AMD plans to deliver a 1.5+ exaflops supercomputer in 2021

[WYP]

Supercomputing with next-generation AI and HPC performance.



Read more about AMD's plans to deliver a 1.5+ exaflops supercomputer in 2021.

 

[NeverBackDown]

Well when you got 128 threads per CPU I guess it's much easier to do than a few years ago when it was less than 80 threads

 

[demonking]

Massive kick in the teeth there for Intel considering that is where summit is located. Glad these guys are getting it, the sort of science these guys study is close to my heart and very current rather than theoretical or purely computational.

 

[WYP]

Massive kick in the teeth there for Intel considering that is where summit is located. Glad these guys are getting it, the sort of science these guys study is close to my heart and very current rather than theoretical or purely computational.

To be honest, I think this is more of a kick in the teeth for Nvidia. They were doing very well in the supercomputer market and now the US has plans to build two exascale systems with Intel and AMD. The scale of these systems will make a mockery of today's Nvidia-powered supercomputers.

TBH I am not 100% sure on the details, but it makes sense that two separate architectures are being used to bring exascale computing to the US. This way there they have a little more redundancy if one manufacturer fails to meet its target and it splits the money between two US companies.

Nvidia, lacking enterprise-grade CPUs, are not in a great position to continue being successful in the supercomputing space, as both Intel and AMD can integrate their CPUs and graphics products together in a way that Nvidia cannot.

 

[NeverBackDown]

Well, you also have to take into account not everything relies on the GPU. So Nvidia are only a viable option in certain tasks that certain supercomputers need. You still need heavy CPU power, either way, so is this the result of not needing the super expensive Nvidia products as they mostly just need CPU power, or is this the result of Nvidia just charging so much they outbid themselves of the run for the GPU manufacturer?

 

[grec]

I think it's probably more to do with the speed & bandwidth of the interconnects and the level of homogeneity between the CPU & GPU/their physical memory pools, AI can be a super bandwidth heavy task, obviously particularly when your data inputs are along the lines of videos, and Intel and AMD both have aggressive projects with custom or progressive technology in these areas. Of course, so do NVidia (NVLink) but the only CPUs this works with are IBM PowerPC I think.

 

[WYP]

I think it's probably more to do with the speed & bandwidth of the interconnects and the level of homogeneity between the CPU & GPU/their physical memory pools, AI can be a super bandwidth heavy task, obviously particularly when your data inputs are along the lines of videos, and Intel and AMD both have aggressive projects with custom or progressive technology in these areas. Of course, so do NVidia (NVLink) but the only CPUs this works with are IBM PowerPC I think.

Agreed, it is Intel/AMD's CPU/GPU interconnectivity that makes them uniquely suited to these kinds of applications. Yes, Nvidia has to that to an extent with NVLink, but they still lack the CPU IP and integration with said CPU IP that AMD and Intel can offer.

It will be interesting to see what comes of these supercomputers, but ATM it seems like Intel and AMD are benefitting from being in both the CPU and GPU markets. It will be interesting to see how this all turns out.