Volta: NVIDIA's Next Generation GPU Architecture (2017-2018)

I'm expecting to see Volta around summer of 2018.

 

Agreed, I expect Pascal to be milked to 2018 at least!!

Sent from my LG-H850 using Tapatalk

 

Considering Pascal didn't exist until after Maxwell was announced, when they updated the road map because HBM wasn't ready, it's my belief that there's not much of a reason to stay with Pascal 16nm when you can launch Volta 16nm, and milk Volta for three generations . Out with the old, in with the new.

AMD is about to release Zen and their HBM2 GPU lineup Q1 2017: Vega. NVIDIA will likely follow and compete with Volta.

 

Ah, and I'm expecting to buy a laptop from you a that time. Need that time to save haha

 

We'll be here when you're ready!

 

Firstly...

In essence, the graphics processor now becomes an SoC instead of a PCB, where everything is laid out wide on a PCB, with the VRAM chips separate from the GPU, and connections between the two. This has several benefits:

1. Smaller physical footprint. Video cards have rivalled their motherboards in size in recent years. I remember having an 8400 GS that was slightly smaller than two credit cards lengthwise. Now, GPUs are behemoths with massive coolers. With HBM/HBM2, something the size of the card above, or even smaller (possibly with the area of a UDIMM or even SO-DIMM) can house processing power well in excess of the current Titan XP.
2. Waaaay faster memory bandwidth, on the order of a few TB/s.

That would mean RAM upgrades would be tied to CPU upgrades, as the system is now an SoC. Then we'd be locked down even more. As far as desktops are concerned - this won't happen for a long, long time, possibly a couple of decades at least.

 

Actually, current plans for stacked ram/3d ram/hbm2 on CPUs are to use it for the iGPU with HMA so that when not used by the igpu, it becomes a fourth cache level while maintaining memory on the board for when it is exhausted. So, buying a processor with 16-32GB HBM2 would still allow for however much ddr4 to also be used...

Sent from my SM-G900P using Tapatalk

 

Second point is false. First point is kinda true...

1. Reduce size is true but thermal will still be an issue. Expect similar size as R9 Fury X for HBM2 cards. You still need spaces to lay out your VRMs/cooler/voltage controllers/etc.

2. Hardly way faster.... 3072bit 12gb HBM2 on PCIE P100 12GB gets you around 540gb/s performance, within reach of good overclocked 10ghz gddr5x on a 384bit. 4096bit 16GB HBM2 on 16GB P100 would give you 720GB/s where a 14ghz GDDR5x on a 384bit bus would give you 672gb/s. Ofcourse HBM2 could be clocked faster in future giving it more edge.

The power consumption reduction/efficiency and bandwidth are likely to be key points. With bandwidth probably not being a keypoint for consumers right now.

The con of HBM2 right now is likely to be horrendous yield. NVIDIA have 2 different PCIE P100 configuration with one of them having only 3/4 stacks of HBM2 enabled.

Do you mean HSA?

 

Yep, sorry. Was typing fast and have been off my game lately. HSA. I got heterogeneous memory architecture in my mind somehow.

Sent from my SM-G900P using Tapatalk

 

The thing is, I think the earliest we will see 3d stacked memory on CPU is Zen+ and thats even a very optimistic guess.

One thing would be interesting... watching Nvidia and Intel fight out the HPC market.