16" MBP Geekbench scores so far show big differences

[DaakuMaujii]

can we get some results with multiple runs back to back.. for instance.. my mid 2018 MacBook Pro i7 dropped over 200 points between the first run and the 4th run.. due to thermals.. it never could boost for longer than 3 seconds without having to throttle

Just did that on the i7-2.3GHz 16GB, 5300M 4GB. First five runs are consecutive (browser window with results pop up, note down results, close browser, hit the run button) on AC charger, then rebooted the system, disconnected the AC charger, let the system settle for a couple minutes before running the second set of five consecutive tests on battery power.

Run	Single-core score	Multicore score
On AC - run 1	1087	5733
On AC - run 2	1164	5699
On AC - run 3	1175	5712
On AC - run 4	1174	5710
On AC - run 5	1168	5724
Run	Single-core score	Multicore score
On battery - run 1	1045	5721
On battery - run 2	1124	5735
On battery - run 3	1113	5710
On battery - run 4	1120	5702
On battery - run 5	1104	5731

I don't have system monitoring tools installed yet, so I don't have detailed information on the load on the system, cpu temperatures or fan speeds. But only during the last two or three multicore tests (typically around the time Speech recognition test is running) do I hear the fans start to kick in, and settle soon after the test is finished.

 

[-narcan-]

1136
Single-Core Score
6834
Multi-Core Score

on an i9 / 16gb / 5500 4gb

 

[gigapocket1]

Just did that on the i7-2.3GHz 16GB, 5300M 4GB. First five runs are consecutive (browser window with results pop up, note down results, close browser, hit the run button) on AC charger, then rebooted the system, disconnected the AC charger, let the system settle for a couple minutes before running the second set of five consecutive tests on battery power.

Run	Single-core score	Multicore score
On AC - run 1	1087	5733
On AC - run 2	1164	5699
On AC - run 3	1175	5712
On AC - run 4	1174	5710
On AC - run 5	1168	5724
Run	Single-core score	Multicore score
On battery - run 1	1045	5721
On battery - run 2	1124	5735
On battery - run 3	1113	5710
On battery - run 4	1120	5702
On battery - run 5	1104	5731

I don't have system monitoring tools installed yet, so I don't have detailed information on the load on the system, cpu temperatures or fan speeds. But only during the last two or three multicore tests (typically around the time Speech recognition test is running) do I hear the fans start to kick in, and settle soon after the test is finished.

appreciative! Def looks like it can sustain loads a lot better than previous MacBook pros. I mean. We all knew that but it’s good to see some scores and see just how good it is at not maxing out.

 

[am2am]

Here are my results (i9 2.4, 5500 M with 8GB, 32RAM, 2TSSD):


out of curiosity I've set fans at max - results below:


Cinebench results here:

the same with fans at max:

 

[miroki]

While sipping my morning tee I had some inspiration, so I have parsed all the currently available Geekbench 5 results for the 16" MBP — see the results below. Some explanation: the dots represent the individual score, the lines are the 25%, 50% (median) and 75% quantiles (that is, everything between the top and bottom line is where 50% of the relevant scores are), and the shapes represent the density distribution of the scores (so called violin plot, which are my favourites)

In addition, I ran a simple unequal variance T-test on the data, with the following result:

- assuming 5% significance threshold, there is no statistical difference between the single-core results for i9-9880H and i9-9980HK, albeit the i9-9980HK is a bit faster on average (95% confidence interval of score differences between -5.52462 and 82.82824 )

- assuming 5% significance threshold, there is a statistical difference between the multi-core results for i9-9880H and i9-9980HK (95% confidence interval of score differences between 497.2288 and 746.6590)

Bottomline: i9-9980HK is around 10% faster on average for multi-core (and 3% faster on average for single-core). This is quite unexpected given the CPU specs. The HK CPUs are rigorously binned and/or Apple might be undervolting them. At any rate,I am starting to regret that I ordered the i9-9880H, it might need to go back 😅

EDIT: added RAM, as requested by @The Mercurian . I also did a quick regression analysis on this, and RAM does not make any difference in a regression model. But at the same time, we can't really analyse the impact of RAM as we don't the enough data points: users who buy the high-end CPU also never seem to get 16GB RAM. Here is a table:

CPU RAM n
i7-9750H 16384 MB 109
i9-9880H 16384 MB 87
i9-9980HK 32768 MB 40
i9-9980HK 65536 MB 1



View attachment 877675View attachment 877676

If there isn’t any significant difference, why do you regret to get the cheap 2.3ghz CPU?

 

[NeroAugustus]

If there isn’t any significant difference, why do you regret to get the cheap 2.3ghz CPU?

There is an actual difference in multi core it seems.

I was hesitating between the 2.3 and 2.4 versions, but I've seen a review showing the 2.4 version actually performs worse overall once the gpu is involved in sustained load like games. As if the 2.4 hoards too much ressources and starves the GPU. It makes sense to me considering how small the available power package is for this laptop (100W is nothing for such a powerful machine.)

If that's not the explanation then it's just CPU binning, in which case I'm fine with playing the lottery and potentially getting a 2.3 that could match/outperform a 2.4 in CPU+GPU loads.

Just did that on the i7-2.3GHz 16GB, 5300M 4GB. First five runs are consecutive (browser window with results pop up, note down results, close browser, hit the run button) on AC charger, then rebooted the system, disconnected the AC charger, let the system settle for a couple minutes before running the second set of five consecutive tests on battery power.

Pretty funny it's as if the laptop needs a "warmup" considering single core scores.

 

[leman]

If there isn’t any significant difference, why do you regret to get the cheap 2.3ghz CPU?

Because lack of significance does not mean that there is no difference And there is a healthy boost in multi-processing score.

Then again, Geekbench is a really bad source to base your purchases on. This benchmark is not representative and woefully unreliable.

 

[bsbeamer]

GB5 CPU for MBP16 with 2.4/8GB/64GB/1TB/10.15.1 reports:
1125 single
7247 multi

Same for METAL for 5500M 8GB:
27580

Same for METAL with 5700 XT 8GB in Sonnet 650 eGPU:
33723

Same for METAL with Intel 630:
4974


For comparison GB5 CPU for MP5,1 with 2 x 3.46 (X5690)/128GB RAM/RX580 8GB/NVMe/10.14.6 reports:
666 single
5043 multi

Same MP5,1 for METAL with RX 580 8GB:
39004

 

[intrepidcase]

Because lack of significance does not mean that there is no difference And there is a healthy boost in multi-processing score.

Then again, Geekbench is a really bad source to base your purchases on. This benchmark is not representative and woefully unreliable.

Yep, it may not be 100% representative but in lieu of other test numbers it's better than simply guessing. At the end of the day, the sampled benchmarks suggest a performance improvement roughly in line (4% single-core) or better (10% multi-core) than the nominal clock difference of 4%. It clearly is somewhat better - its up to each person to decide if $150/$200 is worth it for this increase in raw performance.

 

[Spectrum]

GB5 CPU for MBP16 with 2.4/8GB/64GB/1TB/10.15.1 reports:
1125 single
7247 multi

Same for METAL for 5500M 8GB:
27580

Same for METAL with 5700 XT 8GB in Sonnet 650 eGPU:
33723

Same for METAL with Intel 630:
4974


For comparison GB5 CPU for MP5,1 with 2 x 3.46 (X5690)/128GB RAM/RX580 8GB/NVMe/10.14.6 reports:
666 single
5043 multi

Same MP5,1 for METAL with RX 580 8GB:
39004

I find it astonishing how fast the internal 5500M GPU is when compared to the eGPU 5700XT solution.

Given how little space and cooling the MB Pro has, why are eGPU cards and boxes so ridiculously large?

 

[leman]

I find it astonishing how fast the internal 5500M GPU is when compared to the eGPU 5700XT solution.

Given how little space and cooling the MB Pro has, why are eGPU cards and boxes so ridiculously large?

