Thursday, March 7, 2019

Thinkpad T440p/T540p Iris Pro 5200 GPU Comparison Part 1





Introduction

Figure 1: The 4980HQ (left) and 4900MQ (right) used in this comparison
Prior to the availability of Crystalwell CPUs adapted to Socket G3, the T440p and T540p have had two options for internal graphics processors, with either the Intel HD 4600 or the NVIDIA GT730M. The introduction of the Crystalwell CPUs has added a third option, the Intel Iris Pro 5200, that is theoretically the highest performing GPU (Table 1). 

Table 1: Abridged theoretical GPU performance comparison
However, most benchmarks for this GPU have not included the GT 730M, arguably the most relevant comparison in the context of the ThinkPad T440p and T540p. This is an important comparison because GT 730M motherboards are more widely available, often cheaper, and is fully functional in comparison to the Crystalwell CPUs which have shortcomings of their own, chief among them the loss of the miniDisplayPort output due to a design flaw in the PGA adapters used to adapt the BGA Crystallwell CPUs to Socket G3. Potential T440p/T540p buyers could stand to benefit with the knowledge of another viable GPU option if the GT 730M can be made to perform similarly to the Iris Pro 5200 in practice. 

This blog post series aims to quantify the performance of the three GPU options available to the T440p/T540p in gaming benchmarks. In this first part, we will test the three GPUs pushed as far as possible within the limits of the stock BIOS with no further modifications to serve as a baseline for subsequent blog posts, which will investigate GPU performance with additional hardware and firmware modifications.

Setup

The T440p was used to evaluate Intel HD 4600 and Iris Pro 5200 performance with either a 4900MQ (-60 mV undervolt) or 4980HQ CPU (-70 mV undervolt) at stock clockspeeds. This laptop was equipped with 16GB 1866mhz memory.
A T540p was used to evaluate NVIDIA GT 730M performance with a 4700MQ CPU (-70 mV undervolt) overclocked to 3.4Ghz to reduce the CPU clockspeed deficit compared to the other two 49XX CPUs. This laptop was equipped with 8GB 1600mhz memory.

All laptops used Windows 10 1809 with the latest available drivers at the time of testing with either NVIDIA 418.81 or Intel 5058 drivers and Noctua NT-H1 thermal paste for all CPU/GPUs. Throttlestop was also used to increase the TDP limits from the stock 47/37 long/short term power limits to 58/47, the maximum power limits that are supported by the stock BIOS for both systems. All benchmarks were conducted in a 21C air conditioned room.

Overclocking and Thermals

Overclocking was conducted on all three GPU options. 
The NVIDIA GT 730M stock boost clocks consist of 876mhz core and 1000mhz memory. The core was overclocked to 1006mhz by applying the maximum +135mhz offset. Further overclocking on the core is not possible without a VBIOS modification. The Micron DDR3 memory could overclock to 1126mhz by applying a +125mhz offset. Temperatures stayed under 83C for all testing.

The HD 4600 was unstable and unable to complete the benchmark suite at 1400mhz compared to the stock 1300mhz due to heavy undervolting at 0.880v in the stock BIOS compared to desktop HD 4600 variants running at 1.0v, according to my Gigabyte Z97X SOC Force BIOS. No overclocking results for the HD 4600 will be presented due to the poor power/voltage/clockspeed scaling that could make overclocking have detrimental performance effects due to increased pressure on the limited power budget. Stock clockspeeds stayed under the thermal throttling threshold at 88C.

The Iris Pro 5200 could achieve 1400mhz on the desktop without crashing, however, it is heavily power throttled in benchmarks with the stock peak 1300mhz clockspeed consuming up to 44W (!) in the GPU-Z 3D render test. This meant the overclock only had a marginal effect at best on the performance of the Iris Pro, with unknown stability due to the much lower practical clockspeeds running between 950-1100mhz (while consuming around 30W) in benchmarks. Thermal throttling was also present at these clockspeeds, crossing the thermal throttling point of 92C. No overclocking results will be presented for the Iris Pro due to being a waste of time under severe power and thermal throttling.

Results

Iris Pro "Dominance"?

