Our initial AMD Radeon VII liquid cooling mod was modified after the coverage went live. We ended up switching to a Thermaltake Floe 360 radiator (with different fans) due to uneven contact and manufacturing defects in the Alphacool GPX coldplate. Going with the Asetek cooler worked much better, dropping our thermals significantly and allowing increased overclocking and stock boosting headroom. The new drivers (19.2.3) also fixed most of the overclocking defects we originally found, making it possible to actually progress with this mod.
As an important foreword, note that overclocking with AMD’s drivers must be validated with performance at every step of the way. Configured frequencies are not the same as actual frequencies, so you might type “2030MHz” for core and get, for instance, 1950-2000MHz out. For this reason, and because frequency regularly misreports (e.g. “16000MHz”), it is critical that any overclock be validated with performance. Without validation, some “overclocks” can actually be bringing performance below stock while appearing to be boosted in frequency. This is very important for overclocking Radeon VII properly.
We recently revisited the AMD R9 290X from October of 2013, and now it’s time to look back at the GTX 780 Ti from November of 2013. The 780 Ti shipped for $700 MSRP and landed as NVIDIA’s flagship against AMD’s freshly-launched flagship. It was a different era: Memory capacity was limited to 3GB on the 780 Ti, memory frequency was a blazing 7Gbps, and core clock was 875MHz stock or 928MHz boost, using the old Boost 2.0 algorithm that kept a fixed clock in gaming. Overclocking was also more extensible, giving us a bigger upward punch than modern NVIDIA overclocking might permit. Our overclocks on the 780 Ti reference (with fan set to 93%) allowed it to exceed expected performance of the average partner model board, so we have a fairly full range of performance on the 780 Ti.
NVIDIA’s architecture has undergone significant changes since Kepler and the 780 Ti, one of which has been a change in CUDA core efficiency. When NVIDIA moved from Kepler to Maxwell, there was nearly a 40% efficiency gain when CUDA cores are processing input. A 1:1 Maxwell versus Kepler comparison, were such a thing possible, would position Maxwell as superior in efficiency and performance-per-watt, if not just outright performance. It is no surprise then that the 780 Ti’s 2880 CUDA cores, although high even by today’s standards (an RTX 2060 has 1920, but outperforms the 780 Ti), will underperform when compared to modern architectures. This is amplified by significant memory changes, capacity being the most notable, where the GTX 780 Ti’s standard configuration was limited to 3GB and ~7Gbps GDDR5.
Today, we’re reviewing the GTX 1660 Ti, whose name is going to trip us up for the entirety of its existence. The GTX 1660 Ti is NVIDIA’s mid-step between Pascal and Turing, keeping most of the Turing architectural changes to the SMs and memory subsystem, but dropping the official RTX support and RT cores in favor of a lower price. The EVGA GTX 1660 Ti XC that we’re reviewing today should have a list price of $280, sticking it between the $350 baseline of the RTX 2060 and the rough $200 price-point of modern 1060s, although sometimes that’s higher. For further reference, Vega 56 should now sell closer to $280, with the RX 590 still around the $260 range.
Apex Legends is one of the most-watched games right now and is among the top Battle Royale genre of games. Running on the Titanfall engine and with some revamped Titanfall assets, the game is a fast-paced FPS with relatively high poly count models and long view distances. For this reason, we’re benchmarking a series of GPUs to find the “best” video card for Apex Legends at each price category.
Our testing first included some discovery and research benchmarks, where we dug into various multiplayer zones and practice mode to try and find the most heavily loaded areas of the benchmark. We also unlocked FPS for this, so we aren’t going to bump against any 144FPS cap or limitation. This will help find which cards can play the game at max settings – or near-max, anyway.
Metro: Exodus is the next title to include NVIDIA RTX technology, leveraging Microsoft’s DXR. We already looked at the RTX implementation from a qualitative standpoint (in video), talking about the pros and cons of global illumination via RTX, and now we’re back to benchmark the performance from a quantitative standpoint.
The Metro series has long been used as a benchmarking standard. As always, with a built-in benchmark, one of the most important things to look at is the accuracy of that benchmark as it pertains to the “real” game. Being inconsistent with in-game performance doesn’t necessarily invalidate a benchmark’s usefulness, though, it’s just that the light in which that benchmark is viewed must be kept in mind. Without accuracy to in-game performance, the benchmark tools mostly become synthetic benchmarks: They’re good for relative performance measurements between cards, but not necessarily absolute performance. That’s completely fine, too, as that’s mostly what we look for in reviews. The only (really) important thing is that performance scaling is consistent between cards in both pre-built benchmarks and in-game benchmarks.
