Only costs an arm and a leg. Glad AMD finally moved away from this...

Silva:

Only costs an arm and a leg. Glad AMD finally moved away from this...

They where forced to use HBM to be competitive, they needed the memory speed and lower power usage. HBM is still used on the top workstation cards, unless something bag happens with GDDR7, HBM will still be the best solution for high bandwidth and large capacity.

Silva:

Only costs an arm and a leg. Glad AMD finally moved away from this...

So, that's to cost exactly? Is it more than system memory having same bandwidth? Wait? your dual channel system reaches only 60GB/s and not 460GB/s? What are you paying for here exactly? What about mobile device having APU + memory in 3x4 cm area. Imagine 6C/12T + iGPU performing at around 75% of 5700 XT. Perfect 1080p gameplay at good power draw in thin laptop. Saved space from memory means extra cooling capacity, battery, another M.2, ...

But so far none of the HBM gaming gpu succeed.

sunnyp_343:

But so far none of the HBM gaming gpu succeed.

On other hand, none of them would be even released without it. And it is not that they did not succeed. It is that competition cheaped out for years on compute. GCN was unnecessarily big and power hungry due to compute power nVidia did not care for. And since it was thing just one manufacturer had, games did not used it much either. Therefore quite some portions of those GPUs were worthless to pure gaming crowd. Had it been that nVidia matched AMD's compute several generations before Turing, all of them would have been successful. People do not like Turing much, but it is on par with Vega in computational operations. Imagine your GPU costing $50 more while having 10% higher clock on GPU and 50% higher memory bandwidth at same TBP. Would not be bad on $600 MSRP card, right? It all depends on achieved benefits. Apparently, your GTX 1080 did not have need to save much energy on memories. Neither it needed to be smaller, or had need for 50% higher memory bandwidth. But it would be very unrestricted card with HBM2.

Would love to see 1/10th of this speed in non-volatile memory. Not that networks / internal bandwidth could support it, but it would be nice to look at. 🙂 On a more relevant note, any word on pricing in comparison to the GDDR alternatives? Given HBMs past, I'm sure it's prohibitively expensive, so it won't make it down to the GPUs us mortals use. This part of the post gives me that impression as well:

SK Hynix's HBM2E is an optimal memory solution for the fourth Industrial Era, supporting high-end GPU, supercomputers, machine learning, and artificial intelligence systems that require the maximum level of memory performance.

You can pretty much tell where they plan to make all of their money on this.

Fox2232:

What about mobile device having APU + memory in 3x4 cm area. Imagine 6C/12T + iGPU performing at around 75% of 5700 XT. Perfect 1080p gameplay at good power draw in thin laptop. Saved space from memory means extra cooling capacity, battery, another M.2, ...

I thought the old hades canyon NUC with the Intel and Vega processors on a single "chip" was one of the coolest bits of tech I've seen in quite a while. I am still kinda baffled as to why AMD hasn't tried to pursue this further especially with all their experience with interposers, chiplets, HBM, "Infinity Fabric" etc.

I thought AMD used those HBM chips in their process for memory cache

16gb on a single package will be highly sort after in 2020 trust me. If the price comes down i'd expect to see many top end GPUs using it.

TLD LARS:

They where forced to use HBM to be competitive, they needed the memory speed and lower power usage. HBM is still used on the top workstation cards, unless something bag happens with GDDR7, HBM will still be the best solution for high bandwidth and large capacity.

They were not competitive using that BS, it only sank the GPU side of the company. Plus, they didn't make any money on it.