201–210 of 229 posts
Earlier quoted context omitted.
Infiniband is actually a standard and Mellanox became the go to supplier. However in the early stages there were many more suppliers.
Intel bought one of the IB players and for all I know killed it.
Interesting fact, Mellanox is an employer of Palestinian programmers in the occupied West Bank and Gaza Strip. Sort of a mixed blessing, cheap labor without other options because of the occupation, but real opportunities and a more educated and well off population will be more effective at advocating for its rights: https://www.haaretz.com/israel-news/business/palestinian-hig...
Earlier quoted context omitted.
"real technology" is not an objective measurement of societal value.
Except yes it is, because all these software apps are built on prior “real technology” and would be impossible without it. The flow of capital just doesn’t reflect that (and it’s not clear it should).
This is more like a race car company not being as valuable as Toyota. Yes, the technology of the race car is unbelievable, but there are only a few thousand race cars in the whole world and there's a Toyota in every household.
Earlier quoted context omitted.
Yup, but that's exactly why IB isn't too relevant anymore. Ethernet had plenty of time to catch up because of PCIe-SIG and Intel dragging their feet.
It doesn’t matter nearly all Mellanox ASICs support both IB and Ethernet they are the ones who are driving the speed of Ethernet to this level. You also need to use their cables and transceivers (or a similar alternative) for these speeds doesn’t matter if you are using Ethernet or IB.
Earlier quoted context omitted.
Maybe just maybe this is not the fault of Nvidia but due to the fact that large parts of the Linux ecosystem are fragile, time consuming to configure and break if you look at them in the wrong way. Professional linux distributions like Ubuntu paper over a lot of that fragility, whereas in Arch Linux you can easily burn days getting basic functionality to work (multiple sound cards come to mind) only for it to break w…
> Maybe just maybe this is not the fault of Nvidia but due to the fact that large parts of the Linux ecosystem are fragile, time consuming to configure and break if you look at them in the wrong way. How it works with open source drivers is that you main-line your drivers so that the kernel maintainer maintain the drivers for you, for free. Choosing to keep your drivers closed source means committing to keeping your…
Graphics card drivers happen to be very complex beasts, somewhere along the stack they do need a compiler for multiple custom and often proprietary architectures, which most definitely no-one will maintain for free for you. There is a huge incentive to keep most of the code platform independent and only maintain a minimal kernel specific component. This part of the code base is a comparatively trivial part, essentially the kernel should get out of the way as much as possible. The kernel specific abstractions like KMS are examples of such comparatively trivial things.
Earlier quoted context omitted.
The 10g NICs are toys. ;-) Mellanox provides high speed (40/56/100/200 gigabit) switches and adapters. If you think 10g fiber is big, just wait... I’ve worked on several projects with them, and found they generally do a good job of feeding the beast when the OS and driver’s are properly tuned. I’m not informed enough to call good or bad, but will instead say it’s interesting, especially in the HPC space (and the emer…
Last I was looking you could get a 40GB card for around $300. Very impressive price that kind of speed in a home office.
Earlier quoted context omitted.
It is worth adding that the margin out of an app tends to be A LOT higher than shopping centers, airport, etc while the initial capital tends to be a fraction of it. There are also different ways in which you can use the data to help you create another app with high daily active players. An app can have the whole world as their oyster, unlike shopping centers, etc.
It's worth adding, too, that the airport and building take much larger capital expenditures to build and are far less reproducible as a result... and will probably have decades or centuries of future revenue. For a given game, that's probably a few years of peak popularity and comparatively easy to replicate by competitors. To the point of this article, it's kind of surprising to me at least that licensing companies…
In contrast, Mellanox produces a niche product. It is critical to HPC and some deep pocketed finance people, but it is never going to be at the volume of e.g. ethernet.
The long term aspect is that, as Moore's law scaling becomes harder and harder, distributed computation becomes more essential. Right now we are in a dip in peak compute rate requirements, because we have invented battery powered computers and 4G, and hence cloud. But the work done in the cloud isn't very taxing, mainly single threaded. But I'm confident we'll soon be seeing a lot heavier parallel computation in the cloud soon, and stuff that won't fit into one or four GPUs. The tradeoffs between PCIe and IB after 4*16x start to favour IB, especially if the IB silicon is on the NVLink switch complex. So from the perspective of integration bringing much greater IB volume, the acquisition multiples the worth of MLNX by an order of magnitude -- if nVidia can execute.
Earlier quoted context omitted.
IB has been doubling in bandwidth with every release, 10,20,50,100 and now 200G. Those other interconnect technologies did not do that.
Honestly, they competed on price & marketing at first. The handwriting was already on the wall for these companies long before IB had any real advantage on speed. We were all limited by PCIe bandwidth, after all. So no matter what the link speed is, you're not going to get more bandwidth than what PCIe can provide. I worked for Myricom when Mellanox was starting out with IB. I recall stories from that time of custome…
I don't think it was just the price or marketing (although Mellanox at least had sales people -- never heard from anyone at Myricom even after buying $$ of gear). Mellanox always seemed to be pushing the envelope, which might have been more R&D dollars. Never had problems with IB not working (CX-4 and up) but I guess the gremlins depend on scale.
As and outsider, can someone explain the "synergy" (god I hate that word) for the companies? It looks like Mellanox is primarily a network equipment company. What is the "fit" for that within a graphics/AI chip producer?
Earlier quoted context omitted.
It is, but if you're a large enough company to be buying millions of dollars in adapters and switches, reading a guide from Mellanox to turn on DCB should be fairly seamless. RoCEv2 is API-compatible with IB for the most part, so there is really no configuration on that layer. The other pieces -- not really sure what you're getting at. Most of those are for going across data centers, which IB won't do anyways. At lea…
> The other pieces -- not really sure what you're getting at. What I'm getting at is setting up clusters larger than what you can fit behind a single switch. So you'll want e.g. a CLOS fabric with multipathing (the typical IB setup, FWIW). As Trill and SPB seem pretty dead, it seems the momentum is to do the multipathing at the L3 level, using the aforementioned EPVN+VXLAN+BGP, or something similar.