NVIDIA Beats AMD to Market On HBM2 - Announces Tesla P100

[Kaapstad]

In gaming that's not surprising as dp compute is not much of a factor from what I understand. Fp32 is important while gaming not fp64. HBM reduces power consumption true or false? NV used it for a reason whether that's extra bandwidth, lower power draw or a smaller board. Maybe all 3 combined is why.

Yes it is not surprising and it will also work the other way round too for fp64.

 

[Yaayuh!]

Beats AMD to market, a market that most of us have no interest or experience in.

 

[TheRealDeal]

Yes it is not surprising and it will also work the other way round too for fp64.

I think the point is when the GP100 loads up bother sets of shaders which must happen on different work loads then it will use more power. While gaming Titan Black and the 780ti would just be using Fp32 cores so power draw should be similar. I really don't even know if that's how it works but it makes sense as GP100 draws more power than gp102 while getting the benefits of lower power HBM.

 

[Kaapstad]

I think the point is when the GP100 loads up bother sets of shaders which must happen on different work loads then it will use more power. While gaming Titan Black and the 780ti would just be using Fp32 cores so power draw should be similar. I really don't even know if that's how it works but it makes sense as GP100 draws more power than gp102 while getting the benefits of lower power HBM.

My old Titans did not draw a lot more power when they did compute benches.

 

[TheRealDeal]

My old Titans did not draw a lot more power when they did compute benches.

It depends on the compute work you were doing. Some use fp32 and not fp64 hence why Titan Xp is still considered a very good compute card. I think there is also fp16. Titan Kepler would have murdered the 780ti in anything that used fp64 as the 780ti had it's fp64 capability cut down. This is why the original Titan/Black are considered real Compute cards in comparison to the Maxwell/Pascal cards which are just gaming Gpu's mainly.

 

[Kaapstad]

Titan Kepler would have murdered the 780ti in anything that used fp64 as the 780ti had it's fp64 capability cut out. This is why the original Titan/Black are considered real Compute cards in comparison to the Maxwell/Pascal cards which are just gaming Gpu's mainly.

My old ones are Titan Kepler and I have still got them too.

 

[Mauller]

HBM also provides more memory channels while taking up far less space.

HBM provides 8 native memory channels per die, HBM 2 can separate each into 2 pseudo channels that share the same clocking and refresh, but act as independent 64bit memory channels.

the benefit is lower latency. The fury has 32 memory channels due to how hbm works. P100 using hbm2 should have 64 memory channels.

also p100 derived parts are not consumer orientated. So while they may have released a part with HBM2, AMD will still be first to market with a consumer part that utilises HBM2.

Also people comparing original titan and 780ti power use, the cores are likely software disabled due to how ingrained they are in the power supply with the 32bit cores. Idle ALUs still consume power so a titan and 780ti will consume around the same power under gaming workloads. If the 64bit cores could be fully disabled, the 780ti would have used less power.

 

[SiDeards73]

Does anyone here on these enthusiast gpu forums actually care that a professional card is using HBM2? i mean really? Is anyone here likely to buy one to put into their home PC for home use? realistically i think not... this article was a non issue here before it was ever posted

 

[D.P.]

Does anyone here on these enthusiast gpu forums actually care that a professional card is using HBM2? i mean really? Is anyone here likely to buy one to put into their home PC for home use? realistically i think not... this article was a non issue here before it was ever posted

To a certain extent I do. I am doing some deep NN at work, the P100 is intriguing but too expensive at this stage. HBm2 would make a difference in some of the stuff we do, processing big data funnily enough requires the transfer of a lot of data where memory bandwidth obviosuly helps. But then we alo simply using Amazon AWS GPU instances to good success.

 

[nashathedog]

Unigine Superposition benchmark is coming out very soon, from what I understand of it this bench does show the value of HBM and what it can do at higher resolutions.

That'll be interesting, I'll keep an eye out for that cheers.

 

[nashathedog]

It depends on the compute work you were doing. Some use fp32 and not fp64 hence why Titan Xp is still considered a very good compute card. I think there is also fp16. Titan Kepler would have murdered the 780ti in anything that used fp64 as the 780ti had it's fp64 capability cut down. This is why the original Titan/Black are considered real Compute cards in comparison to the Maxwell/Pascal cards which are just gaming Gpu's mainly.

Unless you ask Nvidia

So in fp64 workloads is a Titan Black faster than any of the Titan X's?

 

[TheRealDeal]

Unless you ask Nvidia

So in fp64 workloads is a Titan Black faster than any of the Titan X's?

http://images.anandtech.com/graphs/graph9059/72530.png

It's not just faster it's way faster. Anand had this to say.

On the other hand with a native FP64 rate of 1/32, the GTX Titan X flounders at double precision. There is no better example of just how much the GTX Titan X and the original GTX Titan differ in their FP64 capabilities than this graph; the GTX Titan X can’t beat the GTX 580, never mind the chart-topping original GTX Titan. FP64 users looking for an entry level FP64 card would be well advised to stick with the GTX Titan Black for now. The new Titan is not the prosumer compute card that was the old Titan.