It's not, really. It's been a while since I've checked pricing so my data might be old, but an IB switch is in the same ballpark as an ethernet switch with the same port speed. Same for HCA's.
There's no analogue in the Infiniband world to dirt cheap 1GbE RJ-45 switches though.
> an IB switch is in the same ballpark as an ethernet switch with the same port speed
And both price tags will make Elon's "someone's scamming me with a 'you're an enterprise customer' surcharge" sense tingle. The price tags for anything enterprise networking related are seriously inflated, and I would not be surprised if just making your own NICs and switches is cheaper once you hit a certain deployment size.
Networking has never been so cheap at the highest end. Look at the road we've been through in the last 8 in years, rapidly going from 40 to 100 to 400 (200 was somewhat of a dud, 400 came too early) to 800 to 1600Gbps. It's amazing.
I'm having trouble feeding things at 400GB/s (not a typo, it's gigabyte/s) per H100 box.
On the other hand, silicon has never been so cheap.
The hardest things would've been the DDR4 and PCIE interface. But as they're using standard interfaces, and last generation. I'm sure they got a good discount on all that IP and it didn't cost them hardly any man hours to integrate. And Tesla might've even already had the licenses and IP setup as they make other ASICs.
I didn't do a budget or anything, but at even 10Ks of units, I could see how this could save money. Or at least not loose money. Assuming a comparable IB network card is ~$1000, which I also didn't price.
And there could be other potential cost offsetting features, like power savings.
I've replied in a thread, and what I've replied to above was about the enterprise tax and that it's surely cheaper to do your own, not if the original article is about enterprise or not.
> > I would not be surprised if just making your own NICs and switches is cheaper once you hit a certain deployment size.
> Ideas you’ll never hear at Google or meta
You'd be surprised. Google has a very strong tradition of "not-invented here" which extends to some of our production networking gear as well.
To be fair, at the time, some of this was justified because the available devices on the market couldn't support our use cases back then.
Per section 3.2 of the 2013 B4 paper [0]:
Even so, the main reason we chose to build our own hardware
was that no existing platform could support an SDN deployment,
i.e., one that could export low-level control over switch forwarding
behavior. Any extra costs from using custom switch hardware are
more than repaid by the efficiency gains available from supporting
novel services such as centralized TE.
You couldn't have picked a better/worse duo in tech to be wrong on such an assertion.
Adding to sibling comment about Google, Meta[1] built 2 large-scale production training clusters for science: one with Infiniband, the other one with a custom RDMA over RoCE fabric.
> Custom designing much of our own hardware, software, and network fabrics allows us to optimize the end-to-end experience for our AI researchers while ensuring our data centers operate efficiently.
> With this in mind, we built one cluster with a remote direct memory access (RDMA) over converged Ethernet (RoCE) network fabric solution based on the Arista 7800 with Wedge400 and Minipack2 OCP rack switches.
Google, Meta and Netflix are among the most obsessive on optimizing their infrastructure - it's bold to assume they haven't looked at their COTS network gear and thought "hmmm..."
I don't believe the pricing is in the same ballpark.
There's also other differences, such as port counts.
AFAICT Spectrum switches at 400Gbps have up to 128 ports whereas equivalent Infiniband NDR Quantum only have 64 [0].
When building clusters of 32K+ GPUs the network cost, power, transceivers etc start to add up.
At least when I looked into it last, it was less about price and more about availability. NIC's are easy to get, but the switches were a 50+ week lead time. I'm sure it has come down now though at least for my business, we are just focused on ethernet now. We can do 128 NIC's @ 400G into a single Dell Z9864F using ethernet. Few will need proprietary solutions for this as there are lower hanging fruit.
