per QphH@1000GB

Its not the highest but comparing with the highest (the Xeon) and second highest (the older SPARC64 VII+), 160 threads and 128 threads versus 256 threads. Looks like it still has some way to go before matching them (clock-to-clock and thread-to-thread). :hmm:

 

[Abwx]

Interesting. This part.....Its not the highest but comparing with the highest (the Xeon) . Looks like it still has some way to go before matching them (clock-to-clock and thread-to-thread). :hmm:

Did you notice that the xeon is a 8 Socket system , while the Sparc
is a 4S one ?.....

Do the maths , now...

 

[theeedude]

Don't blame SUN, their foundry at the time (Texas Instruments) abandoned the pursuit of advanced CMOS nodes beyond 65nm. As such, SUN was forced to change foundries for 45nm and beyond mid-stream their designs for such nodes.

This change in plans completely disrupted SUN's CPU design timeline as they had to adapt designs at the time to the changes in planned design rules between TI and TSMC.

That said, SUN is notorious in the industry, perhaps more so than any other IC design house outside of Bitboys, for having had some of the most over-budgeted and delayed processor designs on record.

They once had a chip internally codenamed the "Millennium" chip. (was to be SUN's UltraSparcV chip) It was called this because when the design was started in 1995 it was supposed to come out in 2000, i.e. at the turn of the millennium.

After purportedly spending $5B developing it, SUN cancelled the chip some 2 weeks after it finally tapeout...in 2004

By that measure, Ellison is already working miracles within SUN's product development timeline.

I don't know if it was $5B, but it was ten figures, and they did it AGAIN with ROCK. Talk about failing to learn a lesson from history and being bound to repeat it. But, the T processor team was a separate team within Sun, they used off the shelf CAD tools that were more inline with modern ASIC design than the ancient full custom CAD internal tools rest of Sun used (for various mainly political reasons that the weak Sun management failed to address, but the Oracle acquisition finally squashed). They were always a lot more efficient than the "Big Iron" Sun guys because of that, and when ROCK was cancelled, they finally emerged as the primary CPU design team within Sun, and it is starting to show.

 

[Idontcare]

I don't know if it was $5B, but it was ten figures, and they did it AGAIN with ROCK. Talk about failing to learn a lesson from history and being bound to repeat it. But, the T processor team was a separate team within Sun, they used off the shelf CAD tools that were more inline with modern ASIC design than the ancient full custom CAD internal tools rest of Sun used (for various mainly political reasons that the weak Sun management failed to address, but the Oracle acquisition finally squashed). They were always a lot more efficient than the "Big Iron" Sun guys because of that, and when ROCK was cancelled, they finally emerged as the primary CPU design team within Sun, and it is starting to show.

Yeah its been so long now that I can't honestly remember if the $5B was an internal number or if it was an external number.

I do remember that internally it was said 5,000 employees were working on the project at some point. Obviously not 5,000 for the 9yrs the project existed but that's a lot of compensation package when you integrate the area under the curve. (5,000 engineers circa 2000 would put the burn-rate at around $0.5B/yr)

Regardless the reality of any of these numbers when juxtaposed against my addled memory, there's a reason SUN is regarded as it is in the industry, and why Ellison was able to buy them for so cheap.

 

[jhu]

Did you notice that the xeon is a 8 Socket system , while the Sparc
is a 4S one ?.....

Do the maths , now...

Yes, that is pretty impressive given the number of physical CPUs in there.

 

[GammaLaser]

Max power 240W

Whoa D: