Intel Xeon "Sapphire Rapids" Processor With 20 Cores Tested

[AleksandarK]

Intel is slowly preparing to launch its 4th generation of Xeon Scalable processors, with it being the first arrival of the 10 nm designs to the server market. Codenamed Sapphire Rapids, these processors are expected to bring much-needed IPC and platform improvements so Intel can keep up with AMD's EPYC processors. Today, we are getting some first performance results as well as some information about a specific 20 core, 40 threaded Intel Xeon Sapphire Rapids SKU. In a leaked Geekbench 4 submission, the latest Xeon processor was tested and we get to see even more details about the processor.

Featuring 20 cores and 40 threads, the CPU has a base clock speed of 1.5 GHz. It features as much as 40 MB of L2 cache and 75 MB of L3 cache spread across the die. The system was tested on an Intel reference platform called VulcanCity, with this configuration carrying 32 GB of DDR5 memory. The reported results of the benchmarks that this processor went through are not very impressive. These numbers are easily beaten by AMD Ryzen 9 5950X, however, this is only an engineering sample with low clock speed and it could be possible that Geekbench is not optimized to run on this processor. You can check out some of the performance numbers below, and see the submitted results here.


View at TechPowerUp Main Site

 

[persondb]

That's pretty terrible even for an engineering sample as clocks are only going to up to around 3GHz or so anyway, while a 5950X is 5-6x the single core perfomance and double the multi core, see https://browser.geekbench.com/v4/cpu/16250163

Of course, not apples to apples comparison as server cpus have generally much lower clocks.

Still, I think that there's something wrong in there just look at the memory latency and bandwidth, with it having 206 ns for latency and 8.9GB/s for bandwidth in single core and 283 ns and 17.5GB/s for bandwidth. I think that there's something going on in the memory that's possibly dragging it down.

 

[Vya Domus]

I mean Cypress Cove was pretty unimpressive, there were even some regressions if I am not mistaken compared to the previous architecture. Since this is basically based on that + low clock speeds, I don't have many expectations.

 

[ZoneDymo]

iirc it was George Michael that said: "you gotta have faith"

 

[Midland Dog]

I mean Cypress Cove was pretty unimpressive, there were even some regressions if I am not mistaken compared to the previous architecture. Since this is basically based on that + low clock speeds, I don't have many expectations.

cache latency ruined cypress cove, other than that faster than skylake in all metrics, cache latency just hides that really well

 

[Chrispy_]

I know it's an ES, and I know GB4 is a near-worthless benchmark - but that's still very disappointing; An 8C/16T 3700X from two years ago scores significantly higher and current-gen EPYC Milan, the *actual* competition is four times faster.

 

[Logoffon]

4th generation of Xeon Scalable processors, with it being the first arrival of the 10 nm designs to the server market.

So Ice Lake-SP on the rather confusing 83xx/63xx/53xx/43xx lineup isn't 10nm after all?

 

[jared889]

That's pretty terrible even for an engineering sample as clocks are only going to up to around 3GHz or so anyway, while a 5950X is 5-6x the single core perfomance and double the multi core, see https://browser.geekbench.com/v4/cpu/16250163

Of course, not apples to apples comparison as server cpus have generally much lower clocks.

Still, I think that there's something wrong in there just look at the memory latency and bandwidth, with it having 206 ns for latency and 8.9GB/s for bandwidth in single core and 283 ns and 17.5GB/s for bandwidth. I think that there's something going on in the memory that's possibly dragging it down.

one thing is geekbench and real life, M1 "beats" intel in geekbench but not in real life

M1 Cant beat 6 core i7


AMD Ryzen 9 5950X