AMD Radeon Instinct accelerators for deep learning

[NumLock21]

AMD has announced their Radeon Instincts accelerators for deep learning, which consist of 3 models, the MI6, MI8, and MI25. The MI6 has a compute capability of up to 5.70TFLOPS, with 224GB/s of memory bandwidth and a TDP of less than 150w. The MI8 has 8.2TFLOPS of computing power, 512GB/s of memory bandwidth and a power consumption of less than 175w. MI25 on the other hand, AMD did not provide any numbers for this card, except for a few notes like 2x Pack Math and High Bandwidth Cache and Controller, with a TDP of <300w.

Out of these 3 accelerators,  the MI6 is based on the current Polaris, MI8 on the older Fiji, and MI25 on the new upcoming VEGA.

Now for those who are excited to get one just to see how fast Crysis will run, don't bother because

Quote

 those cards are not technically graphics cards, but accelerators. In fact they don’t even have any display outputs. In other words, Radeon Instinct is AMD’s response to NVIDIA TESLA, pure computing-oriented accelerator.

 

 

 

 

 

 

 

http://www.anandtech.com/show/10905/amd-announces-radeon-instinct-deep-learning-2017

http://videocardz.com/64677/amd-announces-first-vega-accelerator-radeon-instinct-mi25-for-deep-learning

 

 

[Energycore]

1 minute ago, NumLock21 said:

Will fix and update , on mobile

Please do this piqued my interest

[AresKrieger]

Well unless their pricing is good the only one with potential is the mi25 as the others fall below pascals Tflop ratings and Nvidia has given them the same deep learning compatibility as the titan xp.

[Paytend01]

Just curious, what are these cards for? I dont need a long answer just something to know what they are.

[Space Reptile]

 

 

a nice breakdown for those who havent seen it 

[NumLock21]

1 minute ago, Paytend01 said:

Just curious, what are these cards for? I dont need a long answer just something to know what they are.

Deep learning

[Carclis]

Source if anyone is interested.

[Paytend01]

Just now, NumLock21 said:

Deep learning

Wow... lol, but like what do the cards themselves do? 

[LukaH]

this is cool at least something about vega directly from amd. Still no indication how desktop vega will be but its something...

[Carclis]

So 50% faster than Titan XP for deep learning sounds impressive. Hopefully they do this well for gaming performance too.

[NumLock21]

12 minutes ago, Carclis said:

snip

I will give source when im done with this topic.

[LukaH]

1 minute ago, Paytend01 said:

Wow... lol, but like what do the cards themselves do? 

in the amd video they say stuff like:

we produce over 9000 (it was a big number) bytes a day and we can't examine it by our self fast enough.... that is where these cards come in with AI/deeplearning scanning the data converting into charts and data sheets we can understand easily.

[AresKrieger]

7 minutes ago, Space Reptile said:

 

 

a nice breakdown for those who havent seen it 

Wonder if amd have compatibility with with more ai oriented instruction sets like the one Nvidia was using for pascal, as that provided Nvidia an effective 4x boost for Tflop count in regards to deep learning

[Guest]

Basically Quadro equivalent amirite

[zMeul]

11 minutes ago, Paytend01 said:

Wow... lol, but like what do the cards themselves do? 

compute ... and that's it

they are similar to nVidia's Tesla accelerators or Intel's Xeon Phi accelerators

and most of the times they do not have video outputs

[zMeul]

2 minutes ago, deXxterlab97 said:

Basically Quadro equivalent amirite

no, Quadros are professional graphics cards; these are compute accelerators (see my post above)

the AMD equivalent for the Quadro is the FirePRO series

[Citadelen]

So would I be right is assuming these are the successor to the long time discontinued FireStream GPUs?

[Energycore]

28 minutes ago, Paytend01 said:

Wow... lol, but like what do the cards themselves do? 

Crunch billions of parallel floating point operations every second?

 

The most usual purpose of that power is grahpics shading, but these are specialized for a different purpose.

[NumLock21]

1 hour ago, Space Reptile said:

 

 

a nice breakdown for those who havent seen it 

permission to post this in OP?

[Notional]

1 minute ago, NumLock21 said:

permission to post this in OP?

You don't need permission 

[TheRandomness]

So they're basically an RX 480, a Fury Nano and a Vega 10/11 card... 490/590?

[Mr.Meerkat]

4 minutes ago, TheRandomness said:

So they're basically an RX 480, a Fury Nano and a Vega 10 chip... 490/590?

Literally was bout to post that  

[BuckGup]

So basically it has the same flops as a Titan XP with HBM2. Good

[TheRandomness]

47 minutes ago, LukaH said:

this is cool at least something about vega directly from amd. Still no indication how desktop vega will be but its something...

The highest end Vega 10/11 chip has most likely been used here, and a TDP of <300W indicates maybe a dual GPU solution.. Two 150W Vega 10/11 chips with 'High Bandwidth cache' (probably HBM2) and '2x packed maths' leads to that assumption. So basically a potential Vega 11 chip (unless Vega 10 is really that efficient) with HBM2, and Vega 11 is the codename for the weaker of the two GPUs... I'm probably completely wrong on this

[zMeul]

4 minutes ago, SamStrecker said:

So basically it has the same flops as a Titan XP with HBM2. Good

P100 you mean?