For one, the eGPU is limited by the 4x PCI interface, especially for compute tasks since there is a lot of data transfers going in and the GPU is likely to stall. On the other hand, the mobile chip in the MBP is ridiculously efficient compared to its large brother.

And yes, those GPUs are humongous. They alone pull over twice as power as the entire MBP, they have massive heatthinks and they include an array of power circuit elements to feed all those watts. Besides, the price needs to be affordable so they sacrifice space for cheaper design.

 

[bsbeamer]

I find it astonishing how fast the internal 5500M GPU is when compared to the eGPU 5700XT solution.

With the current driver, yes. The RX 5700 XT driver needs to be improved, should easily be scoring over the RX 580 in METAL benchmarks. I assume in 10.15.2 or 10.15.3 it’ll be addressed. Using the standard build of macOS may help too, vs the specific build for MBP16.

 

[leman]

With the current driver, yes. The RX 5700 XT driver needs to be improved, should easily be scoring over the RX 580 in METAL benchmarks.

I am quite sure that RX 5700XT (Navi 10) and Pro 5500M (Navi 14) use the same driver. I mean, its the same GPU generation and the only real difference between them is that Navi 14 is basically a "half" of the Navi 10 chip.

 

[bsbeamer]

Might be, but this is the first release of the drivers. If nothing is improved, a sub $200 RX580 in Mojave will be almost 33% faster in METAL than ~$400 RX 5700 XT in Catalina. Does not inspire confidence at all. Only thing it does is justify Apple keeping RX580 as entry GPU for MP7,1.

...and this is not even considering all of the RX580 eGPU crashes in Catalina so far.

 

[Spectrum]

For one, the eGPU is limited by the 4x PCI interface, especially for compute tasks since there is a lot of data transfers going in and the GPU is likely to stall. On the other hand, the mobile chip in the MBP is ridiculously efficient compared to its large brother.

And yes, those GPUs are humongous. They alone pull over twice as power as the entire MBP, they have massive heatthinks and they include an array of power circuit elements to feed all those watts. Besides, the price needs to be affordable so they sacrifice space for cheaper design.

Sure...the bigger eGPU boxes have more headroom and features...but for many this really isn't needed. Especially when the gains are relatively modest.
Regarding cost, how much does the 5500M add to the cost of a MBPro? It is absolutely negligible.
One only has to look at the relative cost of a 13 inch and 16 inchMBPro. When similarly specced, the 16 inch is actually cheaper, yet has a 6 core CPU, larger screen, and dedicated GPU.

I guess, I just would like to see the release of sensible eGPU solutions that use modest, good value, mobile-grade GPU components, in a small form factor. The upgrade compared to Intel would still be massive 4-5x.

 

[leman]

Might be, but this is the first release of the drivers. If nothing is improved, a sub $200 RX580 in Mojave will be almost 33% faster in METAL than ~$400 RX 5700 XT in Catalina.

Ah, you were talking about Navi vs RX 580, sorry I was not paying attention. Yeah, that is probably some sort of bug/problem with Geekbench itself. In real-world tests the 5700 XT beats the RC 580 by a large margin, e.g.: https://barefeats.com/nmp-5700-xt-vs-other-gpus.html

 

[bsbeamer]

That’s basically what the Sonnet Puck eGPU was supposed to be - cheap graphics driver GPU for 13” machines and Mac Mini without dGPU.

https://www.sonnettech.com/product/egfx-breakaway-puck.html

It needs to be updated, and needs to still work correctly in Catalina before it can be recommended. MANY users suddenly had them stop working with the 10.15.1 update. Same with RX580’s.

If you’re looking at eGPU for METAL acceleration, this is not the route to go.

 

[leman]

Yep, it may not be 100% representative but in lieu of other test numbers it's better than simply guessing. At the end of the day, the sampled benchmarks suggest a performance improvement roughly in line (4% single-core) or better (10% multi-core) than the nominal clock difference of 4%. It clearly is somewhat better - its up to each person to decide if $150/$200 is worth it for this increase in raw performance.

