AMD RDNA2 will arrive with HBM2E memory in its highest range

[Inkriql]

The Radeon VII has not met expectations, to the point that it has already entered EOL and will soon withdraw from the shelves. In contrast, Navi has produced the opposite effect thanks to its RDNA architecture, but currently does not have high-end GPU. To alleviate this, AMD would introduce with RDNA2 the new HMB2E memories in its top GPUs in order to compete with NVIDIA Ampere.

Last week we already advanced the new AMD roadmap for its processors and graphics cards, where times had somewhat shortened and everything seems to go as planned by the company.

This implies that RDNA has only been the first platoon to attack NVIDIA in regards to graphics cards, where as we know they have arrived with the last lithographic process of TSMC, its 7 nm.

RDNA2 instead, will come with the new Taiwanese lithographic process (7nm +) which will not be a simple restructuring to improve efficiency, but will introduce more improvements and, above all, more density per mm2 (+ -20%). This, together with the improvements of the architecture itself (for now unknown) should be a more than notable leap in performance.

But AMD will not stay there, as we know today it plans to introduce another improvement, at least in its high-end graphics cards. And those of Lisa Su would go from implementing GDDR6 to a new and remodeled HBM2E, the latest version of this type of VRAM that includes two very notable improvements.

First and according to SK Hynix, this new type of memory will have the highest transmission bandwidth in the industry, where it will far exceed the current HBM2 even though it is only an improved version.

The company has offered specific data in this regard: a pin-to-pin transmission rate of 3.6 Gbps with a 1024-bit bus, which would give a bandwidth of 460 GB / s
HBM2E almost doubles the performance of the original HBM2

The truth is that the performance with HBM2E has greatly increased, where, for example, if we compare what a Radeon VII can do with HBM2 and what it could do with HBM2E, the difference is overwhelming.

In the first case, as we saw at the time, the graphics card was able to achieve 1 TB / s of bandwidth through a 4096 bit bus, but with the new HBM2E on the same bus the bandwidth would be increased up to 1.8 TB / s. The improvements do not end here, since the capacity has also been increased.

Thanks to the new TSV techniques, HBM2E will be able to stack up to 8 16 GB chips, which following the previous simulation of the Radeon VII, would not give a total capacity of 64 GB distributed in 4 vertical memory stacks.

The problem, as AMD has been dragging in recent times, will be the cost.

HBM memories in any of its versions have always offered great efficiency and great performance, where their only problem is the cost they have, where AMD has always played with narrower profit margins due to this.

The question, in the event that RDNA2 effectively includes HBM2E is simple, will it influence the price of the high-end AMD and with this we will see more expensive GPUs in the style of NVIDIA with Turing? Or will AMD return to the strategy of minimum benefits but maximum price competition?