• MSI GTX 1050 Ti “OC” ($115 after MIR, ~$130-$140 normally)
• Gigabyte RX 570 Aorus 4GB (~$180)
• MSI RX 580 Gaming X 8GB ($240~$250)
• EVGA GTX 1060 SSC 6GB ($240~$250)
• EVGA GTX 1070 SC ($370-$400)
• EVGA GTX 1080 FTW1 (~$550)

Flooding the CPU with Work

This test’s only goal is to determine a threshold beyond which the laws of diminishing returns come into play: We seek the point at which investment ceases to return linear (or somewhat linear) scaling in performance. At some point, we’d expect, a line has to be drawn where the CPU simply cannot deliver enough frames to the GPU. If it’s juggling Dx11 or unoptimized implementations of newer APIs (or plain wrappers), then it’s possible that the CPU can’t push enough draw calls to the GPU for rendering. This results in an overloaded CPU. For this reason, games which are more geometrically complex will sooner encounter a CPU constraint in the pipe.

We’re not just testing games that are geometrically complex, though – we’ve got a mix of everything. The list includes:

• GTA V (all Very High/Ultra, except Post-FX on High; FXAA)
• Total War: Warhammer (1080p/High preset)
• Sniper Elite 4 (1080p/High preset, Dx12, Async)
• Battlefield 1 (1080p/Ultra preset with tuning to FOV)
• Ghost Recon: Wildlands (1080p/Very High preset)
• Ashes: Escalation (1080p/High, Dx12)
• Excluded: Watch Dogs 2 – we saw some interesting performance with the RX cards here and are excluding it until we’ve figured it out. Presently working with AMD & NVIDIA to further research.

We already know that devices like a GTX 1050, RX 560/550, or similar cards will not produce a GPU bottleneck, and so did not test them.

GPU Testing Methodology

Our other GPU tests are completely incomparable to these results due to a new set of numbers, completely different testing methodology, new game settings, and different test hardware (G4560).

Driver version 382.05 was used for nVidia, 17.5.1 for AMD.

A separate bench is used for game performance and for thermal performance.

Game Bench 

BIOS settings include C-states completely disabled with the CPU locked to 4.5GHz at 1.32 vCore. Memory is at XMP1.

For measurement tools, we’re using PresentMon for Dx12/Vulkan titles and FRAPS for Dx11 titles. OnPresent is the preferred output for us, which is then fed through our own script to calculate 1% low and 0.1% low metrics (defined here).

GTA V G4560 Bottleneck (1050 Ti vs. 1060, 570, 580, 1070)

Starting with GTA V, we immediately see our point of bottlenecking at the GTX 1080 and GTX 1070, where the two perform effectively identically – they are well within test-to-test variance, here. The GTX 1080 FTW and 1070 SC are both locked to around 100-101FPS AVG. The GTX 1060 SSC runs 96FPS AVG, making the GTX 1070 and 1080 4.9% faster – that’s definitely diminishing returns for buying anything more than a 1060. To further illustrate this, we see that the GTX 1050 Ti scales almost completely to the 1060, where the 1060 runs 52% faster than the GTX 1050 Ti. An RX 580 is also nearing the cutoff point of 100FPS, but doesn’t quite push hard enough to get there.

What we learn from this is that anything from a GTX 1060 and down – that’d include the RX 580, 570, 1050 Ti, 560, and 1050 devices – would not be sufficiently bottlenecked by the G4560 in this game. The cutoff point is 100FPS for these settings, at which point we experience diminishing returns.

But that’s just one game. Let’s look at Total War: Warhammer next.

Total Warhammer G4560 Bottleneck (1050 Ti vs. 1060, 570, 580, 1070)

Total War: Warhammer has a similar cutoff point, where we’re still seeing some scaling with the higher-end cards, but it’s to the tune of 1-2FPS once at GTX 1060 levels. Scaling from the 1060 to the 1080 is an improvement of just 4%, which is nothing close to what we see in our GPU benchmarks when the CPU limitation is removed. The 1060, 1070, and 1080 are all functionally equal in this game when paired with a G4560. The RX 580 is not too distant from this group with its 97 FPS.

As previously, you could purchase anything from a GTX 1060 downward and end up fine. The GTX 1060 seems to be about the cutoff point.

