Nvidia Buys Mellanox [Updated]

[Lurick]

Rumors/reports have started surfacing that Nvidia might be in the market to buy Mellanox who are mostly known for their InfiniBand and Ethernet networking gear after earlier reports discussed rumors that Intel was looking to buy the company. It appears that Nvidia has since swooped in with an offer of $7 billion dollars for the company although exact terms are still unknown one person has said the deal will be done in cash. In my opinion this could be interesting to see how Nvidia integrates the mainly networking gear focused vendor into their existing data center portfolio but could also lead to some interesting solutions as well with high bandwidth GPUs coupled with low latency networking gear.

 

Quote

The deal would be Nvidia’s biggest-ever acquisition and boost its business of making chips for data centers, allowing it to reduce its reliance on the video game industry, for which it is best known as a major technology vendor.

 

Nvidia has outbid Intel Corp in the auction for Mellanox and could announce a deal as early as Monday, the person said. The source asked not to be identified because the negotiations are confidential. Intel and Mellanox did not immediately respond to requests for comment. Nvidia declined to comment. Financial news website Calcalist had reported earlier on Sunday that Nvidia had outbid Intel for Mellanox.

 

Link: https://www.reuters.com/article/us-mellanox-m-a-nvidia/nvidia-nears-deal-to-acquire-mellanox-technologies-source-idUSKBN1QR0QP

 

 

Edit:

It looks like the deal has been finalized with a final bid of $125/share or around $6.8 billion and the deal is expected to close by the end of 2019. There are some concerns as Mellanox has large customers in China and due to the recent US/China trade row.

 

New Link: https://www.reuters.com/article/us-mellanox-m-a-nvidia/nvidia-outbids-intel-to-buy-israels-mellanox-in-data-center-push-idUSKBN1QS197

[Fasauceome]

Networking stuff on GPUs?

 

[ki8aras]

6 minutes ago, Lurick said:

Rumors/reports have started surfacing that Nvidia might be in the market to buy Mellanox who are mostly known for their InfiniBand and Ethernet networking gear after earlier reports discussed rumors that Intel was looking to buy the company. It appears that Nvidia has since swooped in with an offer of $7 billion dollars for the company although exact terms are still unknown one person has said the deal will be done in cash. In my opinion this could be interesting to see how Nvidia integrates the mainly networking gear focused vendor into their existing data center portfolio but could also lead to some interesting solutions as well with high bandwidth GPUs coupled with low latency networking gear.

 

 

Link: https://www.reuters.com/article/us-mellanox-m-a-nvidia/nvidia-nears-deal-to-acquire-mellanox-technologies-source-idUSKBN1QR0QP

YAASSS NETWORKING ON GPUS

[NunoLava1998]

@advoli quick patent it

[leadeater]

2 hours ago, Lurick said:

In my opinion this could be interesting to see how Nvidia integrates the mainly networking gear focused vendor into their existing data center portfolio but could also lead to some interesting solutions as well with high bandwidth GPUs coupled with low latency networking gear.

Makes sense thinking about it, systems like the DGX-2 are rather complicated and use lots of expensive PCIe/NVLink switches and the system needs lots of power. I'm thinking Nvidia is looking to offboard the GPUs and run NVLink across IB type technology. I know PCIe rack switches also exist but having not used those I don't know how well they work.

 

Would be even more interesting if Nvidia can make the GPUs work independent of a server system and you can send job to them over network/fabric, that would be awesome. 

[mr moose]

39 minutes ago, leadeater said:

Would be even more interesting if Nvidia can make the GPUs work independent of a server system and you can send job to them over network/fabric, that was be awesome. 

Interestingly that was my second thought, my first thought was with ARM getting better in the power compute space that maybe Nvidia postulate some writing on the wall for Traditional GPGPU and simply want to diversify their product range.   

 

 

 

[Bananasplit_00]

i thought the title said Matrox first and kinda panicked tbh

[FPSwithaWacomTablet]

5 hours ago, leadeater said:

Would be even more interesting if Nvidia can make the GPUs work independent of a server system and you can send job to them over network/fabric, that would be awesome. 

