6 ms·

This comment is off topic, but while I'm listening to the earnings call, I don't hear about specifically official PyTorch and Tensorflow support for AMD graphic

This comment is off topic, but while I'm listening to the earnings call, I don't hear about specifically official PyTorch and Tensorflow support for AMD graphic cards. All the questions and answers are generic with buzzwords like AI, doubling down on our software support, but it doesn't give me confidence to change my NVIDIA GPU to an AMD one for the foreseeable future. I remember the time when Elon Musk said to an analyst that he's asking boring questions to fill in his spreadsheet, and I'm feeling the same thing while listening to the earnings call.

ckastner 6y ago
My impression is that the accelerated computing side of AMD is receiving far, far too little attention. For example, their flagship GPU is still not officially supported by ROCm (AMD's answer to CUDA) [1]. Imagine the 2080 Ti not being supported by CUDA. I've become a huge AMD fan, both because of their hardware, and because of their commitment to open source. But while the battles they have one against Intel on the x86 side are impressive, it seems that CUDA is leaving them far behind. [1] https://github.com/RadeonOpenCompute/ROCm/issues/887 https://github.com/RadeonOpenCompute/ROCm/issues/887
- Tuna-Fish 6y ago
  It likely will never have support. AMD chose to bifurcate their GPU designs into compute (CDNA) and games (RDNA) lines, with different architectures. RDNA sheds all the fancy features needed to support modern compute, thus gets more efficient in games that do not use it, but also cannot support the modern compute APIs.
  - xiphias2 6y ago
    NVIDIA is adding more features, like super-scaling to games, and machine learning models are improving faster than Moore's law. I expect those fancy features, like tensor cores to be a must for 4K gaming in the future. What's funny is that the same strategy (leaving out specialized instructions from consumer level hardware) that worked extremely well for CPUs won't work for GPUs in my opinion. If you look at ray tracing hardware (I have it on my RTX 2070 Max-Q card in my laptop), it sucks right now, but it's improving very fast as machine learning algorithms improve. I just found this: https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-know https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-... One thing that I forgot is that AMD can just focus on inferencing hardware (INT16 operations), and leave out tensor cores...so actually you are right, I'll just stay with NVIDIA GPUs.
- xiphias2 6y ago
  ,,AMD ROCm is validated for GPU compute hardware such as AMD Radeon Instinct GPUs. Other AMD cards may work, however, they are not officially supported at this time.'' It seems like Lisa Su thinks that there's a separate ,,gamer market'' and ,,accelerator market''. Jensen Huang understands that the same person can like to play games and train machine learning models on the same machine. I'd love to switch to AMD CPU to have a portable laptop with low resource usage, as I spend most of my time travelling, but as GPUs in the cloud are overpriced (thanks to Jensen with separated pricing for servers), and internet in hotels are unpredictable, I don't want to train models in the cloud. Anyways, Lisa said that she reads all comments about AMD, so I hope she'll listen :)
ksec 6y ago
Yes I posted another link [1] for earning call but didn't get much traction / upvote. Although I have to admit Acquiring Xilinx is much bigger news. Judging from watching the Financial News and reading analyst's comment for years. My feeling was that their Job was not to push for hard question or an honest answer. Their job is to push whatever interest they had with the company. So a spin for better long term prospect and downplay risk. I was happy with the Enterprise results ( +116% YoY ) until I read this >Revenue was higher year-over-year and quarter-over-quarter due to higher semi-custom product sales and increased EPYC processor sales. Semi-Custom is definitely PS5 and Xbox. Basically I still dont see EPYC making enough inroad in the Server Market. And this is worrying, while the Stocks, Reviews, Hype are all going to AMD. No Results so far have shown Intel is hurt or AMD is making big gains in market shares and revenue shares. The only good part I guess is Ryzen Mobile contribution to Computing and Graphics segment. [1] https://news.ycombinator.com/item?id=24906314 https://news.ycombinator.com/item?id=24906314
- xiphias2 6y ago
  I can imagine that switching from Intel to AMD takes a bit of time for servers, as supporting 2 architectures at the same time is usually bad news for 99%-ile latencies for web services. At the same time x86 is a mature instruction set (as long as you don't use 512 bit vectorization, where things are getting tricky), so the transition shouldn't be that hard.
- hajile 6y ago
  The server market moves and replaces slowly. Even when Intel was beating AMD by 30% or more it still took years for AMD percentages to drop. AMD EPYC on AWS is 10.42% cheaper than Intel per hour across the board for m5 instances (9.6% cheaper for t3 instances). 7nm EPYC saves more than 10% power vs Intel 14nm and per-chip savings from buying AMD are way more than 10% vs similar Intel offerings. Why can Amazon spike prices that much? Because people will pay and still consider it a deal. AMD's main issue still seems to be available due to do much competition for 7nm and tsmc being reluctant to build new fabs.
  - Teever 6y ago

