4 ms·

20c Apple M1 Ultra Nearly Matches 64c Threadripper 3990X in Geekbench 5

15 points by ashvardanian 5y ago

livinglist 5y ago
RIP intel, i will miss you from time to time, probably….hopefully
- ashvardanian 5y ago
  Intels upcoming CPUs are also quite impressive. Apple has great benchmarks, but porting software into their ecosystem is problematic and their SoCs aren’t available on the server market. If only we could have these machines with 200 Gbit networking, CUDA-programmable GPU's and in a 1U rack-mountable chassis...
  - ashvardanian 5y ago
    At least with the networking part, the bandwidth is there, its just spliced between Thunderbolt ports, which are also highly under utilized without proper eGPU support.
  - olliej 5y ago
    If only other companies provided … NVIDIAs proprietary API?
CaptainNegative 5y ago
What a strange comparison. The 3990X is over 2 years old now, and out of stock everywhere except from 3p resellers. Does AMD even make them anymore? We're allegedly about two weeks away from Zen3.
- ashvardanian 5y ago
  Because Zen 3 Threadripper was announced only 2 days ago [1]. It’s not yet available even for preorders. Furthermore, the improvement of Zen 3 over Zen 2 is marginal. [1]: https://www.anandtech.com/show/17296/amd-announces-ryzen-threadripper-pro-5000-wx-series-zen-3-core-for-oem-workstations https://www.anandtech.com/show/17296/amd-announces-ryzen-thr...
  - adrian_b 5y ago
    Having replaced a Zen 2 CPU on my motherboard with a Zen 3 CPU, more than a year ago, I can guarantee that the improvement of Zen 3 over Zen 2 is not marginal. Zen 2 had an IPC similar to Skylake, while Zen 3 has an IPC slightly better than Tiger Lake. The difference between Zen 3 and Zen 2, even at equal clock frequencies, is very noticeable. The benchmark discussed here is not very significant not only because Zen 2 is obsolete, but also because Geekbench does not scale well to more than 32 cores, so the real multi-threading performance of a 64-core CPU is not demonstrated well with Geekbench.
happymellon 5y ago
What sort of real world things does Geekbench represent? Are we saying that it can do CPU ray tracing as well as a threadripper, or is this more lightweight multithreading and would be an excellent webserver? Sun had systems that were much better than x86 in high threaded environments. But ultimately in the real world x86 was better. I know my M1 Pro MacBook Pro is definitely not better than my Intel desktop when I use it for writing code. Although it could also be operating systems and supporting software that are giving me the contrasting experience.
- enduser 5y ago
  It will be interesting to see how the M1 Pro MacBook Pro stacks up against the Studio M1 Pro in real-world performance, given that over half of the Studio is dedicated to thermal management.
- olliej 5y ago
  I have this vague recollection the some benchmarks use long double which is a severe handicap on all x86 platforms other than win64. I can’t recall if the geek bench use of long double is a few of the c++ APIs specified in terms of long double or if it actually does real work with them.
zamadatix 5y ago
Note the 32c 3970X gets nearly 90% the multicore score of the 64 core 3990X i.e. what Geekbench 5 is benchmarking doesn't care about the majority of the cores despite the focus in the headline. What's more impressive here is the comparison the 16 core 5950X where the M1 Ultra still whomps it on single core but now also multi-core on top of having 4 extra cores that probably weren't accounted for much by this benchmark. It'll be really interesting what Raptor Lake vs Zen 4 vs "M2" bring to the table all hopefully by the end of the year.
chlorion 5y ago
Whatever the geekbench multicore test is testing doesn't seem to actually benefit from extra cores very much. Not all real world workloads will scale perfectly with moar cores, but considering that the test is supposed to benchmark multicore performance it should be doing roughly that, otherwise what exactly is the point? This new system is very cool but there seems to be a lot of phony benchmark results with the intention of making it look faster than it actually is. I am very interested in seeing how phoronix results end up comparing to these other tests!

