State of The Art - SPARC S7 & Solaris

[SPARC S7 Processor, Courtesy Oracle Data Sheet]

State of The Art - SPARC S7 & Solaris

Abstract:

The SPARC processor was developed by Sun Microsystems and had existed since the exit of major systems manufacturers from the Motorola 68K environment. Multiple manufacturers had always existed in this environment, to provide to consumers multiple supply chains in this commodity hardware market. The migration from 32 bit to 64 bit computing in SPARC occurred decades ago, as current computing systems still wrestle with the complexities. Oracle ceased producing horizontal scaling CPU sockets once purchasing Sun Microsystems. In June 2016, Oracle decided to re-enter the commodity market with the SPARC S7. NetMgt had published an article, but it was lost, and it was decided it was time to re-publish it again.

[Labeled Die Photo, courtesy Oracle Hot Chips 27 Presentation]

S7 "Sonoma" Floor Plan:

It can be clearly seen that the new die photo shows the use of 2x 4 Core Clusters. The cores are nearly the same 4th generation S4 cores, bundled in it's 32 core sister M7 processor. Glueless Coherency links were bundled, to scale an S7 system from 8 to 16 cores, with little external circuitry. With DDR4 memory interfaces on-chip, latency is cut down by eliminating external chips. Database Analytics Accelerators have been included, although not as many per-core as with the larger M7.

Close to 20% of the space used by an Infiniband network interface. The last time integrating network on-board silicon occurred was with the UltraSPARC T2+ (which integrated 10 Gig Ethernet) - but it was not released to customer facing production system.This was a significant disappointment to NetMgt, since an new Blade system with an Infiniband backplane allowing scalability to thousands of sockets would have been a welcome addition for cloud computing.

[Courtesy, Oracle's SPARC S7 Servers Technical Overview]

Architectural Changes:

The S7 is a smaller processor, returns to 8 cores on a die, similar to the T4 from back in 2011, but even outperforms processors from 2013 with higher cache, clock speed, and 4th generation core. It was designed to be a competitive socket (in price/performance) to commodity proprietary CPU's (instead of being the fastest performing socket in the market.)

[Dual-Die Photo, courtesy Oracle Hot Chips 27 Presentation]

Dual Socket Configuration:

The glueless dual-socket configuration allows for outstanding communication speeds between sockets. It is apparent that more bandwidth may be available over PCIe links than over the un-exposed Infiniband. Without the Infiniband exposed, the S7 looks more like an UltraSPARC IIIi.

[Infiniband Performance, courtesy Oracle Hot Chips 27 Presentation]

Unexposed Infiniband Performance:

The unexposed infiniband offered the possibility of significant packet performance improvement, over a PCIe card, even under high load. Note, the red Sonoma IB line mostly maintaining between 20-60 Millions of Packets per second.

A blade chassis connecting all minimal Sonoma "S7" blades (holding memory & 2 sockets) connecting back to infiniband storage in an MPP environment scaling across thousands of nodes could have been an amazing entry into Cloud Computing or Super Computing environments.

Conclusions:

While  NetMgt welcomes the new low-cost "Sonoma" S7 options, we see the form-factor produced as a "game changer" unrealized, by not placing the chip in a socket, exposing Infiniband, and into a chassis form factor which could most leverage it's strength. Such a socket could have easily replaced & out-performed the Intel based Storage subsystem Oracle's Infiniband native Engineered Systems. A such a socket in a blade chassis would have filled-out a gaping hole in Oracle's systems portfolio. Furthermore, the cost of memory is so high in the chassis that the cost difference between SPARC "Sonoma" S7 and SPARC M7 are marginal if selecting a chassis to perform LDom virtualization. It is a beautiful chip, for what it is, but a great opportunity lost.