GTX 1650 vs GTX 750 Ti
Grid Hardware GPGPU
Published: 2019-05-19

After months of pondering what GPU to use to get my compute farm back into crunching GPGPU workunits, I was persuaded by a recent review to order a GTX 1650 and give it a go.

Background and lead-up

Years ago, I was reading stories about how Linux support on Steam had really improved. I decided to build a Linux box capable of running modern-ish games. This was in 2015, a time when AMD GPU support on Linux was abominable, so my choice for a video card boiled down to “which Nvidia card?” My answer was the GTX 750 Ti.

It wasn’t a new card at the time, but I was coming from an iMac, so it was good enough for me. It was also readily available and cheap – this was an experimental build, after all. If the results were not good, I didn’t want to be out a thousand bucks.

The 750 Ti worked fine, and that machine is where I did all my gaming until a friend talked me into buying a PS4. By that time, I had already gotten into grid computing, but switching to gaming on the PS4 let my old gaming rig become my new dedicated crunchbox, and the 750 Ti became my first GPGPU.

I’ve always had a thing about efficiency in computing. For me the question is neither “what’s got the most grunt on the market”, or “what’s the lowest-power thing available”. For me the question is: “Given my compute requirements, what gets me there with the fewest watts burned?”

As such, I’ve never been interested in GPUs with external power cables. It turned out that the 750 Ti was pretty good by this metric, cranking out respectable numbers on a ~40W power budget. When the 1050 Ti came out, I upgraded to it, and put the 750 Ti in a spare machine to continue crunching.

But then I got really into World Community Grid, which has no GPGPU subprojects, and I started to drift away from GPU based projects. And then I did the Greenhouse build – my current compute setup – which is four machines, in a tower, in my living room. The power and cooling needs from that didn’t make me hurry to add GPUs back into the mix.

For a while I was considering GT 1030s. They are approximately 0.8X as powerful as a 750 Ti, but only consume about 30W. Unfortunately, they also currently cost $95.

Eventually I decided that power and cooling would be fine, and started looking for the right card. 750 Tis cost as much used as they did new in 2015, and 1050 Tis have just become very hard to find (and are also overpriced, in my opinion).

When the GTX 1650 was released, I was hopeful but not excited. Early reviews boiled down to “Meh. I guess, if you need something that runs on PCI power only”. I did need that, but I didn’t want to spend money on a thoroughly mediocre card.

Finally, earlier this week, Anandtech published an in-depth review of the card, including lots of information on compute benchmarks and power usage. They found that, in synthetic benchmarks, the 1650 was between 2X and 3X the performance of a 750 Ti (with an average of something like 2.3X). The specific card they reviewed was the Zotac variant, which they noted ran particularly cool. I decided to give it a try.

Specs and operating data

GTX 750 Ti

My 750 Ti card is the EVGA 750 Ti SC. As you can see, at full load, it uses 36-37W. The fan is tuned to be on the conservative side, resulting in a card that runs at between 64C and 66C under load.

GeForce GTX 750 Ti (driver version 430.14, CUDA version 10.2, compute capability 5.0, 2000MB, 1963MB available, 1606 GFLOPS peak)

Sun May 19 21:44:36 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.14       Driver Version: 430.14       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 750 Ti  Off  | 00000000:26:00.0 Off |                  N/A |
| 53%   64C    P0    37W /  38W |     55MiB /  2000MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+

GTX 1650 Ti

The Zotac Gaming 1650 uses 1.7X the power of the 750 Ti, but runs 5-6 degrees cooler. I’m unsure if this is because of the more aggressive fan profile, or a better heat sink design, or both, but under full load it has not yet hit 61C, and is usually at 59C.

GeForce GTX 1650 (driver version 430.14, CUDA version 10.2, compute capability 7.5, 3909MB, 3850MB available, 3037 GFLOPS peak)

Sun May 19 19:00:30 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.14       Driver Version: 430.14       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1650    Off  | 00000000:26:00.0 Off |                  N/A |
| 64%   58C    P0    66W /  75W |     71MiB /  3909MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+

Cooling considerations

This has less to do with the cards, and more to do with the Greenhouse design. It’s based on mini-ITX boards, in an open enclosure, on trays that are just barely bigger than the mobos themselves. Originally, 120mm fans were arranged on either side of a system tray, like this:

---     --------     ---
|F|     |      |     |F|
|A| <-- | MOBO | <-- |A|
|N|     |      |     |N|
---     --------     ---

(Arrows indicate airflow/fan direction)

I knew it would be tight, but I thought I had left enough room on the left side of the tray for a GPU to be plugged into the mobo’s single PCIe slot. And I had, in a technical sense, but not in a realistic sense.

Before ordering the 1650, I plugged in my venerable 750 Ti to test things out. The GPU fan blades and the exhaust fan blades were basically touching, which was not good.

Since the chassis was open, I had been wondering if the “exhaust” fan was even contributing much. It was further away from the CPU than the “intake” fan. So as a first step, I simply removed it.

I still don’t think that the exhaust fan had been contributing much, but replacing it with a wall of PCB which self-heated to 66C, sure made a difference. CPU temps on that machine went from about 62C to 70C within minutes.

I decided to replace the fan, situating it at the back of the mobo, blowing inward. My thought was that (again) due to the open chassis, this would produce sort of a cross-breeze.

    -------
    | FAN |
    -------
       |
       V
--- --------     ---
|G| |      |     |F|
|P| | MOBO | <-- |A|
|U| |      |     |N|
--- --------     ---

Instead, this caused CPU temps to rocket to 74C within a minute of turning the machine back on. I think that the IO ports, combined with the fact that the CPU is closer to the back of the mobo than the side, led to the rear fan causing hot air from the CPU’s HSF to become trapped. I decided to try flipping the rear fan’s direction, returning it to functionally being an exhaust fan:

       ^
       |
    -------
    | FAN |
    -------
--- --------     ---
|G| |      |     |F|
|P| | MOBO | <-- |A|
|U| |      |     |N|
--- --------     ---

This brought CPU temps down lower than they had been before: about 59C.

Performance comparison

I let the 750 Ti crunch some Primegrid WUs, to establish a basis for comparison. Then I installed the 1650 and let it get to work as well. Here are the results:

Project Workunit 750 Ti time 1650 time Perf
Primegrid PPS Sieve 177s 140s 1.26X
Primegrid Genefer 15 92s 36s 2.55X
Primegrid Genefer 17 Low 21min 23s 10min 03s 2.12X
Primegrid Genefer 17 Mega 23min 09s 10min 50s 2.14X
Primegrid Genefer 18 82min 52s 37min 47s 2.19X
Primegrid Genefer 20 12h 24min 22s 5h 27min 09s 2.27X
Einstein@Home GRPBS1-GPU 54min 03s 26min 32s 2.03X

I would have included GPUGrid results as well, but as of the time of writing, they have suspended Linux crunching.

Conclusion

I’m happy with this card. It would be better if it beat up the 750 Ti while using the same amount of power, but right now semiconductors just aren’t in a space where we’re gonna get those kind of gains.

I do like that this particular card runs so cool while putting in the work though, since I run GPUs 247.

Unless a 1650 Ti comes along, or AMD pulls something amazing out of their hat with Navi, this is as good as it gets without exceeding motherboard power limits.