Thread by @jwangARK

1/ The old Sun Microsystems adage: “The network is the computer” is perhaps the best lens to view Nvidia’s decision to acquire Israeli interconnect maker Mellanox for $7 billion.
bloomberg.com/news/articles/…

2/ The slowdown in Moore’s Law has forced the industry to migrate to parallel computing. The key to making lots of chips work efficiently together is fast interconnects.

3/ Nvidia has already been a pioneer of interconnects at the server level. NVlink allows multiple GPUs to work together in a single box. Last year Nvidia built NVSwitch, a custom silicon that allows 16 GPUs to work together in one server.

4/ This is still not enough. To train the world’s most complex neural nets—or better still—to use AI to search for neural net architectures—requires hundreds of GPUs. And the only way to hook up dozens of such boxes is to use fast interconnects such as Mellanox InfiniBand.

5/ For example the Summit and Sierra supercomputers—which have very similar designs to AI training clusters are based on lots of GPUs wired through NVlink and lots of nodes wired by InfiniBand.

6/ Fast interconnects are clearly very important for AI training. What’s not clear is why Nvidia has to *buy the company* rather than just buy the components as they do for CPUs, memory etc.

7/ Nvidia is not doing it for ‘diversification’ or ‘synergy’. That’s not Jensen’s style—Nvidia is not Broadcom. The only reason why Jensen would make an acquisition is if he believed it is highly strategic to the company’s future.

8/ We can see this through their past M&A history, which quite frankly, has been pretty shoddy though perhaps no worse than its peers. In each case there was great strategic merit, but execution and traction never followed.

9/ In 2006 Nvidia acquired PortalPlayer—the application processor provider of the iPod. At the time app processors were seen as the next big thing and Nvidia saw it as their next opportunity. Apple soon swapped out PortalPlayer for Samsung.

10/ In 2008 Nvidia acquired Ageia—a provider of game physics software and silicon in response to Intel acquiring Havoc a year earlier. The idea was to make the GPU the physics engine. Game developers never really bought in and the tech never gained strategic importance.

10/ In 2011 Nvidia acquired Icera—a software modem company based in the UK. Mobile was all the rage and Nvidia needed a modem for its Tegra SoC. Icera’s modem ultimately didn’t keep up with QCOM and Mediatek. The Icera team was laid off in 2015 (scooped up by Graphcore).

11/ Which brings us to today’s acquisition of Mellanox. At $7B, that’s 7x sales or 50x earnings—a rich multiple the result of a bidding war with Intel and Xilinx. But once again, what are they really buying?

12/ While InfiniBand is Mellanox’s crown jewels, it’s mostly a mature business. Ethernet is growing quickly thanks to a data center refresh to 25/50/100gbps adaptors, but that has no obvious synergies with GPUs.

13/ On the call with analysts, Nvidia emphasized not hardware but software integration. For example, GPU Direct allows GPU -> {GPU, memory, storage} transfers without going through the CPU. Bypassing Intel has been an existential goal for Nvidia forever. developer.nvidia.com/gpudirect

14/ In all goes right, a combined NV+MLNX tech stack could perhaps make GPU programming transparent and scale seamlessly across hundreds of servers, making the data center one giant “GPU brain”. That would be the vision.

15/ But it’s not at all obvious how this would work and given Nvidia’s M&A history, one has the right to be skeptical.

16/ A more cynical take is that Nvidia acquired Mellanox just so that Intel wouldn’t have it. Facebook’s manic $19B acquisition of WhatsApp was partly motivated by a desire to keep it out of Google’s hands. Google still has no viable messaging app today.

17/ With the #1 interconnect company in the bag, Nvidia now has 2/3 of the most important lego pieces to win in the data center. What it really now needs is a CPU…which perhaps will be what it acquires next. /end

Epilogue: @TDaytonPM has a great story on the acquisition with a direct quote from Jensen on why he's not interested in investing in CPUs.
nextplatform.com/2019/03/11/con…