NVIDIA's GeForce GTX 1660 Ti Breaks Cover, Shaking Up Mainstream Gaming

👤by Tim Harmer Comments 📅22.02.2019 14:02:12



After six months NVIDIA's GeForce RTX 20-series is now complete, with GPUs ranging from the £320 RTX 2060, up to the £900+ GeForce RTX 2080 Ti. At the same time, the GTX 10-series is now almost fully consigned to the history books, production having ended and the remaining stock likely depleted in the near future. That means a mainstream replacement is needed, and it just so happens that today marks its introduction.

Slotting in just below the GeForce RTX 2060 is the GeForce GTX 1660 Ti, a broadbase replacement for the GTX 1060 6GB at the $279 excl. tax/£259.99 inc V.A.T. price point. It marks a distinct separation in the GeForce brand: GeForce RTX for performance-class GPUs with NVIDIA RTX features, and GeForce GTX for more affordable mainstream GPUs that don't support those specialist features.

NVIDIA have come in for plenty of flak in the past for confusing naming schemes, so it's absolutely a breath of fresh air to see such a clear distinction made between RTX capable and incapable classes of GPUs. It would have been easy for NVIDIA to call this the GTX 2060 6GB and potentially trigger to a whole lot of confusion, and they deserve credit for not doing so.


TU116 Block Diagram


The GTX 1660 Ti has a brand new GPU at its heart, rather than a repurposed TU106. The TU116 is a smaller, more streamlined chip with neither Tensor cores nor RT cores, and so has a substantially different Shader Module. Nor it is a rebadged Pascal GPU (and shame on you for assuming it might be); the design still incorporates Turing generation architectural improvements that unlock new features that haven't had the limelight, at least compared to their RTX-specific brethren.

One aspect that still holds up is the memory allocation. The 6GB on the GTX 1060 proved to be ample for 1080p gaming, and this frame buffer size is repeated in the GTX 1660 Ti. However the latest generation GDDR6 memory is used, increasing available bandwidth significantly and allowing it to keep up with the pace of games released in 2019.

Unlike RTX 20-series releases, the GTX 1660 Ti will not be available as a Founder's Edition with reference PCB and cooler design. Partners will be releasing their own variants, complete with their own cooler options and appropriate factory overclocks.

As the name implies, the GTX 1660 Ti is something of a mid-point generationally between the GTX 1060 6GB and RTX 2060. As with prior mainstream releases NVIDIA are aiming to satisfy consumers who currently own similarly positioned cards that are a couple of generations (3+ years) old. In this instance that means the GTX 960, a workhorse card in its time but now showing its age even at 1080p.

So, 1080p gaming in the latest titles with some Turing next-generation rendering features. There's plenty to cover, so let's get cracking.

Technical Specifications

Manufacturing Node:- TSMC 12nm FFN
Shader Modules:- 24
CUDA Cores:- 1560
Base Clock:- 1500 MHz
Boost Clock:- 1770 MHz

Memory:- 6GB GDDR6
Memory Speed:- 12Gbps (effective)
Memory Bus Width:- 192-bit
Memory Bandwidth:- 288 GB/s

FLOPS:- 11 TOPS (5.5 TFLOPS FP32 / 5.5 TOPS INT32)
FP16 FLOPS:- 11 TFLOPS
Texture Units:- 96
L1 Cache (total):- 1536 KB

Transistors:- 6.6 bn
Die Size: 284mm^2
TDP: 120W



Comparative Specs.


New, but still familiar, Shader Module

One of the core factors involved in Turing's development was the minimum necessary performance level needed for RTX features such as Ray-tracing and Deep Learning Inferencing to actually improve the user experience. They discovered that the level was pretty high, and that the RTX 2060 was realistically the lowest possible without compromising both rasterisation and hybrid rendering features too far. More affordable markets would therefore need either a TU106 with huge portions of the chip fused off (which may still be expensive), or a new GPU tailored to the market. In the TU116, they chose the latter.

At first glance the TU116 is starkly different to the TU106. It's more than 30% smaller with 6.6 bn compared to 10.8 bn transistors, and doesn't incorporate Tensor or RT cores into the Shader Module design. It also has just 1560 shaders compared to 1920 on the RTX 2060, so pure rasterisation performance too should be lower than its sibling. That said, its raw specifications indicate a major step up from the GTX 1060 6GB, which featured just 1280 Pascal generation shaders.



Changes to the Shader Module itself are even more substantial, particularly when comparing the cache system between Turing and Pascal SMs. Turing is much more heavyweight thanks to 4x the L1 Cache Bandwidth and 3x the L1 Cache Size, increasing the speed at which data is loaded and the chance for an L1 cache 'hit' (and reducing number of slower L2 Cache accesses necessary). Overall FPS performance improvements are not linear, but they do stack up. In terms of shader performance specifically NVIDIA assess an increase in IPC of around 40%, compounding with other features depending on the game.

The GTX 1660 Ti has all the SMs on the TU116 enabled, leaving scope for lower performance SKUs with fewer SMs enabled. That, however, is a conversation for the future.

---

Turing-specific Features

The practicalities of lacking both Tensor and RT Cores is that the GTX 1660 Ti is capable of neither Real-time or Hybrid Raytracing, Deep Learning Super Sampling, DLAA, and any other technologies which leverage the raytracing and inferencing aspects of NVIDIA's RTX technology suite. The GTX 1660Ti does however incorporate other features exclusive to the Turing architecture. We've discussed those with immediate importance in the past, but it's worth going over them again.

Coherent Float and Int operations



The pascal architecture is capable of executing both Floating and Integer operations, but only sequentially by switching between the two modes. Turing shader processors by contrast can execute batches of mixed operations at the same time, greatly increasing total operations per second in mixed workloads.

While floating point operations have tended to dominate game design, modern titles are beginning to roll in integer operations where it's deemed appropriate as GPU hardware broadens its mixed workload capabilities. Shadow of the Tomb Raider is a classic example, and this mixed approach on the Turing architecture represents an almost 1.5x perf improvement over Pascal in a representative scene.

Unified Cache

Turing moves to a Unified L1 Cache, a schema which offers lower latency, higher bandwidth, and higher capacity compared to Pascal and earlier architectures. Once again the impact on performance will differ according to the game, but with Call of Duty: Black Ops 4 NVIDIA highlight a 1.4x uplift in throughput.

Adaptive Shading



Another novel technique made possible by Turing is Adaptive Shading. The principle is simple. Sections of a scene will differ in complexity, from heavily textured objects and surfaces to more plain walls with less detail. From frame to frame large areas with less detail may change very little over time. Adaptive Shading is aware of that contextual difference, and can divert shading resources away from these zones dynamically to boost overall frame rates.

Adaptive Shading is one aspect of Variable Rate Shading possible under the RTX technologies umbrella. It's not clear which aspects won't be possible with the GTX 16-series, but that will no-doubt become clear as it is used in more games. For now, Adaptive Shading is present in Wolfenstein: The New Colossus, and is a factor in the NVIDIA's projection of a 1.4x performance advantage vs the 1060 6GB.

New NvEnc Encoding Engine

In common with the RTX 20-series, the GXT 1660Ti supports NVIDIA's new NvEnc Hardware Encoder. Designed specifically for game broadcasting via Twitch and other streaming platforms, it significantly improves the video stream (x264 fast) encoding bit rates to up to 6000KB/s. Video Quality is also improved, now exceeding the quality of a i9-9900K using the same encoding as well as offloading load from the CPU. Definitely a feature to look out for if you're a budding streamer.



OC Scanner

Although not strictly exclusive to Turing, OC Scanner was introduced alongside the RTX 20-series cards as a means of offering automated (one-click) overclocking for a 'free' performance boost. It works by scanning for artefacts and other inconsistencies as operating frequencies increase, easing off when detected and failing safely rather than causing hard lockups. OC Scanner is now incorporated into the major aftermarket overclocking tools from EVGA, MSI and ASUS, and remains a nice 'bonus' for the GTX 1660 Ti no matter your manufacturer of choice.

--

Power Efficiency

Currently Turing GPUs are manufactured using TSMC's 12nm lithography, which has many similarities with the GTX 1060's 16nm manufacturing process. As a result power efficiency improvements will tend to be down to architectural changes rather than the process node, but higher frequencies on similar voltages aren't out of the question.

NVIDIA boast of a 1.4x perf/watt improvement in optimal workloads over the GTX 1060 6GB, which is more than just a little significant. The GTX 1660 Ti has a reference TDP of 120W, the same as the GTX 1060 6GB, which speaks to the expected performance improvement you might see when comparing with the previous generation.



Target Market

As we've mentioned, NVIDIA are really pushing the GTX 1660Ti as an upgrade for those in the GTX 960-class of system performance. They expect these users to be most comfortable with 1080p resolutions at high quality settings, pushing upwards of 120fps in some popular and fast-paced titles. According to the Steam Hardware Survey over two thirds of PC gamers are equipped with a GTX 960 or lower GPU, representing a massive potential market. The picture is very complex at this price point however.

Currently you're able to purchase an AMD Radeon RX Vega56 for a touch under £250, and a small selection of GTX 1070 models have begun to flirt with a sub-£270 price point. These are the cards the GTX 1660 Ti will need to compete with, not the GTX 1060 6GB or Radeon RX 590. Straying from the MSRP and into the territory of £300+ will push it into the clutches of the GeForce RTX 2060, a card that should beat it all ends up.

With that in mind, premium GTX 1660 Ti's kitted out with cutting edge (and potentially over-specced) cooling solutions might be a tough sell unless pricing is kept under control. However, manufacturers who can adhere to NVIDIA's suggested pricing as well as offer good quality cooling (which ZOTAC and Palit succeeded at during the RTX 2060 launch) could be rewarded with significant sales.

As always more information is available at the official NVIDIA Geforce site, and you can check out our lauch day review of the ASUS ROG STRIX GTX 1660 Ti for a more complete picture of relative performance.




Related Stories

Recent Stories

« NVIDIA Releases GeForce MX250 and MX230 Mobile Graphics · NVIDIA's GeForce GTX 1660 Ti Breaks Cover, Shaking Up Mainstream Gaming · ZOTAC Outs GAMING GeForce GTX 1660 Ti Graphics Cards »