What I mean is that I ran Geekbench multiple times on my machine (there were no background tasks interfering), and I got a widely different result every time. There were deltas of 400-500 points between consecutive runs. A benchmark that is not able to produce consistent results is quite useless.

And yes, the nominal clock is totally irrelevant: it is just a marketing metric that only applies to minimal bounds in certain scenarios. It won't say much about the CPU performance. Observations so far suggest that the 2.4 model seems to be more efficient on average.

 

[bsbeamer]

Ah, you were talking about Navi vs RX 580, sorry I was not paying attention. Yeah, that is probably some sort of bug/problem with Geekbench itself. In real-world tests the 5700 XT beats the RC 580 by a large margin, e.g.: https://barefeats.com/nmp-5700-xt-vs-other-gpus.html

Believe it’s a Catalina issue or issue with 10.15.1 or the specific build for MBP16. I’m the first to say I really don’t like GB5 scores, but this shows SOMETHING is up that needs to be addressed. Hopefully by 10.15.2 release.

 

[leman]

Believe it’s a Catalina issue or issue with 10.15.1 or the specific build for MBP16. I’m the first to say I really don’t like GB5 scores, but this shows SOMETHING is up that needs to be addressed. Hopefully by 10.15.2 release.

You might be right but so far my Pro 5500M outperforms — by far — my Vega Pro 20 in every scenario (and I did run some OpenCL and Metal workflows), yet their GB5 compute score are very close. So there might be some sort of driver inefficiency that shows itself in GB5 specifically, or (a simpler hypothesis), Geekbench is a bad benchmark

 

[Chronos___]

Has anyone tried clamshell mode (closed - connected to peripherals etc) vs normal mode (just the MacBook opened)? When I use my Macbook in clamshell mode it get twice the temperature compared to "normal" mode. (with pictures).
This could explain at least some variance.

 

[bsbeamer]

You might be right but so far my Pro 5500M outperforms — by far — my Vega Pro 20 in every scenario (and I did run some OpenCL and Metal workflows), yet their GB5 compute score are very close. So there might be some sort of driver inefficiency that shows itself in GB5 specifically, or (a simpler hypothesis), Geekbench is a bad benchmark

I wish that was the only issue. Does not explain why RX580/RX570/RX560 in eGPU in Catalina have suddenly stopped working. It’s a widespread issue that indicates something with graphics drivers in the Catalina OS needs to be fixed and/or improved. The low RX5700xt scores in Catalina vs RX580 in Mojave just add additional data. Unfortunately, no way to downgrade MBP16 to Mojave.

 

[NeroAugustus]

I guess, I just would like to see the release of sensible eGPU solutions that use modest, good value, mobile-grade GPU components, in a small form factor. The upgrade compared to Intel would still be massive 4-5x.

I think the issue is that one of the main argument for eGPUs is their upgradability. Upgradable means you can buy off the shelf GPUs and those are only desktop-grade, meant to be used in a huge case, GPUs.

The eGPU market is still very small, so there is no incentive to provide small form factor GPUs. Your best option right now is to look at GPUs meant for micro ITX systems, that's as small as it gets in the desktop PC space. But then there's the enclosure issue, most manufacturers will want to ensure their enclosure fits the most common GPU sizes, meaning nothing too small... so yeah basically, eGPUs are so niche that nothing is really optimized with them in mind.

 

[bsbeamer]

Nothing has really changed or improved from 10.15.1 > 10.15.2 in terms of benchmark results:

MBP16,1 CPU single 1134 and CPU multi 7288 in 10.15.2
MBP16,1 26963 is best METAL benchmark with 5500M 8GB in 10.15.2
MBP16,1 33742 is best METAL benchmark with RX 5700 XT 8GB via eGPU in 10.15.2

 

[iemcj]

Just for comparison to the maxed out 2017 imac in geekbench 5.

1222
Single-Core Score

4649
Multi-Core Score

So somehow this macbook pro is BETTER than a full fledged imac. Wow technology moves fast!