12.8 Tbps Line Rate L3 Ethernet Switch ASIC
- tailored to the needs of hyperscalers
- Tomahawk 3 is the world’s first 12.8Tbps switch chip, now sampling to customers, that will accelerate the next
- wave of 100GbE/400GbE hyperscale cloud build-out
- high port radix requires low latency,
- buffer architecture,
- 40% less power,
- 75% cheaper cost per 100Gpbs through the use of 50G PAM4 signalling
- Up to 256 x 50G-PAM4/25G-NRZ Dual-Mode Serde
Shrinks an OCP Backpack switch (which is basically 6 x 32 100 front facing switches with an internal spine) into a single switch. So what used 10 ASICs before, is now a single ASIC unit.
- architected for hyperscale – performance oriented, latency, L3/label switched fabrics, load balancing.
- But no reqt for programmable data plane.
- driven by Deep Learning / ML / AI
- Hyperscale DC market opportunity -growing
- NVMe over fabrics – the use of SSD in distributed storage means that network performance must be low latency
- Storage advances – NVMe over Ethernet
Compare Trident 3
- Trident 3 – has a programmable pipeline and lower latency but matters less.
- Tomahawk 3 at same power and cost but twice the bandwidth (but no programmability)
- Note that high radix, multi-hop Ethernet is their operation. HPC is building Slim Fly/DragonFly
- 16nm die process, gives an 18 month advantage (?)
Interpreting what this means
- focussing on low latency so its not a flexible architecture for overlays etc needing programmability. You likely won’t see this in the Enterprise (although Cisco could waste money building it into a spine only unit and claim its got magic inside)
- building nearly dedicated ASICs for cloud providers is a reflection of how much they are spending on data centres and the need to build products that are suited to their exact requirements.
- Broadcom has three ASIC families : Tomahawk is for speed and port density. Trident is for programmable features at medium density – ie data centres. Jericho is an ASIC targeting carrier grade needs such as routers.
- See Broadcom is the incumbent, it doesn’t have to be better than the competition. It needs to be good enough so that people don’t switch away (see MS Windows). With Marvell buying Cavium we are seeing consolidation in the ASIC market.
- The ASIC market is a scale up business, the larger volume you can manufacture the better your competitive advantage and that why Broadcom bought Brocade to scale up its Fibrechannel ASIC business and reduce costs.
- I don’t believe that Cisco can compete with Broadcom and the custom cloudscape chips on Nexus (which might be a slightly customised version of Marvell silicon) doesn’t seem practical anymore when most people are focussed not he software defined parts and not on the hardware. In Enterprise, most people haven’t even gotten to 10G so a 32 port 400G silicon isn’t relevant.
- If you are in the Enterprise, this likely doesn’t matter since most of you don’t need more than 2 switches anyway. Having 2 switches with 32 x 400G ports with limited features but going very very makes no sense because you haven’t scaled out.