networkZONE Products for the week of May 21, 2007

Broadcom Corporation Says…
10 Gigabit Ethernet Connectivity Driven Into Mainstream Servers With Single Chip, Dual-Port 10GbE C-NIC
New converged controller extends Broadcom's NetXtreme II C-NIC family, adding a 10 Gbit/s solution, and completes its 10GbE portfolio featuring market leading C-NICs, enterprise switches and physical layer devices

Broadcom Corporation has announced the industry's first true single-chip, dual-port 10 Gigabit Ethernet (10GbE) converged network interface controller (C-NIC) specifically developed for high volume server designs. Leveraging two field proven generations of Broadcom NetXtreme II Gigabit Ethernet (GbE) C-NICs, the introduction marks an important milestone for Broadcom -- successfully delivering the industry's first fully functional, single-chip C-NIC at 10Gbit/s rates, with no external memory required. This new device completes the company's portfolio of 10GbE end-to-end solutions featuring market leading C-NICs, switches and physical layer devices (PHYs), enabling OEM partners to enhance next generation servers with a complete portfolio of 10GbE network infrastructure solutions from Broadcom.

In today's enterprise network, a standard server equipped with current Ethernet controller silicon is typically limited to 1Gbit/s rates. Moving to support a 10 times speed improvement to 10 Gbit/s rates would require additional support from the host CPU to process the data. This causes significant increase in the amount of processing power CPU would have to dedicate to communication versus its primary duty of running applications.

To address this challenge, Broadcom's new 10GbE C-NIC enables simultaneous processing of network, storage, processor clustering and management traffic on chip while enabling convergence of different traffic types over a single Ethernet fabric. By supporting Microsoft's Windows TCP chimney engine, iSCSI block storage and remote direct memory access (RDMA), the new C-NIC enables network protocol processing on-chip, thereby saving the server's CPU and memory I/O resources to perform their primary tasks -- running applications. As a result, IT professionals can simplify their network designs by providing network, storage, clustering and management capabilities over existing and familiar TCP/IP and Ethernet infrastructures while boosting server performance through added network bandwidth and improvements in CPU utilization.

"Enterprise connectivity today is ready for a speed upgrade as the number of GbE ports required to support the server's bandwidth need has increased beyond 4GbE, driven by virtualization and I/O intensive applications running on multi-core CPUs. The market is now ready to adopt 10GbE as the next plateau for industry standard server connectivity with blade servers leading the transition," said Greg Young, Vice President and General Manager of Broadcom's High-Speed Controller line of business. "The increased bandwidth (to 10GbE) also provides an opportunity to converge disparate networking fabrics (data, storage, management, and processor clustering) by providing sufficient bandwidth for all traffic types. It is expected that most server OEMs will migrate to 10GbE as the new baseline for connectivity."

Announced is the Broadcom NetXtreme II BCM57710 -- the industry's first fully functional dual port C-NIC that runs at 10Gbit/s rates and supports small form factors and PCB footprints necessary for mass adoption and deployment. The BCM57710, a highly cost-optimized solution, is also part of the NetXtreme II C-NIC family -- that allows BCM57710 to provide an easy migration path to 10GbE by leveraging the NetXtreme II C-NIC family software already deployed in more than 70% of enterprise servers shipping today.

"Broadcom continues to define and drive Ethernet technology for the enterprise server space and has introduced many technological firsts over the years," said Bob Wheeler, Senior Analyst for The Linley Group. "The company is now taking its Ethernet solutions portfolio to the next level by including 10GbE offerings enabling a continuation of its technological leadership in Ethernet, and in particular, the enterprise server space."

Broadcom's new BCM57710 10GbE C-NIC is designed for bidirectional line rate performance, not seen from any other 10GbE device to date. A server running Broadcom's 10GbE C-NIC technology with Microsoft Windows Server 2003 with TCP Chimney shows performance gain up to 6 times with C-NIC over standard layer 2 -- with layer 2 delivering 4.7 Gbits/s at 99% CPU utilization while C-NIC technology with TCP Chimney support delivering line rate performance with 33% CPU utilization.

"Through the combination of our Scalable Networking technologies for Windows Server 2003 and the upcoming release of Windows Longhorn Server 'Longhorn,' IT professionals can cost-effectively easily benefit from 10Gbits/s performance by enabling offload with the Broadcom's BCM57710 C-NIC," said Henry Sanders, General Manager of Windows core networking group, Microsoft Corporation. "Microsoft and Broadcom have enjoyed a long relationship developing TCP offload solutions and we are pleased to now enable bidirectional line-rate offload at 10Gbit/s speeds."

