Hacker News Reader

> You could pave a road with the gravestones of technology vendors who pursued industry dominance with faster, more expensive chips. If IBM, Sun, Unisys, HP, Digital, SGI and a dozen others failed, why would a company with almost zero datacenter presence succeed?

I have very much the same opinion, and I think that Huang is at least aware of that.

> Nvidia would probably pursue a similar strategy -- sell chips with lots and lots of cores and lease capacity on demand.

But this will be pretty much the same IBM thing. Some banks, and Oracle/SAP buyer types may buy in, but that will be it. It does not change anything to how the companies in the above list of "big iron" vendors attempted to do it before.

I want to remind that Huang is also, a very big "investment relations" player, and he likes to make an impression of "grandiose" plans to impressionable investor guys.

> impressionable investor guys.

Impressionable investor guys tend not to be investor guys for very long. Big-vision bluster can can work with media and some end users but it turns off real market analysts.

Seems to be working fine for Elon, Miyoshi Son & gang.

Elon seems to have delivered as opposed to Son, no?

Huh? Mellanox (now part of Nvidia) has a massive data center presence, and it's growing their NICs are leaps and bounds above Intel's.

Have you used Infiniband? 40Gbps cards are $30 on ebay, right now for tech that is what 3 generations back. Currently shipping gear is 200Gbps. Everything flows from, in, out or to the network. Cores don't matter. All future workloads will be done by GPUs.

Sorry for the snark, but what if my workload contains branches?

GPUs have excellent support for branches. Pascal has divergent thread execution, Volta added forward progress guarantees, Ampere improves on this further.

There is a lot of research about doing databases on GPUs, and Apache Spark runs on GPUs today, with much higher performance than on CPUs.

GPU databases are limited primarily by memory constraints on the card (e.g. ~32GB maximum per card for GV100 or whatever) and interconnect latency/bandwidth, not by raw parallel scan speed. If scan speed was all that mattered, we'd have had GPU-like parallel database hardware decades ago. You can crunch rows, but only as long as it fits in memory. Once your working set exceeds the provided RAM and has to page out data to the CPU over PCIe or some other link, the numbers and utilization begin looking much worse. "Every benchmark looks amazing when your working set fits entirely in cache."

But even more than that, for the price of a single high end Tesla (approx. 10k USD), you can build a high-end COTS x86 machine with a shitload of RAM, NVMe, and then install ClickHouse on it. That machine will scale to trillions of rows with ease and millisecond response times, whether or not everything fits in memory. It will cost less money and also cost less energy and it will scale out easier, and have better utilization of the hardware.

I'd wager that unless you have infinite money to dump on Nvidia or exceedingly specific requirements, any GPU database will get soaked by a comparable columnar OLAP store in every dimension.

> GPU-like parallel database hardware decades ago

https://en.wikipedia.org/wiki/Netezza#Technology

Furthermore, that line of reasoning is fallacious. The non-existence of a thing doesn't prove that the idea is bad.

I have in HPC and backup scenarios.

I also ran datacenter services on Sun and IBM hardware that smoked anything offered by Intel platforms. But at the end of the day, commodity beats premium for 80%+ of the market. We run Linux on POWER exclusively as a cost-savings mechanism to reduce the Oracle bill.

I have been prototyping all of my new applications in context of commodity x86 with 10GbE links and NVMe storage.

I personally feel like distributed (coherent) shared memory is a bit overkill. I would much rather the kind of grain that a solid 10gbps link works well with. I.e. anything that is sensitive to latencies measured in microseconds or lower just lives on the same node where it needs to be sampled.

I can't think of too many applications that couldn't be designed to work just as well using 10GbE vs Infiniband/RDMA/shared memory. Even most real-time simulations could be designed to run in a practically syncrhonous fashion across a number of distributed hosts using stuff you can buy at Newegg or Amazon.

In exactly the same way, "treat your servers as cattle, not pets" is massive overkill for your average wordpress site, but Amazon pulled an entire industry out of that hat.