Welcome to systems-on-chip, where the embedded space has been enjoying it for years. Considering that  Xilinx and Intel-Altera have been strapping ARM CPUs beside their FPGAs (Zynq and Cyclone SoCs come to mind), I don't see a conceptual issue strapping a Linux-capable CPU right beside a full nVidia GPU through some proprietary interconnect. Heck, nVidia's Tegra CPU might even play a role in such a possible development.

 

(I am in a uni course where we play around with such SoCs; check the Terasic DE1-SoC evaluation board for more info).

[leadeater]

19 minutes ago, FPSwithaWacomTablet said:

Welcome to systems-on-chip, where the embedded space has been enjoying it for years. Considering that  Xilinx and Intel-Altera have been strapping ARM CPUs beside their FPGAs (Zynq and Cyclone SoCs come to mind), I don't see a conceptual issue strapping a Linux-capable CPU right beside a full nVidia GPU through some proprietary interconnect. Heck, nVidia's Tegra CPU might even play a role in such a possible development.

 

(I am in a uni course where we play around with such SoCs; check the Terasic DE1-SoC evaluation board for more info).

Shouldn't actually need to pair the GPU with anything that high in the stack though implementing that much, there's a huge amount of work in the industry to get around that. Have a look at Gen-Z, CCIX etc or NVMe over Fabrics. Products implementing those could still fit under the term SoC but you don't need a CPU or to run up an OS.

 

The only problem I can see for now is just how tied GPUs are to the operating systems, driver and API wise. Getting them to the point where they are standalone addressable entities on a fabric is likely rather challenging, but then it might not be or it won't take too long to make it happen. Would love to see it in the next 3 years but I wouldn't bet money on that time frame.

[straight_stewie]

1 hour ago, leadeater said:

Would be even more interesting if Nvidia can make the GPUs work independent of a server system and you can send job to them over network/fabric, that would be awesome.  

 

I think that Nvidia has always been after something like that. If you look at some of their compute space marketing materials you'll see this idea of some blade servers with centrally located GPGPU units. With NVLink being embedded in some versions of PowerPC chips now, they may be looking to further sink their teeth into this space.

It may not even have anything to do with improving their GPU offerings, and just be about strengthening their position in the industry.

1 hour ago, mr moose said:

Interestingly that was my second thought, my first thought was with ARM getting better in the power compute space that maybe Nvidia postulate some writing on the wall for Traditional GPGPU and simply want to diversify their product range.    

I would tend to think that on-the-fly reconfigurable FPGA technology would be the threat they are worried about. AFAIK ARM chips aren't encroaching on the GPGPU space, they are encroaching on the traditional processor space.

[PlayStation 2]

I'm thinking that Nvidia would buy them for datacenter usage as well as potential cloud computing, too. I'd be a little surprised if some important shit came down to the consumer GPUs.

[NumLock21]

Nvidia did have their own networking chip during the nforce 3 days.

[mr moose]

1 hour ago, straight_stewie said:

 

 AFAIK ARM chips aren't encroaching on the GPGPU space, they are encroaching on the traditional processor space.

Maybe not now but I can see it in the future, and NVIDIA like most companies plan 30 years out.

 

I think had you said ARM would be encroaching on desktop processor space 10 years ago you would have been laughed off the forums for being ignorant,  but now we are looking at 1000 petaflop super computers being built around them.

 

https://www.datacenterdynamics.com/news/fujitsu-joins-linaro-to-develop-hpc-software-for-arm/

 

 

[Salv8 (sam)]

3 hours ago, leadeater said:

-snip-

actually this would be an excellent move for the server space, imagine a server with just GPU's, a simple arm CPU, and high speed low latency networking

render farms with this would be more efficient and faster since the workload is mostly GPU bound and don't require much CPU power and the data that needs to be processed can be sent directly to the GPU's VRAM and/or cache without requiring the CPU or any other device to get involved in the process, allowing fast transmission of data that needs to be processed quickly.

this would be a great move for nvida to acquire and use this tech to improve server render farm technologies

[leadeater]

2 minutes ago, Salv8 (sam) said:

a simple arm CPU

Why have the arm CPU at all . Once you have the fabric controller on the device and itself can communicate over the fabric there is no longer any need for a CPU.