Sniper Elite 4 G4560 Bottleneck (1050 Ti vs. 1060, 570, 580, 1070)

Sniper Elite is the odd man out, where the game’s DirectX 12 native integration and asynchronous compute functions permit the CPU draw call workload to largely shift to the GPU. In this instance, we see scaling all the way up to a GTX 1080, surprisingly, where we’ve got a 170FPS AVG at 1080p with high settings. That puts the 1080 27% ahead of the GTX 1070, at 134FPS AVG, which is 38.5% ahead of the GTX 1060 SSC. The RX 580 Gaming X runs a 109FPS AVG, but none of these cards are at the limit of what the CPU can handle. We’d strongly recommend not basing purchases on this game, though, as this is a rare case in the gaming world right now. Granted, not many people would consider coupling a G4560 with a GTX 1080 to begin with.

Ashes Escalation G4560 Bottleneck (1050 Ti vs. 1060, 570, 580, 1070)

Let’s look at another DirectX 12 game next.

These results are for Ashes: Escalation’s GPU benchmark, which may as well be a synthetic test at this point. Escalation has the GTX 1080, 1070, 1060, RX 580, and RX 570 all performing mostly the same, with no difference visible until the GTX 1050 Ti. The cutoff point seems to be the 570 here. The 1070 and 1060 are within test-to-test variance, as indicated by the error bars.

Battlefield 1 G4560 Bottleneck (1050 Ti vs. 1060, 570, 580, 1070)

Battlefield 1 with DirectX 11 and 1080p/Ultra settings posts some scaling for the GTX 1080 from the 1070, though it is limited to 11% -- we normally see more than this when not CPU-limited. Although a 1080 posts improvement, it is well into the point of diminishing returns.

The 1070 holds an improvement of 18% over the GTX 1060 SSC, or 12.5% over the RX 580. Scaling seems to choke past a GTX 1070 in this game, although you’d get better value with a 580 or 1060.

Ghost Recon G4560 Bottleneck (1050 Ti vs. 1060, 570, 580, 1070)

Ghost Recon: Wildlands is another title where brute force gets us some extra framerate, but we start encountering limited gains once again. The GTX 1080 runs about 9.7% faster than the GTX 1070 in this game, which in turn runs 15% faster than a GTX 1060 SSC.  

We experienced a weird, reproducible issue with Ghost Recon: Wildlands with the RX 580. This isn’t something we saw when testing with the i7-7700K and only encountered with the G4560.

Here’s a look at the numbers in order of test execution. We ran two runs, each after system reboot. The first pass spikes, second pass drags, and then things begin to level out after that. With the higher end CPUs, we were seeing very steady, reliable performance. We’re not sure what’s causing this on the G4560, but we also observed interesting performance in Watch Dogs 2 with the same card. We are seeing completely repeatable (flat, unchanging) results on the nVidia devices with the G4560. We’re awaiting further thoughts from our industry contacts to hopefully better understand this.

Conclusion: When Does the Intel G4560 Bottleneck the GPU?

It all depends on the games, as usual, but we’ve also got to keep reality in mind: With the G4560, buying anything beyond a 1060 or 580 is sort of unrealistic. Going $300 on a GPU and $70 on a CPU, unless planning to swap the G4560 at a later date, is not going to be a common scenario.

Fortunately, that also happens to be around where the cutoff point is in several games. GTA V, AOTS, and Total War all show almost no scaling beyond GTX 1060 class hardware. Sniper Elite 4 is a rare, weird exception – unless this is the only game you’re ever going to play, don’t base a GPU purchase on it. Although we scale up to a GTX 1080, no other game shows performance like that. It is impressive, but that’s to do with how Sniper is built more than anything else. Sniper also exhibits near 1:1 multi-GPU scaling, another thing that few games can replicate. That’s not to minimize how well-built Sniper Elite 4 is, but just to remind why we use multiple games.

Other games show some scaling into the 1070/1080 class hardware, but returns diminish to 10% or less. At that point, it makes far more financial sense to buy a GTX 1060 or RX 580 maximally, then invest the extra $100 into a better CPU.

This was a fun feature benchmark to conduct. We’re back to normal reviews shortly.

Editor-in-Chief, Test Lead: Steve Burke
Tester: Patrick Lathan
Video Producer: Andrew Coleman