"Broadcom's flow-through architecture for 10GbE C-NIC is a great achievement. The ability to enable a 10GbE C-NIC in a small package without requiring external memory is ideal for space-optimized server applications requiring 10Gbit/s networking speeds," said Steven Hunter, Distinguished Engineer & CTO, IBM BladeCenter, IBM Corporation. "IBM currently offers 1Gbit/s C-NIC solutions in its servers and looks forward to expand its current 10Gbit/s offering in the future. Having the protocol processing occur on the C-NIC enables line rate networking bandwidth without over-burdening the host resources."

"We are pleased to collaborate with Broadcom to validate PCI Express version 2.0 designs and to provide our common customers a smooth transition to our next generation technologies," said Jim Pappas, Director of Technology Initiatives, Intel Corporation. "PCI Express version 2.0 increases bus speed from 2.5 Gbits/s to 5 Gbits/s per lane enabling much higher I/O densities for next generation servers."

"Broadcom's BCM57710 adopts an innovative follow-through, no external memory architecture and has a full suite of software," said Vinod Lakhani, Director of Marketing for Broadcom's High-Speed Controller line of business. "Since competitive solutions do not include full C-NIC functionality, they lack some level of protocol processing or require additional external components that cannot be integrated in a single-chip design. As a result, today's host CPUs are busy terminating protocol stacks in the absence of on-chip protocol processing technologies, making Broadcom's 10GbE C-NIC technology essential for next generation servers."

Technical Information
Broadcom's single-chip BCM57710 C-NIC converges TCP, iSCSI, RDMA, and remote management traffic over 10GbE with PCI Express (a high-speed peripheral interconnect technology designed to accommodate higher speed CPUs) interface to the host. The BCM57710 includes key features that set it apart from the competition and is based on a flow-through architecture that eliminates the need for external memory, which is essential for lowering the cost of 10GbE to enable its widespread adoption. The BCM57710's intelligent and innovative protocol stack processing completely eliminates the need for large memory banks otherwise required for 10 Gbit/s data rates support. It is also optimized for line rate performance for small I/O sizes which are predominant in the data center and a challenging achievement for a 10GbE C-NIC solution.

The BCM57710 is also optimized for low latency that is not available from any other 10GbE controller solution. Clustering applications benefit directly from low latency and the BCM57710 lower total cost of ownership (via Ethernet) makes it a preferred solution for RDMA applications.

The BCM57710 supports a complete ecosystem for server LOMs. This ecosystem includes features such as iSCSI boot, pre-boot execution environment (PXE), and universal management port (UMP) or DMTF's Network Controller - Sideband Interface (NC-SI) to enable interfacing to baseboard management controllers (BMCs). Broadcom's BCM5221 single-port 10BaseT/100Base-TX transceiver is also available to connect directly to a system management processor compatible with the DMTF NC-SI specifications.

EN-Genius Says . . .

In reviewing Broadcom's dual-10GbE port, iSCSI controller and RDMA offload engine, I find myself in a real predicament of having news about a potentially significant product but insufficient information to give you a proper analysis of it. As anyone who is a regular reader of networkZONE knows, Broadcom often puts out some great products but has a tradition of reluctance to share meaningful details about them that has made their product reviews a challenge. Unfortunately trying to review the BCM57710 10GbE C-NIC has been especially frustrating because my briefer shared even less beyond the sparse information in his PowerPoint presentation than usual. This being said, I'll give you the small tidbits I was able to obtain and give you some idea of what I think the controller is really capable of. I'll also take Broadcom to task for what I feel is an unfair comparison of its part against a so-called competing product.

The BCM57710 builds on technology developed for their 2004 NetXtreme II 1G controller. Since the block diagram in the briefing (see Fig. 1a) was utterly useless, I extracted something a bit more helpful from their product brief (See Fig. 1b) to help me make sense of this very complex chip. It's difficult to tell whether the interconnect bus scheme shown in the detailed block diagram is anywhere close to the physical reality of what goes on inside the chip but it at least gives you a rough idea how the functional blocks are connected.

If the diagram in Fig. 1b can be believed, the controller's designers have taken a very straightforward approach to delivering a unified network that efficiently transports storage, network, management and compute cluster traffic over the same Ethernet pipes with a minimum of overhead or blocking. Its RDMA elements appear to be dedicated hardware cores that relieve the host system of much of the burden of TCP overhead processing and free up precious bandwidth across the 8X PCIe host interface as well.

