AMD Introduces Radeon Instinct Accelerators, One Based On Vega 10

[cangelini]

Eager to get in on the machine intelligence era with a graphics architecture known for compute prowess, AMD is introducing three deep learning accelerators to its new Radeon Instinct family. It's also beefing up its software stack with MIOpen.

AMD Introduces Radeon Instinct Accelerators, One Based On Vega 10 : Read more

 

[IceMyth]

Wow, I think AMD now is in the right track to beat NVidia and stand on the top (AMD fanboy comment). Deep learning and neural networks are the kind of things companies should strive for either NVidia, Intel, or AMD. What I would love to see a comparison of what the comming AMD GPUs will introduce to the public in term of Gaming, Design and all fields but as the article said still need sometime.

 

[Guest]

AMD's CEO Lisa Su seems to really know what she's doing. It should be a model example of why sometimes you need an engineer CEO instead of a sales type.

 

[captaincharisma]

AMD has turned into a company that's all talk but as always the reults will be lackluster. but hey, it makes the shareholders happy

 

[jaber2]

Does this mean my next CPU will be a GPU?

 

[falchard]

But AMD has been in this market for years, how are they late when they are several years early? They were even competing with nVidia on the contracts they got recently.

 

[JakeWearingKhakis]

JABER2 - No, its just that the way GPU's are made to calculate they just happen to be very good at neural networks, and things like bitcoin mining. I believe they will be (and already are) helping the CPU when it needs the help and is not being used for graphics processing at the time the CPU needs it.

CAPTAINCHARISMA - They are not all talk. They just went the wrong way with CPUs for a while. Which is why hyperthreading is coming back with ZEN. They aren't all talk with GPUs either. What you've seen this year is their low-mid range line up. once VEGA and the 490X come, there is no doubt it will be wonderful tech.

Nvidia doesn't make CPUs (other than Tegra?) and AMD being the smaller company can keep up without stealing technology and keeping it open source.

Intel is a mega corporation, and AMD having less than I think 10k employees is on their ass. AMD gets way less credit than it deserves.

 

[falchard]

It's amazing how much changed in the last couple months with the RX 480 and GTX 1060. Before you could say the 1060 will perform better in DX11 and worse in Vulkan/DX12. Now it's the RX 480 will perform the same in DX11 and better in Vulkan/DX12. I guess this means the actual competition for the RX 480 8GB was the GTX 1070.

 

[alidan]

Wow, I think AMD now is in the right track to beat NVidia and stand on the top (AMD fanboy comment). Deep learning and neural networks are the kind of things companies should strive for either NVidia, Intel, or AMD. What I would love to see a comparison of what the comming AMD GPUs will introduce to the public in term of Gaming, Design and all fields but as the article said still need sometime.

tomorrow should be the stream for zen, given they also showed off vega gpu getting 70-100fps at 4k ultra in doom, may see vega there too.

It's amazing how much changed in the last couple months with the RX 480 and GTX 1060. Before you could say the 1060 will perform better in DX11 and worse in Vulkan/DX12. Now it's the RX 480 will perform the same in DX11 and better in Vulkan/DX12. I guess this means the actual competition for the RX 480 8GB was the GTX 1070.

nah, the 1070 to far ahead hardware wise, there is only so much ground to make up in drivers, but the 480 being between a 1060 and 1070 is about right, to beat a 1070 they would need to refine the process and make more of the silicons higher end, some of them can get damn near 1500mhz clock rate some of them can barely go above base, if the process gets improved and they can put out more higher end parts, it could compete with a 1070, but likely wont happen unless they rebrand for navi.

 

[omikronsc]

"AMD said the MI6 and MI8 are ideal for GPU-accelerated inference, which is what a neural network does after it has gone through its more intensive training stage."

So MI6 and MI8 is a crap. Amdahls law points that the best speed up you get from speeding up the LONGEST operation in a process. Image shows that training takes hours (days in fact) and inference takes seconds (milliseconds in fact) and AMD is best in speeding up inference part.

I'm wondering if this is what AMD really presented or author of this article messed it up.

 

[problematiq]

"AMD said the MI6 and MI8 are ideal for GPU-accelerated inference, which is what a neural network does after it has gone through its more intensive training stage."

So MI6 and MI8 is a crap. Amdahls law points that the best speed up you get from speeding up the LONGEST operation in a process. Image shows that training takes hours (days in fact) and inference takes seconds (milliseconds in fact) and AMD is best in speeding up inference part.

I'm wondering if this is what AMD really presented or author of this article messed it up.

The problem with benching the training portion is that it's heavily dependent on what you are training. It also depends on your training style, my brother and I trained that same thing (face recognition) on two boxes with identical hardware yet it took his box a whole day an a half longer to train than mine. (his performed the task better in the end though.)