[Salv8 (sam)]

4 minutes ago, leadeater said:

Why have the arm CPU at all . Once you have the fabric controller on the device and itself can communicate over the fabric there is no longer any need for a CPU.

i suggest arm since some server software already and OS's already exist for the architecture. due to it's low power and software support it could be an excellent option for CPU operations, RISC could also work as well although some existing software may need to be ported. from what i have read a fabric controller may not be applicable since all i can find on it is Microsoft's implementation of it in azure, it could work but it would again require custom software that doesn't exist as far as i know; that can interface with these custom devices. with my idea, existing solutions will work the same way, only the render node hardware would be changed and the software would be modified to work with the new hardware and architecture.

ya see where i'm getting at?

[leadeater]

Just now, Salv8 (sam) said:

i suggest arm since some server software already and OS's already exist for the architecture. due to it's low power and software support it could be an excellent option for CPU operations, RISC could also work as well although some existing software may need to be ported. from what i have read a fabric controller may not be applicable since all i can find on it is Microsoft's implementation of it in azure, it could work but it would again require custom software that doesn't exist as far as i know; that can interface with these custom devices. with my idea, existing solutions will work the same way, only the render node hardware would be changed and the software would be modified to work with the new hardware and architecture.

ya see where i'm getting at?

Well I'm more meaning Nvidia can make the GPUs addressable resources for a server. By that I mean you can have a rack chassis of GPUs attached to an IB/NVLink fabric that servers also connect to and you can send tasks directly to those GPUs, each piece can take care of themselves and doesn't require any extra handling. That's the basis of Gen-Z, every device has a Gen-Z controller and they can all talk Gen-Z and pass information between each other but nothing is reliant on each other to function.

[Salv8 (sam)]

1 minute ago, leadeater said:

Well I'm more meaning Nvidia can make the GPUs addressable resources for a server. By that I mean you can have a rack chassis of GPUs attached to an IB/NVLink fabric that servers also connect to and you can send tasks directly to those GPUs, each piece can take care of themselves and doesn't require any extra handling. That's the basis of Gen-Z, every device has a Gen-Z controller and they can all talk Gen-Z and pass information between each other but nothing is reliant on each other to function.

ah i get ya now...

anyways both of our solutions work to solve the same problem, how the industry proceeds is up to them.

i wanna see how this goes.

[leadeater]

1 minute ago, Salv8 (sam) said:

ah i get ya now...

anyways both of our solutions work to solve the same problem, how the industry proceeds is up to them.

i wanna see how this goes.

Well both Gen-Z and CCIX are being worked on and there is actual working examples of both. I think CCIX is further along and it's 'rumored' to be in EPYC2.

[Salv8 (sam)]

55 minutes ago, leadeater said:

Well both Gen-Z and CCIX are being worked on and there is actual working examples of both. I think CCIX is further along and it's 'rumored' to be in EPYC2.

cool

[SpaceGhostC2C]

I'm calling it now: "download more GPU", meme of the 2020s  

[Mesterial]

One day it's rumor from Intel, today Nvidia...

I guess that's good for Mellanox share rate !  

 

While NVlink over IB sounds nice, they should focus on porting codes for supporting NVlink first. I don't know a single scientific software using it (even HPL does not !). 

[williamcll]

We could see a NVS 2000? 

[TVwazhere]

The time change is really fucking with me, I thought the title was "Netflix to possibly buy mellanox"

 

I am too tired  

[Tribalinius]

We should also not forget they partnered with Aquantia for autonomous car networking last year. They probably have fingers in multiple pies at the moment.

 

https://hothardware.com/news/nvidia-selects-aquantia-for-multi-gig-connectivity-on-drive-px-autonomous-vehicle-platforms