Why Ethernet is the Winner: A Love Letter to AI Networking’s Next-Gen Beast, the Tomahawk 5

There’s something almost poetic about the march of technology — the way it pushes boundaries, defies expectations and revolutionizes everything. As someone who’s been in the trenches with these beasts for years, I’ve seen this evolution firsthand. And let me tell you, there’s nothing quite like the thrill of watching the Broadcom Tomahawk series grow from its humble beginnings to the powerhouse it is today.

A Decade of Unrivaled Innovation: From Trident to Tomahawk 5

Broadcom’s journey from the Trident to the Tomahawk 5 was nothing short of a technological rollercoaster. A thrilling ride spanning over a decade, filled with twists, turns, and lots of doubling (seriously, those folks at Broadcom seem to love doubling things). Let’s dive into the timeline that got us here:

1. Trident (2010): 640 Gbps Ah, 2010! A simpler time. The Trident series marked a new era in Ethernet networking with a capacity of 640 Gbps. And who could forget the legendary Edgecore 10G switches like the AS5600-52X switch with Trident+ and a Freescale CPU, still boasting 4 x 40G uplinks? Ah, those were the days! I still remember the smell of fresh silicon and the sound of those switches humming in the data center like it was yesterday!
2. Trident 2 (2012): 1.28 Tbps Fast forward to 2012, and Broadcom had doubled the bandwidth with Trident 2, bringing us to 1.28 Tbps. Those were the times when the AS5712-54X switch made its debut — 48 x 10G with 6 x 40G uplinks and an Intel CPU! What a beast and smarter, too! (Cue a little nostalgia for those good old tech days.)
3. Tomahawk (2014): 3.2 Tbps Then came 2014, a year that holds a special place in my heart. Why? Because it’s when I joined Edgecore, right when we launched the AS7712-32X, our first 32 x 100G switch on the mighty Tomahawk! And guess what? It’s still in production today — proof that a great design never dies!
4. Tomahawk 2 (2016): 6.4 Tbps By 2016, the Tomahawk 2 had doubled its predecessor’s capacity to 6.4 Tbps. This was when the famous AS7816-64X burst onto the scene — offering 64 x 100G from a single chip! Talk about efficiency! And yes, they’re still going strong, proving their mettle day in and day out.
5. Tomahawk 3 (2018): 12.8 Tbps Oh, 2018… when Tomahawk 3 strutted in, doubling the capacity to 12.8 Tbps and introducing us to the 400G era with 32 x 400G. Remember the feeling of unboxing those beauties? It was like opening a treasure chest!
6. Tomahawk 4 (2020): 25.6 Tbps In 2020, while the world faced… well, everything else, Tomahawk 4 quietly doubled its capacity to 25.6 Tbps. Imagine that — doubling while everything else seemed to be halving! Now we had 64 x 400G or even 32 x 800G setups — just like the wizards over at Celestica could provide!
7. Tomahawk 5 (2022): 51.2 Tbps And then came 2022, when the Tomahawk 5 hit the scene with a jaw-dropping 51.2 Tbps capacity, once again doubling the previous generation’s bandwidth. Now, we’re talking 800G — 64 ports of it, no less! I mean, what’s not to love?
8. Tomahawk 6: The Future is Near And if you think the story ends there, think again. The Tomahawk 6 is already on the horizon, promising yet another doubling — 64 x 1.6T! I can’t wait to see which ports will become the stars of this show — will it be OSFP224 or QSFP-XD? Let the race begin!

Figure 1: Tomahawk: Ethernet Network for AI at Scale.

(And yes, yes, I know there’s also the Trident 3 family that came along in the meantime, from the X2 to the X7, and the quirky Trident 4 — and there’s even a Trident 5 lurking somewhere in the near future! But for now, let’s stay focused on the Tomahawks 😉)

Also, I know we might be a bit delayed with all these dates… — it takes time from the first chip announcements to when the actual hardware rolls out. But the good news is that you can already get your hands on Tomahawk 5 systems this year, not just from Edgecore but also from Celestica, Dell, UfiSpace and others! Whether you’re looking at the powerhouse 64 x 800G configuration or opting for the more modest (but still impressive) 32 x 800G setup if the first is a bit too much, I’ve got a feeling we’ll see even more exciting port configurations coming soon like native 400G: 128 x 400G — why not!

