@flazza Thank you, for the link. The comparison charts to Pascal Tesla (P100) are especially interesting.
Speculation that Nvidia Volta will still use GDDR:
Motley Fool: Why NVIDIA Corporation’s Volta Gaming Products Won’t Use HBM2
Reading the above, I wonder if the High Bandwidth Cache Controller requires HBM or will it also work with GDDR? The article above mentions the use of HBM2 increasing memory bandwidth and efficiency, but I suspect that the memory latency is where the real difference between HBM2 and GDDR6 will be.
- Nvidia Pascal works with GDDR and HBM.
- Nvidia Volta will also work with both, if this article is to be believed.
- AMD Vega and Navi will work only ell with HBM, assuming both:
- their architectures are inseparable from the HBCC
- the HBCC requires HBM due to its lower latency
So assuming that games eventually do optimise for HBM and/or the HBCC AMD is stuck selling cards where a large part of the cost (HBM) will only be useful to the consumer in the future.
Nvidia seems to have no plans for consumer card HBM and is willing to sell GDDR cards for the time being, making higher profit due to GDDR’s lower cost to manufacture.
So AMD seems to be betting on a future where HBM2 is game-engine-relevant, while Nvidia is not, but has the option to switch to HBM2 if that future comes about. Maybe Nvidia will suffer in this case without a HBCC, but they might have something comparable now or by then.
However, the worst case scenario would be that the pricing bait and switch on Vega56 and Vega64 drives away enough buyers that game developers never program for HBCC or HBM.
^ If Vega/Navi cannot use GDDR
Also, can anyone make sense of these memory bandwidths:
- Radeon Instinct - MI125 (HBM2 - 484 GB/s) src
- Tesla - P100 (HBM2 - 732 GB/s) src
- Titan Xp (GDDR5x - 547.7 GB/s) src
Why does HBM2 (High Bandwidth Memory 2) have less bandwidth than GDDR5x on the Titan Xp?