Review: Nvidia GeForce RTX 2080 Ti and RTX 2080

Introduction

Unless you have been living under a large rock, it won't have escaped your attention that Nvidia has brought its next-generation graphics microarchitectuer, known by the codename Turing, to market with the GeForce RTX 2080 Ti and GeForce RTX 2080 cards. The more mainstream GeForce RTX 2070 is due to follow next month.

It can be easily argued that Turing is the biggest mind shift in Nvidia's GPU thinking for a long time; Turing borrows far more from workstation/server-focussed Volta than it does from Pascal. Not only that, Nvidia makes a bold play in the ray tracing stakes by dedicating specific RT cores for the purpose, while more than dabbling in the deep-learning field by integrating Tensor cores.

Appreciating the number of new technologies baked into RTX silicon requires reading of the architecture piece from last week. Summarising what you will find there, Nvidia packs in additional CUDA cores that are more efficient, improves caching, bumps up memory bandwidth, introduces deep-learning hardware in the form of those Tensor cores and, of course, that play for ray tracing.

Such observations automatically lend themselves to statements such as 'forward-looking design' and 'all-new thinking'. That's all well and good, but the gamer of today wants superlative performance now and in the future. Turing-based cards need to hit the ground running.

We say this because they are not cheap. Today's dynamic duo, RTX 2080 Ti and RTX 2080, ship at a minimum $999 and $699, respectively - you can attach a 1:1 dollar-to-pound exchange rate, as well - and the truth of the matter is that retailer price gouging in the first few weeks means you will probably pay at least 10 per cent more than what's quoted.

Early adopters can either go for an add-in card design from the likes of EVGA or Asus, or lay cash down on Founders Edition GPUs available from Nvidia directly. What's also new this time around is FE cards are overclocked on the core, putting Nvidia's partner cohort in a sticky situation. So without further ado, let's formally introduce the FE pair and see how they perform against a number of other high-profile cards from Nvidia and AMD.

Tale of the Tape

GeForce RTX 2080 Ti Founders Edition

The GeForce RTX 2080 Ti and RTX 2080 use the Turing TU102 and TU104 dies, respectively. Said dies are truly massive, even on a 12nm process, and it's almost scary to imagine the complexity of the 18.6-billion transistor, 754mm² TU102. What's clear by looking at the specs is that the additional RT and Tensor hardware, plus beefed-up caches, creates a monster of a GPU in every sense of the word.

The RTX 2080 Ti is not the full implementation of the TU102 die. Rather, with it having a couple of SM units switched off and a narrower 352-bit memory bus - TU102 is specced with 384 bits - we can say that, clock for clock, it ought to provide about 95 per cent of the possible performance. Nvidia, it seems, keeps the full-fat die for the Quadro RTX 6000 chip that costs a whole heap more.

Other than that, specifications are robust, and it either matches or 'out-tables' the incumbent mainstream champ, GeForce GTX 1080 Ti, in every spec that matters. Knowing the CUDA cores are now more efficient due to integer processing moving on to a different pipe, and with due consideration of the extra muscle, one would expect RTX 2080 Ti to be 35-50 per cent faster at 4K. Enough for a 50 per cent price hike, to a place where spending £1,000 buys you the cheapest RTX 2080 Ti around? The numbers will tell all.

Funnily enough, the Founders Edition is one of the most expensive, priced with a $200 premium over bone-stock models. Nvidia reckons those who want the best will be willing to shell out that bit more for the creme de la creme, and this is why it has, for the first time, overclocked the boost speed for an FE offering.

Of course, RTX 2080 Ti is really for that small sliver of the gaming market who has the desire and, more importantly, necessary funds for such a GPU. Anyone not comfortable with shelling out four figures for a graphics need not apply. Perhaps the next rung down of the RTX ladder is a better bet.

GeForce RTX 2080 Founders Edition

Switching to the TU104 die reduces the financial barrier to $699 for the cheapest cards and $799 for the FE. On paper, it looks vastly more powerful than the GTX 1080 and, knowing how the architecture works, probably faster than the GTX 1080 Ti - it ought to make up for its small GFLOPS and memory-bandwidth disadvantage by dint of its more efficient design. But knowing the Turing and Pascal architectures are different means we have to look at each game's performance on a case-by-case basis.

It is good to see Nvidia keep the performance up by maintaining that 14Gbps GDDR6 memory, though this time run through a 256-bit bus. And it's also telling that this second-rung GPU, based on a lesser die, is still physically larger than the GTX 1080 Ti. Nvidia is putting a heck of a lot of store in non-CUDA performance.

This variant of the FE has healthier-than-stock frequencies, too. Our educated guess is that Nvidia wanted the RTX 2080 to be faster than GTX 1080 Ti in almost all of today's games and much quicker for workloads of tomorrow, ones that take advantage of DLSS processing and ray tracing. If that turns out to be true, Nvidia can genuinely say that its new architecture represents a clear leap over the old.

There is significant potential in both new cards, so let's now see what they're all about by taking a quick look at the FE cards and then running them through the performance wringer.