ckastner 6y ago

My impression is that the accelerated computing side of AMD is receiving far, far too little attention. For example, their flagship GPU is still not officially supported by ROCm (AMD's answer to CUDA) [1]. Imagine the 2080 Ti not being supported by CUDA. I've become a huge AMD fan, both because of their hardware, and because of their commitment to open source. But while the battles they have one against Intel on the x86 side are impressive, it seems that CUDA is leaving them far behind. [1] https://github.com/RadeonOpenCompute/ROCm/issues/887 https://github.com/RadeonOpenCompute/ROCm/issues/887

Tuna-Fish 6y ago
It likely will never have support. AMD chose to bifurcate their GPU designs into compute (CDNA) and games (RDNA) lines, with different architectures. RDNA sheds all the fancy features needed to support modern compute, thus gets more efficient in games that do not use it, but also cannot support the modern compute APIs.
- xiphias2 6y ago
  NVIDIA is adding more features, like super-scaling to games, and machine learning models are improving faster than Moore's law. I expect those fancy features, like tensor cores to be a must for 4K gaming in the future. What's funny is that the same strategy (leaving out specialized instructions from consumer level hardware) that worked extremely well for CPUs won't work for GPUs in my opinion. If you look at ray tracing hardware (I have it on my RTX 2070 Max-Q card in my laptop), it sucks right now, but it's improving very fast as machine learning algorithms improve. I just found this: https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-know https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-... One thing that I forgot is that AMD can just focus on inferencing hardware (INT16 operations), and leave out tensor cores...so actually you are right, I'll just stay with NVIDIA GPUs.
xiphias2 6y ago
,,AMD ROCm is validated for GPU compute hardware such as AMD Radeon Instinct GPUs. Other AMD cards may work, however, they are not officially supported at this time.'' It seems like Lisa Su thinks that there's a separate ,,gamer market'' and ,,accelerator market''. Jensen Huang understands that the same person can like to play games and train machine learning models on the same machine. I'd love to switch to AMD CPU to have a portable laptop with low resource usage, as I spend most of my time travelling, but as GPUs in the cloud are overpriced (thanks to Jensen with separated pricing for servers), and internet in hotels are unpredictable, I don't want to train models in the cloud. Anyways, Lisa said that she reads all comments about AMD, so I hope she'll listen :)

ksec 6y ago

Yes I posted another link [1] for earning call but didn't get much traction / upvote. Although I have to admit Acquiring Xilinx is much bigger news. Judging from watching the Financial News and reading analyst's comment for years. My feeling was that their Job was not to push for hard question or an honest answer. Their job is to push whatever interest they had with the company. So a spin for better long term prospect and downplay risk. I was happy with the Enterprise results ( +116% YoY ) until I read this >Revenue was higher year-over-year and quarter-over-quarter due to higher semi-custom product sales and increased EPYC processor sales. Semi-Custom is definitely PS5 and Xbox. Basically I still dont see EPYC making enough inroad in the Server Market. And this is worrying, while the Stocks, Reviews, Hype are all going to AMD. No Results so far have shown Intel is hurt or AMD is making big gains in market shares and revenue shares. The only good part I guess is Ryzen Mobile contribution to Computing and Graphics segment. [1] https://news.ycombinator.com/item?id=24906314 https://news.ycombinator.com/item?id=24906314

xiphias2 6y ago
I can imagine that switching from Intel to AMD takes a bit of time for servers, as supporting 2 architectures at the same time is usually bad news for 99%-ile latencies for web services. At the same time x86 is a mature instruction set (as long as you don't use 512 bit vectorization, where things are getting tricky), so the transition shouldn't be that hard.
hajile 6y ago
The server market moves and replaces slowly. Even when Intel was beating AMD by 30% or more it still took years for AMD percentages to drop. AMD EPYC on AWS is 10.42% cheaper than Intel per hour across the board for m5 instances (9.6% cheaper for t3 instances). 7nm EPYC saves more than 10% power vs Intel 14nm and per-chip savings from buying AMD are way more than 10% vs similar Intel offerings. Why can Amazon spike prices that much? Because people will pay and still consider it a deal. AMD's main issue still seems to be available due to do much competition for 7nm and tsmc being reluctant to build new fabs.
- Teever 6y ago