Unfortunately, the sparse information I got is not enough to give me a warm fuzzy feeling about several key elements of the device. For example, Broadcom claims that their custom packet engine is able to sustain line rate on all connections and that the bottlenecks that occur are on the PCIe side simply because the 8x interface has bus transaction overhead and line coding loss that don't let it quite handle the full 20 Gbit/s worth of traffic. The only other information I was able to get came straight from their slide set that made the following claims:

• Line rate performance with Layer 4 offload
• Line rate performance for Layer 2
• Optimized for small I/O sizes that are very important in data center settings
• Full IPv6 offload support
• Low power -- 7W

I was disappointed that the product manager I talked to would not go into the architectural details which would reassure me that this performance wasn't just true for a limited set of conditions and that that it would choke on everyday corner-cases such as lots of short packets, heavy encapsulation, or other less-straightforward challenges. There is some level of confidence to be granted in their favor, from their claims that it has been demonstrated for customers and from their invitation to see the chip run for myself (if I was willing to come to Las Vegas for Spring Interop -- my least favorite show in my least favorite town). This is tempered by the fact that demos are generally set up to show a product in the most favorable manner possible, and I've that seen enough impressive seeming hardware set-ups running at trade shows (including several by Broadcom) that are specifically designed to obscure many inconvenient truths.

I am a bit more confident in their claims for their offload engine which provides full host offload for Microsoft TCP Chimney-compliant operation. When pressed, my briefer said the chip should also be able to support non-Chimney/non-Microsoft offload work, although some details are in flux as the Linux community converges on a preferred solution. It's a bit tougher to completely buy their claims that that their RDMA offload mechanism can run simultaneously with TOE and, in fact, all protocols can run at once without any further information.

With the high levels of paranoia and hype that have crept back into the networking market over the last few years, it's tempting to forgive some of the problems I had getting a straight picture of what, exactly, is in this chip, and how it works. What is much harder to overlook is what seems to be a deliberately-mismatched comparison between the BCM57710 and a competitor's chip that is supposed to be roughly equivalent by picking NetXen's NX2031, which the company describes as an intelligent NIC chip (see Fig. 2 -- lifted from my briefing slides, but not under NDA)

For one thing, Broadcom takes NetXen to task for using external memory for store-and-forward process, a choice which usually results in more latency and some increased power consumption. This is a somewhat disingenuous claim because NetXen does a lot of packet inspection and processing in addition to its TCP offload duties (see the September 2006 review) -- functions that the Broadcom device does not even pretend to support. This could well be the thing that makes the NetXen chip require more memory than the buffering used by the BCM57710 iSCSI and TCP offload functions. I'll be the first to admit that Broadcom is absolutely brilliant at integrating stuff, including memory, but the disparate nature of the jobs they do, and lack of information about their embedded RAM size, makes any meaningful comparative arguments about NetXen shortcomings difficult at best, and of questionable value.

I'm also having difficulty figuring out where Broadcom came up with a power consumption figure of 20 - 21 W for the NX2031 when the figure I was given was 10 - 11 W (plus a bit more for their external DRAMs). Given the fact that there is no mention of power on their current product brief, it's quite likely that NetXen gave me an overly-optimistic estimate, but it's hard to believe they would be off by 100%.

If it does anything close to what it claims to be capable of, the BCM57710 could be a secret weapon that makes the company a fortune as it blows away the bottlenecks in blade servers, storage systems and the like. Unfortunately, between the fact that the BCM57710 and the NX2031 have such different functionality and the power numbers that seem to contradict whatever information I have, it's difficult to understand what Broadcom had in mind when they made the comparison.

If I was of a more suspicious nature, I'd wonder if this was not a deliberate mismatch, calculated to be picked up and passed on by a busy journalist who does not have time to do a little fact checking. One could also speculate that Broadcom's unwillingness to share details was in good part due to the fact that the silicon is not quite ready for prime time and they have issued a very early release in an attempt to lock up the market until they get it right. It certainly would not be the first time a major chip maker tried to game the market with this tactic although I would hope my friends at Broadcom would be above such tactics.

Regardless of its particular technical merits, many of the ideas behind the BCM57710 that are intended to make 10G unified networks practical and affordable will probably end up helping drive the technology into all form factors, including blade servers, NIC cards, embedded compute platforms or iSCSI NAS storage boxes. Let's hope the chip lives up to its promises so Broadcom can reap some of the rewards for all their good efforts.

With so little hard information provided to me, I can't draw any firm conclusions and decline to post a Saltshaker Rating for Broadcom's BCM57710 at this time.

Broadcom says the BCM57710 10GbE C-NIC is sampling to early access customers. Volume production is expected during Q4 of 2007. OEM volume pricing will be less than $100.

Product Page