AMD Navi might be faster than the RTX 2080

Until AMD launches Navi on the market, it will probably take until the second half of 2019. The 7nm GPU could even position itself above the RTX 2080.

Nvidia and the high-end monopoly

Although the graphics card market is not at a complete state of stagnation at the moment, many customers are still very disappointed. While AMD hasn’ t been able to come up with a new graphics card since 2017, Nvidia has taken the chance and introduced the new graphics card generation called Turing in the run-up to Gamescom 2018. In the high-end sector, the manufacturer is now virtually a monopolist. This has also led to the fact that Nvidia can demand very high prices for the new graphics cards. For example, the RTX 2080 Ti cost at least 1,249 Euro when introduced, while the RTX 2080 cost at least 849 Euro. Although the prices are slowly falling a bit, the lack of competition is still noticeable.

As a result, AMD is currently forced to take action on the one hand, but is also unable to deliver an opponent on the other. The Vega architecture was rather a flop for gamers, as it did not approach the high-end model GTX 1080 Ti from Nvidia and only ran at the level of the GTX 1080. Delivery problems and the mining boom also caused big problems for AMD. However, the next step is the shrink from the 14nm process to 7nm – and that takes time. The Vega architecture will be the first, and will be launched this year as the Vega 20 with a 7nm structure width. However, AMD is only launching the graphics card as a Radeon Instinct for server customers – the gamers are left empty-handed. Navi, Vega’s successor, will not be launched until the second half of 2019. If you look at the production process, Navi could also become an opponent for the high-end division of Nvidia.

Production process: smaller is better

If we now look at Nvidia’s current high-end segment, one thing is particularly remarkable: the chips are very large. The chip of the RTX 2080 Ti measures 754 mm² and accommodates 18.6 billion transistors, while the predecessor GTX 1080 Ti has a 471 mm² chip with 11.8 billion transistors. The transistor density of the RTX 2080 Ti is thus approximately 24.7 million transistors per square millimeter, while the predecessor GTX 1080 Ti even has a transistor density of 25.1 million transistors per square millimeter. The same applies to the other Turing models:

This shows that the shrink from 16nm to 12nm has achieved relatively nothing. However, since 12nm is only a slightly improved 16nm process, this was to be expected. Nvidia compensates the almost constant transistor density with significantly larger dies and sometimes more clockspeed. How many of these transistors are owed to the tensor and RT cores is difficult to estimate. Nevertheless, Nvidia shows quite well that the Turing generation mainly relies on the larger chips, since the chip of the RTX 2070 is almost as big as the chip of the GTX 1080 Ti.

A good example for the optimization through the shrink to 7nm is offered by Apple. The smartphone manufacturer exclusively ordered the chips A10 Fusion, A11 Bionic and A12 Bionic from TSMC. For the A10 Fusion, a 16nm structure width was used, which accommodates 3.3 billion transistors or 26.4 million transistors per square millimeter with a chip size of 125 mm². The A11 Bionic has already been manufactured with a 10nm structure width. There are 4.3 billion transistors on a chip size of 87.66 mm², which means that there are around 49 million transistors per square millimeter. The current A12 Bionic with 7nm structure width has the highest transistor density due to its smallest structure width. On 83.27 mm² 6.9 billion transistors are integrated, which corresponds to about 83 million transistors per square millimeter. The switch from 16nm to 7nm thus resulted in a transistor density per square millimeter that was more than three times as high.

7nm might kick-start Navi

For the 7nm products, AMD also relies on the services of the foundry TSMC instead Globalfoundries. Their optimized 7nm process could also rely on EUV lithography for Navi, which is probably not yet used for Vega 20. This makes even finer structures and a higher number of transistors possible. If we now assume that AMD makes good use of the 7nm process, the same number of transistors could be achieved with a third of the surface area of the Turing graphics cards. With a 250 mm² die 20.7 billion transistors would be possible. A 200 mm² die would offer space for 16.6 billion transistors. This would give AMD a chip that even with a size of 200 mm² could be positioned above the RTX 2080 and just below the RTX 2080 Ti. It would be very unlikely that AMD would not take advantage of this potential.

Of course, the sheer number of transistors does not make a powerful graphics card. The RX Vega 64 has 12.5 billion transistors, which is more than the GTX 1080 ti with 11.8 billion. At that time, the structure widths were still close, as Vega was manufactured with 14nm and Pascal with 16nm. With the switch to 7nm production, AMD now has the real opportunity to overtake Nvidia in transistor density.

AMD’s possible product strategy

It would therefore be logical to position a high-end model with the biggest Navi chip between the RTX 2080 and RTX 2080 Ti. A slightly trimmed chip might still fit between the RTX 2070 and RTX 2080. Depending on how cheap the production is, the small Navi variant could then already have the price of a mid-range card. This would also be possible by using comparatively inexpensive GDDR6 instead of expensive HBM2 memory. AMD will probably continue to serve the area below with the recently leaked RX 590. An expansion upwards with significantly larger chips and more transistors is of course also possible. For the server market, a Navi GPU with HBM2 or even HBM3 can be launched.

The whole potential of the 7nm production can be seen for the first time with the release of Vega 20, which is also manufactured in 7nm. On the other hand, we still have to wait until the second half of 2019 for Navi. But then AMD could finally be competitive against Turing again.