Geekbench 5 iMac Metal Rough Comparisons - which card to get?

[PhotoVid]

We've discussed this a bit in other threads and here's the first comparison I've been able to do for those wondering what to do owning a 580x iMac.

5700 16gb is 57898, 59417, 56291
5700 8gb scores 53232, 49344 (https://browser.geekbench.com/v5/compute/1369517)
5500 scores 41727, 41283
The 5300 36169.
The 580x 33462.


The difference in upgrades:

580x to base mode: approx 7.5% gain

Base model to 5500: approx 13.4% gain

5500 to 5700: approx 21.7% gain

5700 to 5700xt 16gb: approx 11.3% gain

Your best bang for the buck seems to be between moving to the 5500 or grabbing the 5500 to 5700 8gb upgrade.

Again rough numbers - anyone have any other scores to share?

 

[wave84]

In all the tests I've seen, the 580x is pretty much on par with the 5500 at about 41K. Where did you get that score from?
The rest of your numbers are in line with what I've seen.

 

[creativityrules]

When looking at Geekbench, you need to make sure the CPU is roughly the same across 5500XT, 5700, and 5700XT.
Below is my best effort filtering through the published results (Metal with i7 core) on geekbench:
Unlike OpenCL results, the Metal results don't always list the GPU spec.

 

[KrazyKanuck]

I was just looking through some of the Geekbench results, and I noted that in OpenCL (sorry, a little off-topic ?), in the System Information section, the 5700 XT is profiled as having 40 compute units, which makes sense. But the 5700 is profiled as having only 32, rather than the 36 it should have. Now, I know only enough about all of this to know how ignorant I am, but can somebody suggest a reason that the other 4 compute units aren't showing up? This seems to be consistent across all 5700 OpenCL scores.

 

[KrazyKanuck]

The 5300 OpenCL results show 20 compute units, which is correct. And the 5500 XT OpenCL results show 24 compute units, which is also correct. It's just the 5700 that shows a number of compute unites that doesn't match Apple or AMD's specifications. I'm guessing that those "missing" 4 compute units on the 5700 would make at least some difference in these OpenCL scores, if perhaps less difference in real world use. I'll stress that I'm not terribly comfortable in my understanding of these synthetic benchmarks or the intricacies of GPUs, but the significant difference between the 5700 and 5700 XT benchmarks that we've seen in these new iMacs has struck me as odd, especially given how much closer they preform relative to each other as drop-in cards. I wouldn't have thought that 8 vs.16GB of VRAM would make quite that substantial a difference in these benchmarks. Might AMD or Apple have disabled those 4 compute units in the 5700 to more evenly space the performance gains from moving from the 5500 XT to the 5700 to the 5700 XT (ie. 24 CUs, 36 32CUs, and 40CUs, for jumps of 8CUs) and forgotten to tell us all about it? ?

 

[creativityrules]

The 5300 OpenCL results show 20 compute units, which is correct. And the 5500 XT OpenCL results show 24 compute units, which is also correct. It's just the 5700 that shows a number of compute unites that doesn't match Apple or AMD's specifications. I'm guessing that those "missing" 4 compute units on the 5700 would make at least some difference in these OpenCL scores, if perhaps less difference in real world use. I'll stress that I'm not terribly comfortable in my understanding of these synthetic benchmarks or the intricacies of GPUs, but the significant difference between the 5700 and 5700 XT benchmarks that we've seen in these new iMacs has struck me as odd, especially given how much closer they preform relative to each other as drop-in cards. I wouldn't have thought that 8 vs.16GB of VRAM would make quite that substantial a difference in these benchmarks. Might AMD or Apple have disabled those 4 compute units in the 5700 to more evenly space the performance gains from moving from the 5500 XT to the 5700 to the 5700 XT (ie. 24 CUs, 36 32CUs, and 40CUs, for jumps of 8CUs) and forgotten to tell us all about it? ?

That’s very interesting find! I’m wondering if this is a Geekbench 5 issue where it is not picking up the 4 compute units during the test? Maybe it is an AMD driver issue? Apple advertises there are 36 compute units in the 5700 card. Could explain why the performance gap is smaller when comparing Unigine heaven results.

 

[KrazyKanuck]

That’s very interesting find! I’m wondering if this is a Geekbench 5 issue where it is not picking up the 4 compute units during the test? Maybe it is an AMD driver issue? Apple advertises there are 36 compute units in the 5700 card. Could explain why the performance gap is smaller when comparing Unigine heaven results.

I would like to think that this is just some quirk with the benchmarking app. I really want that to be the case. Is there some reason that the other 4 compute units aren't made available for the benchmarking app on this particular card? This sort of makes sense, since the cards do seem to perform more alike during gaming, as you suggest. Surely it's not the case that Apple or AMD has actually disabled the cores. That would be false advertising. Like buying an 18-core CPU and discovering that only 16-cores are enabled.

 

[KrazyKanuck]

When looking at Geekbench, you need to make sure the CPU is roughly the same across 5500XT, 5700, and 5700XT.
Below is my best effort filtering through the published results (Metal with i7 core) on geekbench:
Unlike OpenCL results, the Metal results don't always list the GPU spec.
View attachment 947467

If, say, Geekbench isn't using those additional 4 compute engines, then that might explain why the spread between these iMac cards doesn't mirror the spread between the drop-in versions. The 5700 shouldn't fit right in the middle between the 5500 XT and the 5700 XT, even with the 5700 XT's extra VRAM. The 5700 and the 5700 XT should be far more similar in terms of their compute scores than these scores seem to show. Let's say those extra 4 CUs aren't reflected in these scores. Add them back in and that moves the 5700 closer to 54000-55000, which is about where it should be compared to the 5700 XT. My understanding, and someone correct me if I'm wrong, is that 8GB vs 16GB of VRAM shouldn't really impact Geekbench's compute scores all that much, since the benchmark is designed to use a set amount of RAM and VRAM based upon an "average" of what is commonly available.