TSMC Lands Orders for HPC Chips from NVIDIA: Volta in 12nm?

[Megalith]

Some folks think that a new chip order sent to TSMC is hinting that Volta will, in fact, be forged with 12nm technology, as rumored by earlier reports. VideoCardz has noted that these chips must be the Volta GV100, Xavier SoC, or both. The latter was originally said to be manufactured in the 16nm FinFET process.

Taiwan Semiconductor Manufacturing Company (TSMC) has landed orders for high performance computing (HPC) chips for AI applications from NVIDIA and Qualcomm, respectively, according to a Chinese-language Commercial Times report. TSMC will fabricate NVIDIA’s next generation Volta GPU using a 12nm process, said the paper, which added that the Volta GPU will be paired with NVIDIA’s Xavier supercomputer chips for self-driving car applications. TSMC also produces the 256-core NVIDIA Pascal GPU and dual-core Denver 2 CPU for NVIDIA’s recently released JetsonTX2 supercomputer platform, using a 16nm process, the paper indicated.

 

[Quartz-1]

So why aren't they using the 10nm process? What does this mean for Volta's release date?

 

[TAP]

So why aren't they using the 10nm process? What does this mean for Volta's release date?

Nvidia has been taking the conservative route with regard to node changes lately which you can't blame them. Nothing worst than having a design ready to go and having your product delayed 9-12 months due to delays in the new node change. And even then, going from 16nm to 12tnm is still a huge jump relatively speaking.

 

[N4CR]

According to earlier roadmaps, Paxwell was never supposed to exist. So I wonder if they've had issues producing Volta at a smaller node. It was for a while allegedly launching at 16nm, which makes no sense after a 1080Ti, so now we are not sure according to more BS. So really this is all just FUD?

 

[EdKiefer]

Probably more because HBM2 wasn't mature enough yet.

 

[N4CR]

Probably more because HBM2 wasn't mature enough yet.

But then the Nvidia apologists say it's a waste of time with GDDR10 million..

Can't have it both ways. Vega seems to be coming to market with HBM2 in the next few months. Same old I guess time will tell but I find it very unusual how this mixed up and panned out.

 

[DPI]

But then the Nvidia apologists say it's a waste of time with GDDR10 million..

Can't have it both ways. Vega seems to be coming to market with HBM2 in the next few months. Same old I guess time will tell but I find it very unusual how this mixed up and panned out.

I'm sure there's a point in there somewhere, but in the end, performance is all that matters. Whether nVidia uses steam pipe based memory or AMD uses quantum memory, if the performance isn't there nobody will give a shit which memory tech was utilized.

 

[Sycraft]

But then the Nvidia apologists say it's a waste of time with GDDR10 million..

Can't have it both ways. Vega seems to be coming to market with HBM2 in the next few months. Same old I guess time will tell but I find it very unusual how this mixed up and panned out.

Well for a desktop GPU, it may be. Memory bandwidth is one of those things where having enough is extremely important, but more doesn't really help. Same deal with amount of RAM. Basically you need enough bandwidth to keep the processor fed with data. So long as it can do that, more isn't going to help. You see that in HardOCP's memory speed tests all the time. While a synthetic subsystem benchmark shows linear improvements on faster RAM settings, actual applications show nothing because the RAM is already fast enough.

So it may be for desktop GPUs that the bandwidth offered by GDDR5X currently is enough, and that HBM2 while neat, isn't necessary. That can certainly change in the future as needs increase, but it may be for now that the impressive bandwidth numbers don't translate to a meaningful performance increase. Certainly we have evidence to suggest that given the R9's performance with HBM vs other chips with GDDR5.

 

[EdKiefer]

Well for a desktop GPU, it may be. Memory bandwidth is one of those things where having enough is extremely important, but more doesn't really help. Same deal with amount of RAM. Basically you need enough bandwidth to keep the processor fed with data. So long as it can do that, more isn't going to help. You see that in HardOCP's memory speed tests all the time. While a synthetic subsystem benchmark shows linear improvements on faster RAM settings, actual applications show nothing because the RAM is already fast enough.

So it may be for desktop GPUs that the bandwidth offered by GDDR5X currently is enough, and that HBM2 while neat, isn't necessary. That can certainly change in the future as needs increase, but it may be for now that the impressive bandwidth numbers don't translate to a meaningful performance increase. Certainly we have evidence to suggest that given the R9's performance with HBM vs other chips with GDDR5.

Right, and GDDR5x has enough BW for chips out now, at normal res.
But put much faster GPU with higher res and we may see a need.
It was GDDR5x was fast enough and had decent BW so they added nice amount, if went HBM2 price would go way up and having a lot more BW that GPU can't utilize is a waste. Its all a balancing game, we will see how it pans out.

 

[Sycraft]

Right, and GDDR5x has enough BW for chips out now, at normal res.
But put much faster GPU with higher res and we may see a need.
It was GDDR5x was fast enough and had decent BW so they added nice amount, if went HBM2 price would go way up and having a lot more BW that GPU can't utilize is a waste. Its all a balancing game, we will see how it pans out.

For sure not trying to say GDDR5 will be all we'll ever need. If nothing else, a new GDDR standard based on DDR4 is in development, though HBM may make it a non- issue. What I mean is that you can make a valid claim that HBM is not needed for current products. You can use it, but the extra bandwidth is not useful currently and thus holding off the products for its readiness/release is not useful.

 

[SighTurtle]

But then the Nvidia apologists say it's a waste of time with GDDR10 million..

Can't have it both ways. Vega seems to be coming to market with HBM2 in the next few months. Same old I guess time will tell but I find it very unusual how this mixed up and panned out.

People can be wrong. Its ok. Unless your Nvidia, no one else is gonna be able to 100% predict what is going to happen.

 

[EdKiefer]

For sure not trying to say GDDR5 will be all we'll ever need. If nothing else, a new GDDR standard based on DDR4 is in development, though HBM may make it a non- issue. What I mean is that you can make a valid claim that HBM is not needed for current products. You can use it, but the extra bandwidth is not useful currently and thus holding off the products for its readiness/release is not useful.

Right, I agree with you, when new faster GPU by large margin then you might need more memory BW.

Both HBM and GDDR will go up in BW.

 

[Presbytier]

But then the Nvidia apologists say it's a waste of time with GDDR10 million..

Can't have it both ways. Vega seems to be coming to market with HBM2 in the next few months. Same old I guess time will tell but I find it very unusual how this mixed up and panned out.

When you look at projected bandwidth of HBM memory on Vega compared to that meomry bandwidth on the Ti it does not seem to be that much faster. HBM is the future but not right now it was a smart cost saving idea to stick with faster GDDR5X