Key Drivers of Broadcom’s Success

What’s Broadcom’s secret sauce, you ask? It’s a mix of several key ingredients:

• Efficient, Scalable Architecture: Designed to expand seamlessly with our ever-growing hunger for data.
• Leading-Edge Process Technology: Using the latest semiconductor tech to crank out better performance and lower power consumption.
• Best-in-Class Custom Physical IP: Optimized for high performance and efficiency, like a custom-tailored suit for your data needs.
• Physical Design Expertise: Delivering products that meet the highest standards of reliability and performance — and trust me, I’ve put them to the test!

80x Bandwidth Increase and 90% Improved Energy Efficiency: What It Means

Now, let’s talk about numbers. From the Trident to the Tomahawk 5, Broadcom has achieved a mind-boggling:

• 80x Bandwidth Increase: Allowing hyperscale data centers to handle massive data volumes with fewer devices. That means reduced capital and operational expenses (and more budget left for the really fun stuff).
• 90% Improved Energy Efficiency: Lower operational costs and a greener footprint. It’s like getting the best of both worlds: performance and sustainability.

Figure 2: Tomahawk 5 Advantage — 2x Bandwidth, 6x Fewer Switches, and Much More!

But wait, there’s more! When Broadcom doubles that speed, it’s not just about replacing two old switches with a new one. No, no, no! Think about it: to achieve the same throughput under a non-blocking topology, you’d need to connect not two, but six old switches. We’re talking massive savings here — fewer licenses, less rack space, reduced power consumption, fewer optics and cables, and simpler automation. Simple math, but worth shouting about!

Why Tomahawk 5 is the Ideal Choice for AI Clusters

Unmatched Bandwidth and Port Density

Broadcom’s Tomahawk 5 supports 256 ports of 200G, doubling the capacity of competitors like NVIDIA’s InfiniBand, which maxes out at 128 ports with a 25.6 Tbps chip. This translates to more traffic management with fewer devices, reduced infrastructure costs, and enhanced network efficiency. (Translation: more bang for your buck!)

Superior Scalability

Do you like math? I do! So, let’s geek out a bit:

• 256 Nodes @ 200G:Tomahawk 5: 1 Switch, 1 Tier InfiniBand: 6 Switches, 2 Tiers
• 32K Nodes @ 200G:Tomahawk 5: 384 Switches, 2 Tiers InfiniBand: 1,280 Switches, 3 Tiers

See what I mean? Twice the scalability per device, over three times fewer switches than InfiniBand, and a fourfold improvement in system efficiency for 2-tier scaling. It’s like magic, but with data! And again – it’s not only about the number of switches, but also the required connectivity, software, power, and so on.

Why Single-Chip Tomahawk 5 Outshines Multi-Chip Rivals

Now, I know what you’re thinking — “But wait, how does this compare to NVIDIA’s Quantum-X800 InfiniBand?” Great question! And here’s the thing: when I make this comparison, I’m talking specifically about a single-chipset system. You see, the beauty of the Tomahawk 5 is that it achieves all this unmatched performance, scalability, and efficiency using a single chip. Meanwhile, platforms like the Quantum-X800 typically rely on multiple chipsets to handle complex tasks like adaptive routing and congestion control. But with Tomahawk 5, we’re getting all this power and more, without the extra hardware complexity. It’s like comparing a finely-tuned sports car to a bulky, multi-engine truck — sure, they both get the job done, but the elegance of the single-chip Tomahawk solution is something to behold!

A Note on Port Configurations

Now, before we get too carried away with all this scalability talk, it’s worth mentioning that the Tomahawk 5 does have a limitation in terms of maximum logical ports: 320. So, while each of those 800G ports is incredibly powerful, you can use them in different configurations — either as 800G, split into 2 x 400G, 4 x 200G, or even 8 x 100G or 8 x 50G — but the total number of logical ports you can configure is capped at 320.

So, yes, it’s like having an all-you-can-eat buffet, but you only have so many plates. Choose wisely, and you’ll still come out a winner!

Enhanced Energy Efficiency

With a 75% reduction in power consumption, Tomahawk 5 is not just fast; it’s also green. A smarter choice for next-gen data centers and a step towards a sustainable future.

