Nvidia announces new Tesla Accelerators

[patrickjp93]

@Patricjp93

i expected the professional maxwell to be 1/16th FP64.... why did they nerf it this badly? I know the architecture itself werent focused on FP64, but this seems a bit harsh, even for professional grade stuff.

Not harsh at all. That DCC tech and the high cache took up space. Nvidia cut out some stuff from Maxwell to make it fit for gamers since HPC was taking a breather. Nvidia does what is best for it as a business, no two ways about it.

[Jirne]

Alright guys, let's set this straight. 

M40 and M4 are Gm200/206 based cards(new and cool and maxwell), targeted at AI/Machine Learning/Skynet/ that kind of stuff. For this application learning fast is more important than precise fp64 calculations, THAT is why low fp64 numbers are irrelevant, stop bitching.

 

The K40 is GK110 based (old kepler stuff) and is indeed target for compute and yes it has therefor higher fp64 numbers but yes, that AMD card mentioned earlier theoretically beats it... But it's kepler... So stop bitching about maxwell cuz it isn't maxwell!

[Notional]

Alright guys, let's set this straight. 

M40 and M4 are Gm200/206 based cards(new and cool and maxwell), targeted at AI/Machine Learning/Skynet/ that kind of stuff. For this application learning fast is more important than precise fp64 calculations, THAT is why low fp64 numbers are irrelevant, stop bitching.

 

The K40 is GK110 based (old kepler stuff) and is indeed target for compute and yes it has therefor higher fp64 numbers but yes, that AMD card mentioned earlier theoretically beats it... But it's kepler... So stop bitching about maxwell cuz it isn't maxwell!

 

Doesn't change the fact that single precision fp32 is still much worse than AMD equivalents. These cards better be cheap by professional standards. Also AI/Machine Learning/Skynet is still compute.

[patrickjp93]

Doesn't change the fact that single precision fp32 is still much worse than AMD equivalents. These cards better be cheap by professional standards. Also AI/Machine Learning/Skynet is still compute.

Only on paper, which doesn't matter. In real life AMD gets creamed on every compute metric that matters to the enterprise market.

[rhyseyness]

Only on paper, which doesn't matter. In real life AMD gets creamed on every compute metric that matters to the enterprise market.

Source?

[don_svetlio]

Source?

Asking Patrick for a source is pointless - he always repeats "I am my own source!"

Ask him to do math instead, I did that once - was not disappointed.

You see friend, the math in school is fundamentally wrong since it uses white paper!

On red paper 2+2=3

On green paper 2+2=5

You see? That's the secret to math! Use the correct type of paper! /s

[Notional]

Only on paper, which doesn't matter. In real life AMD gets creamed on every compute metric that matters to the enterprise market.

 

Even so, Maxwell GPU's are so weak in compute that most AMD high end cards in praxis will perform better. Do you honestly think these Tesla's are good? Really?

 

 

Asking Patrick for a source is pointless - he always repeats "I am my own source!"

Ask him to do math instead, I did that once - was not disappointed.

You see friend, the math in school is fundamentally wrong since it uses white paper!

On red paper 2+2=3

On green paper 2+2=5

You see? That's the secret to math! Use the correct type of paper! /s

 

On blue dragon paper 2+2 = over 9000 :lol:

[don_svetlio]

Even so, Maxwell GPU's are so weak in compute that most AMD high end cards in praxis will perform better. Do you honestly think these Tesla's are good? Really?

 

 

 

On blue dragon paper 2+2 = over 9000 :lol:

He forgot blue - I assume he sees in R and G only /s

[don_svetlio]

Exactly what I wanted to say to you. I had to search up what kek means since i don't play WoW or LoL

I just had no other response than "Kek" to the SpongeBob GIF

[patrickjp93]

Source?

Linpack compute benchmarks, which are the gold standard. http://streamcomputing.eu/blog/2014-04-08/accelerators-hpc-2014/

[patrickjp93]

He forgot blue - I assume he sees in R and G only /s

I'm R/G colorblind actually. Also, eat it. Linpack benches are the gold standard compute benchmarks. AMD can't even compete with Xeon Phi. http://streamcomputing.eu/blog/2014-04-08/accelerators-hpc-2014/

You lot are a particular breed of jerks. You must be teenagers.

[don_svetlio]

I'm R/G colorblind actually. Also, eat it. Linpack benches are the gold standard compute benchmarks. AMD can't even compete with Xeon Phi. http://streamcomputing.eu/blog/2014-04-08/accelerators-hpc-2014/

You're forgetting Luxmark where they reign

Also - F@H is better on Kepler and GCN

[Notional]

Linpack compute benchmarks, which are the gold standard. http://streamcomputing.eu/blog/2014-04-08/accelerators-hpc-2014/

 

But Maxwell still has worse compute performance than Kepler. What was your point again?

[patrickjp93]

You're forgetting Luxmark where they reign

Also - F@H is better on Kepler and GCN

A dual-290X beats a 780TI, what a shocker...

 

LuxMark is as AMD biased as SuperPI is Intel biased. It's not worth a thought.

[patrickjp93]

But Maxwell still has worse compute performance than Kepler. What was your point again?

No it doesn't after you take GPU Boost 2.0 into account. The 980 trounces the 780TI by about 15% in Linpack.

[don_svetlio]

A dual-290X beats a 780TI, what a shocker...

 

LuxMark is as AMD biased as SuperPI is Intel biased. It's not worth a thought.

580 beating a 980 - care to comment?

[Notional]

No it doesn't after you take GPU Boost 2.0 into account. The 980 trounces the 780TI by about 15% in Linpack.

 

I doubt Tesla cards overclock.

[don_svetlio]

No it doesn't after you take GPU Boost 2.0 into account. The 980 trounces the 780TI by about 15% in Linpack.

You DO realise 780 Ti can also OC, right? You also realise that the 780 Ti is getting crappier drivers now, right?

[patrickjp93]

I doubt Tesla cards overclock.

They do. They have the same thermal/electrical headroom constraints as their consumer counterparts. They're just rated for full-tilt usage at 100*C 24/7 since they're enterprise parts.

 

You DO realise 780 Ti can also OC, right? You also realise that the 780 Ti is getting crappier drivers now, right?

No one's proven the 780TI being gimped. Is the performance degrading with new drivers? I doubt it.

[patrickjp93]

580 beating a 980 - care to comment?

Nvidia not caring about a benchmark 90% of people also don't care about? Thermal throttling? Fluke run? There are a million better explanations than yours.

[Thony]

A type of artificial intelligence that provides computers with the ability to learn without being explicitly programmed.

So, Skynet.