15 points by ashvardanian 5y ago

livinglist 5y ago

RIP intel, i will miss you from time to time, probably….hopefully

ashvardanian 5y ago
Intels upcoming CPUs are also quite impressive. Apple has great benchmarks, but porting software into their ecosystem is problematic and their SoCs aren’t available on the server market. If only we could have these machines with 200 Gbit networking, CUDA-programmable GPU's and in a 1U rack-mountable chassis...
- ashvardanian 5y ago
  At least with the networking part, the bandwidth is there, its just spliced between Thunderbolt ports, which are also highly under utilized without proper eGPU support.
- olliej 5y ago
  If only other companies provided … NVIDIAs proprietary API?

CaptainNegative 5y ago

What a strange comparison. The 3990X is over 2 years old now, and out of stock everywhere except from 3p resellers. Does AMD even make them anymore? We're allegedly about two weeks away from Zen3.

ashvardanian 5y ago
Because Zen 3 Threadripper was announced only 2 days ago [1]. It’s not yet available even for preorders. Furthermore, the improvement of Zen 3 over Zen 2 is marginal. [1]: https://www.anandtech.com/show/17296/amd-announces-ryzen-threadripper-pro-5000-wx-series-zen-3-core-for-oem-workstations https://www.anandtech.com/show/17296/amd-announces-ryzen-thr...
- adrian_b 5y ago
  Having replaced a Zen 2 CPU on my motherboard with a Zen 3 CPU, more than a year ago, I can guarantee that the improvement of Zen 3 over Zen 2 is not marginal. Zen 2 had an IPC similar to Skylake, while Zen 3 has an IPC slightly better than Tiger Lake. The difference between Zen 3 and Zen 2, even at equal clock frequencies, is very noticeable. The benchmark discussed here is not very significant not only because Zen 2 is obsolete, but also because Geekbench does not scale well to more than 32 cores, so the real multi-threading performance of a 64-core CPU is not demonstrated well with Geekbench.

happymellon 5y ago

What sort of real world things does Geekbench represent? Are we saying that it can do CPU ray tracing as well as a threadripper, or is this more lightweight multithreading and would be an excellent webserver? Sun had systems that were much better than x86 in high threaded environments. But ultimately in the real world x86 was better. I know my M1 Pro MacBook Pro is definitely not better than my Intel desktop when I use it for writing code. Although it could also be operating systems and supporting software that are giving me the contrasting experience.

enduser 5y ago
It will be interesting to see how the M1 Pro MacBook Pro stacks up against the Studio M1 Pro in real-world performance, given that over half of the Studio is dedicated to thermal management.
olliej 5y ago
I have this vague recollection the some benchmarks use long double which is a severe handicap on all x86 platforms other than win64. I can’t recall if the geek bench use of long double is a few of the c++ APIs specified in terms of long double or if it actually does real work with them.

zamadatix 5y ago

Note the 32c 3970X gets nearly 90% the multicore score of the 64 core 3990X i.e. what Geekbench 5 is benchmarking doesn't care about the majority of the cores despite the focus in the headline. What's more impressive here is the comparison the 16 core 5950X where the M1 Ultra still whomps it on single core but now also multi-core on top of having 4 extra cores that probably weren't accounted for much by this benchmark. It'll be really interesting what Raptor Lake vs Zen 4 vs "M2" bring to the table all hopefully by the end of the year.

chlorion 5y ago

Whatever the geekbench multicore test is testing doesn't seem to actually benefit from extra cores very much. Not all real world workloads will scale perfectly with moar cores, but considering that the test is supposed to benchmark multicore performance it should be doing roughly that, otherwise what exactly is the point? This new system is very cool but there seems to be a lot of phony benchmark results with the intention of making it look faster than it actually is. I am very interested in seeing how phoronix results end up comparing to these other tests!