Finding something to actually leverage the increased memory bandwidth of Radeon VII is a challenge. Few games will genuinely use more memory than what’s found on an RTX 2080, let alone 16GB on the Radeon VII, and most VRAM capacity utilization reporting is wildly inaccurate as it only reports allocated memory and not necessarily used memory. To best benchmark the potential advantages of Radeon VII, which would primarily be relegated to memory bandwidth, we set up a targeted feature test to look at anti-aliasing and high-resolution benchmarks. Consider this an academic exercise on Radeon VII’s capabilities.
Our AMD Radeon VII review is one of our most in-depth in a while. The new $700 AMD flagship is a repurposed Instinct card, down-costed for gaming and some productivity tasks and positioned to battle the RTX 2080 head-to-head. In today’s benchmarks, we’ll look uniquely at Radeon VII cooler mounting pressure, graphite thermal pad versus paste performance, gaming benchmarks, overclocking, noise, power consumption, Luxmark OpenCL performance, and more.
We already took apart AMD’s Radeon VII card, remarking on its interesting Hitachi HM03 graphite thermal pad and vapor chamber. We also analyzed its VRM and PCB, showing impressive build quality from AMD. These are only part of the story, though – the more important aspect is the silicon, which we’re looking at today. At $700, Radeon VII is positioned against the RTX 2080 and now-discontinued GTX 1080 Ti (the two tested identically). Radeon VII has some interesting use cases in “content creation” (or Adobe Premiere, mostly) where GPU memory becomes a limiting factor. Due to time constraints following significant driver-related setbacks in testing, we will be revisiting the card with a heavier focus on these “content creator” tests. For now, we are focusing primarily on the following:
The AMD Radeon VII embargo for “unboxings” has lifted and, although we don’t participate in the marketing that is a content-filtered “unboxing,” a regular part of our box-opening process involves taking the product apart. For today, restrictions are placed on performance discussion and product review, but we are free to show the product and handle it physically. You’ll have to check back for the review, which should likely coincide with the release date of February 7.
This content is primarily video, as our tear-downs show the experience of taking the product apart (and discoveries as we go), but we’ll recap the main point of interest here. Text continues after the embedded video:
GPU manufacturer Visiontek is old enough to have accumulated a warehouse of unsold, refurbished cards. Once in a while, they’ll clear stock by selling them off in cheap mystery boxes. It’s been a long time since we last reported on these boxes, and GPU development has moved forward quite a bit, so we wanted to see what we could get for our money. PCIe cards were $10 for higher-end and $5 for lower, and AGP and PCI cards were both $5. On the off chance that Visiontek would recognize Steve’s name and send him better-than-average cards, we placed two identical orders, one in Steve’s name and one in mine (Patrick). Each order was for one better PCIe card, one worse, one PCI, and one AGP.
The AMD R9 290X, a 2013 release, was the once-flagship of the 200 series, later superseded by the 390X refresh, (sort of) the Fury X, and eventually the RX-series cards. The R9 290X typically ran with 4GB of memory, although the 390X made 8GB somewhat commonplace, and was a strong performer for early 1440p gaming and high-quality 1080p gaming. The goal posts have moved, of course, as time has mandated that games get more difficult to render, but the 290X is still a strong enough card to warrant a revisit in 2019.
The R9 290X still has some impressive traits today, and those influence results to a point of being clearly visible at certain resolutions. One of the most noteworthy features is its 64 count of ROPs, where the output is converted into a bitmapped image, and its 176 TMUs. The ROPs assist in improving performance scaling as resolution increases, something that also correlates with higher anti-aliasing values (same idea – sampling more times per pixel or drawing more pixels). For this reason, we’ll want to pay careful attention to performance scaling at 1080p, 1440p, and 4K versus some other device, like the RX 580. The RX 580 is a powerful card for its price-point, often managing comparable performance to the 290X while running half the ROPs and 144 TMUs, but the 290X can close the gap (mildly) at higher resolutions. This isn’t particularly useful to know, but is interesting, and illustrates how specific parts of the GPU can change the performance stack under different rendering conditions.
Today, we’re testing with a reference R9 290X that’s been run through both stock and overclocked, giving us a look at the bottom-end performance and average partner model or OC performance. This should cover most the spectrum of R9 290X cards.
We moderate comments on a ~24~48 hour cycle. There will be some delay after submitting a comment.