Cognitive Routing: Smarter Networks for Smarter AI

Tomahawk 5 brings some serious brainpower with its cognitive routing features:

1. Global Load Balancing: No more bottlenecks — this baby knows how to spread the load.
2. Reactive Path Rebalancing: Always on its toes, adjusting paths to keep things flowing smoothly.
3. Fast Link Failover: Like a ninja, it reroutes traffic around failures in a flash.
4. Drop Congestion Notification: Sends trimmed packets with metadata to keep everything running smoothly and proactively.

Ethernet vs. InfiniBand: A Clear Choice for AI Networking

InfiniBand, it’s been real, but Ethernet is the new star of the AI networking show:

• 10%+ Improvement in Job Completion Time: Faster job completion means quicker AI training and deployment.
• 30x Faster Failover: Ethernet recovers in 53 milliseconds compared to InfiniBand’s 1600 milliseconds. That’s blink-and-you’ll-miss-it fast!

Industry Adoption: Why Tech Giants Choose Ethernet

Just ask the big players:

• Amazon: 60,000+ Ethernet nodes powering AWS and Alexa.
• Oracle: 30,000+ nodes boosting cloud services and real-time analytics.
• Meta (Facebook): 20,000+ nodes supporting AI research and the metaverse.
• ByteDance: 10,000+ Ethernet nodes driving TikTok’s magic.

Broader Industry Implications: The Rise of Ethernet Across Sectors

It’s not just the tech giants:

• Financial Services: Real-time analytics and trading algorithms love Ethernet.
• Healthcare: Faster diagnostics and personalized treatments? Yes, please!
• Automotive: Real-time data processing for self-driving cars? You bet!
• Retail and E-commerce: Smarter supply chains and customer experiences.

Growing Ultra Ethernet: A New Frontier

Figure 3: New Members of the Ultra Ethernet Consortium (August 2024) *Please note that not all new members are displayed on this page. Full list available: Ultra Ethernet Consortium.

And Ethernet’s evolution doesn’t stop there. The Ultra Ethernet Consortium (UEC) is making big waves, welcoming 40 new industry leaders like NVIDIA, Lenovo, and Micron Technology, to name just a few, since March 2024. This expansion strengthens the UEC’s mission to deliver an open, interoperable, high-performance Ethernet-based architecture. With a total of 97 members now, the consortium is better equipped than ever to meet the escalating network demands of AI and HPC at scale. The future of Ethernet looks ultra-bright!

Broadcom Tomahawk 5: The Future of AI Networking

The Tomahawk 5 is not just another Ethernet switch; it’s a revolution. With 51.2 Tbps of bandwidth, exceptional port density, and cognitive routing, it’s ready to take on today’s AI workloads and whatever comes next.

Ethernet as the De-Facto AI Network: A Growing Consensus

More scalability, better performance, and lower costs — no wonder Ethernet is the new standard for AI infrastructure.

The Future is Private: A Vision from Hock Tan, Broadcom’s CEO

In a keynote at VMware Explore 2024, Hock Tan, President and CEO of Broadcom, emphasized the shift toward a future centered on private infrastructure: “Private cloud, private AI, your own private data.” Companies are re-evaluating the costs, complexities, and compliance challenges of the public cloud, with 8 out of 10 CIOs moving workloads back on-premises. And I couldn’t agree more — in a hybrid world, the private cloud is where true innovation and control thrive.

Conclusion: Why Ethernet is the Winner

So, why is Ethernet the winner? It’s simple. Ethernet, with the Broadcom Tomahawk 5 leading the charge, offers unmatched scalability, superior performance, and cost-efficiency that other technologies, like InfiniBand, just can’t beat. It’s the network standard that grows with your data needs, from AI research labs to hyperscale data centers, while keeping costs in check and energy consumption low. Ethernet delivers faster job completion, better failover times, and greater adoption across industries — proving it’s not just the present, but the future of AI networking.

When it comes to building the backbone of tomorrow’s AI infrastructure, Ethernet is the clear choice. It brings together innovation, reliability, and a commitment to pushing the boundaries of what’s possible — all in one neat, powerful package.

And Just to Hammer It Home…

Think of Ethernet like the underdog sports team that trains harder, plays smarter, and ultimately, leaves the flashy players eating its dust. InfiniBand may have its day, but Ethernet is here to stay — running circles around the competition, saving you time, money, and a lot of headaches. So, let’s cheer for the true champion: Ethernet. It’s faster, stronger, and definitely cooler than you might have thought!