Broadcom announced June 11 its customers are now sampling a new Trident 4 compiler-programmable network switch that relies on a 7 nm chip for the first time.
While the 7 nm chip alone might seem to garner the most excitement from the industry, it is the programmable switch capability with Broadcom’s open source Network Programming Language that should entice enterprise customers the most.
With programming prowess, Broadcom raises its competitive profile against Barefoot Networks and its P4 network functions language, especially now that Intel just announced it will purchase Barefoot.
“I do think Barefoot becomes a more significant competitive threat [to Broadcom] as part of Intel,” said Bob Wheeler, a networking analyst for The Linley Group. Customers will be assured of an ongoing supply of products and a clear roadmap with the Intel acquisition, meaning that Broadcom “absolutely” should worry about Barefoot as part of Intel, he added.
But Broadcom is safe for now. It currently controls about 80% of the $3 billion merchant Ethernet switch market, according to Linley.
Asked how big a threat Barefoot is, Broadcom principle product line manager Robin Grindley chuckled and said, “We have a dominant share of the Ethernet silicon market and we’re pretty confident.”
The advent of programmability with Broadcom and Barefoot is pronounced for enterprise users, moreso than for cloud providers, Wheeler said. “The compelling reason for programmable is that it’s future-proofing,” he explained. “If there’s a new protocol and your enterprise wants to take advantage of it, you don’t have to buy new hardware and can get a software upgrade from the OEM, which enables the protocol without a huge hit. That’s good for tunneling protocols for virtualizations or enabling new features.”
Programmability can also help make quick adjustments to protect against a DDoS attack or other cyber exploit. And, it helps with telemetry so a network analyst can study and react to network traffic problems where packets are being dropped or app performance is poor.
By contrast, hyperscale cloud providers want “as high a bandwidth as they can get” and are less worried about programmable capabilities, Wheeler said. “They work very fast and relatively simply and then pack the buffer memory.”
Even so, Broadcom’s Trident 4 family ranges across switches with 2 to 12.8 Tbps aggregate bandwidth. The chip has 21 billion transistors and manages up to 5 billion packets per second on a single chassis. Barefoot, meanwhile, has recently begun shipping a 16-nm Tofino chip.
Broadcom’s Grindley said that Trident 4 with NPL is truly programmable, comparable to what Barefoot’s Tonfino and P4 offers. Broadcom’s approach is compiler programmable and supports run time programmability, Grindley was careful to say. “If you want run time to change something on the fly, you have a table and you add something in that table and add something as it’s running, so we support run time as well,” he said.
Wheeler said the Trident 4 approaches the same level of programmability of the Tofino chip, but he said if frankly difficult to make an exact comparison because of somewhat limited disclosures by the companies.
Broadcom also argued it can save enterprises 75% of the capex cost of OEM gear, taking three chassis down to a single chassis with a single chip. In one PowerPoint slide, Broadcom showed that three OEM chassis would cost $244,736, but would be reduced by 75% with a single Trident 4 chip in a single chassis. Wheeler said Broadcom’s approach might be less costly, but added, “I’m not really buying that 75% savings” because Broadcom is comparing its new chip to two-year-old OEM gear that will three years old once Trident 4 actually ships in about a year.
More than anything, Broadcom’s T4 shows that programmable chips are the way to go when compared to FPGA (Field Programmable Gate Array). Broadcom’s Grindley argued that FPGA can still be relied upon but only “in theory” since “the power efficiency and number of ports you can cram in would be much lower and power would skyrocket.”
Wheeler was more emphatic. “No, no, no. FPGA and a programmable chip don’t even come close,” he said. “The only place FPGAs are interesting is at the network edge in a data center interconnect of two data centers. There you might need do something like encrypting a link and I could see using an FPGA. But FPGA is not competitive with these programmable chips.”