NVIDIA Ampere A100 GPU Specs, Ray Tracing, 200,000 DGX-3, & EGX GTC 2020 Keynote Recap
video description
Date: 2020-05-14
Related videos
Comments and reviews: 10
Geoff
I do AI and CNN medical imaging research. In terms of useful specs, FP64 FLOPS do not matter explicitly, as almost all computations are done using FP32. FP64 may only be relevant in relative performance when comparing only models and interpolating some FP32 numbers. FP16 and tensor-cores are useful, but right now on the software side, their actual implementations are still very much in development stage (ie, APEX package by NVIDIA themselves) so they are not as utilized as NVIDIA likes to think. For my personal work, GPU memory is crucial, but that isn't common for everyone.
reply
I do AI and CNN medical imaging research. In terms of useful specs, FP64 FLOPS do not matter explicitly, as almost all computations are done using FP32. FP64 may only be relevant in relative performance when comparing only models and interpolating some FP32 numbers. FP16 and tensor-cores are useful, but right now on the software side, their actual implementations are still very much in development stage (ie, APEX package by NVIDIA themselves) so they are not as utilized as NVIDIA likes to think. For my personal work, GPU memory is crucial, but that isn't common for everyone.
reply
WandererOfWorlds0
Real world AI applications are starting to become bottlenecked by I/O performance, not GPU performance, at least for models that are not gigantic in size. Nevertheless, if the real world performance is even close to what Nvidia claims, this is a huge milestone for ML engineering and deployment. Sadly, the closest most people will come to a DGX computer is a ssh link in a command line interface.
reply
Real world AI applications are starting to become bottlenecked by I/O performance, not GPU performance, at least for models that are not gigantic in size. Nevertheless, if the real world performance is even close to what Nvidia claims, this is a huge milestone for ML engineering and deployment. Sadly, the closest most people will come to a DGX computer is a ssh link in a command line interface.
reply
DRWNING
i have a 5700xt and im tired of the constant crashing and driver problems. Im just trying to figure out if its worth upgrading to a 2080 super and call it a day or will the new cards be around the same price and just wait. This wait for new hardware is so damn frustrating.
reply
i have a 5700xt and im tired of the constant crashing and driver problems. Im just trying to figure out if its worth upgrading to a 2080 super and call it a day or will the new cards be around the same price and just wait. This wait for new hardware is so damn frustrating.
reply
Cristian
Am I the only one that wants Nvidia to get into making VR STANDALONE headsets, using their GPUs and DLSS 2.0/3.0 (3.0 will be amazing), to be able to get high fps using low res input images? It is the only way to get high quality, high fps, realistic graphics right now.
reply
Am I the only one that wants Nvidia to get into making VR STANDALONE headsets, using their GPUs and DLSS 2.0/3.0 (3.0 will be amazing), to be able to get high fps using low res input images? It is the only way to get high quality, high fps, realistic graphics right now.
reply
SteelSkin667
To me it would make sense that Ampere would follow Volta as a professional part and Hopper would follow Turing into consumer hardware, just because of the people referenced in the code names, but there is indeed no indication that it's actually the case.
reply
To me it would make sense that Ampere would follow Volta as a professional part and Hopper would follow Turing into consumer hardware, just because of the people referenced in the code names, but there is indeed no indication that it's actually the case.
reply
SouthPlanObservation
It's all a fake scam. My polygraph gave flashing lights the moment Jensen was wearing no leather jacket in his kitchen taking the DGX out of the oven. No NASA supercomputer can render a jacket like that.
reply
It's all a fake scam. My polygraph gave flashing lights the moment Jensen was wearing no leather jacket in his kitchen taking the DGX out of the oven. No NASA supercomputer can render a jacket like that.
reply
Guest
I ve always thought of Nvidia as the bigger company for gaming, but between Xbox 360/one and PS4 I think more people probably game on AMD gpus. The Switch is probably making up ground there for Nvidia, but still....
reply
I ve always thought of Nvidia as the bigger company for gaming, but between Xbox 360/one and PS4 I think more people probably game on AMD gpus. The Switch is probably making up ground there for Nvidia, but still....
reply
Marco
In my opinion, the 3080 and 3070 will be released in June-July 2020. According to me, it makes no sense to leave a 5-6 month time frame between the presentation of the architecture and the release of the cards.
reply
In my opinion, the 3080 and 3070 will be released in June-July 2020. According to me, it makes no sense to leave a 5-6 month time frame between the presentation of the architecture and the release of the cards.
reply
DNAReader
Thanks for making quite a lot of explanation on the DL part. Still a bit disappointed we didn't get much of a core deep-dive like we usually do, but I guess HPC is a unique priority currently.
reply
Thanks for making quite a lot of explanation on the DL part. Still a bit disappointed we didn't get much of a core deep-dive like we usually do, but I guess HPC is a unique priority currently.
reply
salmiak
About the saltiness of a potential client at 12:46 I think the DGX-2 had 2 staks of 8 GPUs, 16 GPUs un total twice of the new one, so I think it's to be expected the new one cost half as much
reply
About the saltiness of a potential client at 12:46 I think the DGX-2 had 2 staks of 8 GPUs, 16 GPUs un total twice of the new one, so I think it's to be expected the new one cost half as much
reply
Add a review, comment
Other channel videos