Total War: Shogun 2 (900p DX11 low preset), Tomb Raider (2013) (768p NotebookCheck settings), Sniper Elite 3 (720p low preset), and Shadow of Mordor (720p low preset) show the Iris Pro 5200 can take the performance lead with a substantial difference compared to the stock GT 730M where overclocking is not enough to fully close the gap. In these situations, the overclocked GT 730M is 5-10% behind in performance with the exception of Shadow of Mordor, where it is about 20% behind. With the exception of Total War: Shogun 2, the HD 4600 also has strong performance compared to the stock GT 730M. 

Too close to call...

Company of Heroes 2 (768p, low preset), Deus Ex Mankind Divided (720p low preset), and Hitman (2016) (720p low preset) show the stock GT 730M to be marginally slower than the Iris Pro 5200 but slightly faster when overclocked. Deus Ex is not really playable for any of the GPUs while other two games are on the cusp of crossing the 30 FPS bar. The performance gap between the HD 4600 and GT 730M at stock is not substantial enough for the GT 730M to pull away with practical, playable frame rates, however close to 30 FPS it may be after overclocking.

Underdog victories

The GT 730M matched or outperformed the Iris Pro 5200 in Just Cause 2 (1080p, default settings, no PhysX), Sleeping Dogs (768p NotebookCheck settings), Dirt Rally (1080p, low preset), World of Tanks (768p/1080p), and Final Fantasy XV (720p, lite). Just Cause 2 is an interesting result as it shows the GT 730M is heavily affected by enabling anti-aliasing due to the limited memory bandwidth, however, we will still hand the win to the GT 730M since it could cross the 30 FPS bar without anti-aliasing. Sleeping Dogs is marginally slower on the GT 730M at stock compared to the Iris Pro 5200 but the minimum frame rates are substantially greater when it is overclocked which leads to less stuttering in game. Dirt Rally is a substantial victory for the GT 730M with stock performance matching the Iris Pro, however, either GPU option will present a playable experience. World of Tanks is a similar story, where either preset seems playable for either GPU but the experience is better on the GT 730M, especially when overclocked. Final Fantasy XV is not really playable, with the benchmark deeming all GPUs insufficient and is included for academic purposes. The difference between the HD 4600 and GT 730M is substantial for this set of benchmarks with a practical real world result of playable versus unplayable framerates for many of the titles.

Discussion

The Iris Pro 5200 theoretically won the performance crown but things are less clear in reality when placed in a more realistic context where users just want to install the CPU and go, setting up nothing more than only the most necessary software tweaks to get going. The very high 44W power draw of the Iris Pro 5200 prevents it from achieving it's peak performance which leads to the smaller than expected performance improvement over the HD 4600. Various degrees of power throttling will be a fact of life for the Iris Pro 5200 as the T440p VRM is not capable of supplying more than 93W to the CPU without releasing the magic smoke in my experience. Assuming infinite cooling for a moment, this means there is only enough power for either the CPU or GPU to run at full speed, but not both. This is not a problem for the GT 730M, which has it's own VRM circuitry on both T440p and T540p boards. 

The Iris Pro 5200 is truly hindered by the loss of the miniDisplayPort output because it loses the one unique saving grace that the GT 730M could never contest which is simply better performance driving dual 1440p monitor setups since the GT 730M is not connected to the ThinkPad's display outputs and the HD 4600 struggles with interface lag while a YouTube video is running. Without that capability, performance must be something it should be able to fall back upon but the current results cast doubt upon even that. Future blog posts will attempt to improve it's performance with various hardware and firmware modifications. Failing that, it should still have some niche appeal to those who dislike NVIDIA drivers, such as Linux users who can accept it's shortcomings. 

One thing to note is that while Lenovo has kindly made the overpriced 47x2MQ CPUs highly redundant with the stock TDP settings of 47/37 short/long power limits, it is highly advisable that Crystalwell users increase the power limits as much as possible as the Iris Pro 5200 is more limited by power than thermal throttling. 

The GT 730M performed surprisingly well for a GPU that was expected to underperform. It features substantial headroom for overclocking, relatively decent performance, and wide availability. However, it is also apparent that the claims of the GT 730M underperforming relative to the HD 4600 are not totally baseless despite at no point does the HD 4600 ever outperform the GT 730M. Future blog posts will investigate modding the VBIOS in order to fully realise the GK108 core and to keep things interesting against the improved Iris Pro 5200.

Thinkpad T440p/T540p Iris Pro 5200 GPU Comparison Part 1

Introduction Figure 1: The 4980HQ (left) and 4900MQ (right) used in this comparison Prior to the availability of Crystalwell CP...