Intel Ice Lake-SP Processors Get Benchmarked Against AMD EPYC Rome

[AleksandarK]

Intel is preparing to launch its next-generation for server processors and the next in line is the Ice Lake-SP 10 nm CPU. Featuring a Golden Cove CPU and up to 28 cores, the CPU is set to bring big improvements over the past generation of server products called Cascade Lake. Today, thanks to the sharp eye of TUM_APISAK, we have a new benchmark of the Ice Lake-SP platform, which is compared to AMD's EPYC Rome offerings. In the latest GeekBench 4 score, appeared an engineering sample of unknown Ice Lake-SP model with 28 cores, 56 threads, a base frequency of 1.5 GHz, and a boost of 3.19 GHz.

This model was put in a dual-socket configuration that ends up at a total of 56 core and 112 threads, against a single 64 core AMD EPYC 7442 Rome CPU. The dual-socket Intel configuration scored 3424 points in the single-threaded test, where AMD configuration scored notably higher 4398 points. The lower score on Intel's part is possibly due to lower clocks, which should improve in the final product, as this is only an engineering sample. When it comes to the multi-threaded test, Intel configuration scored 38079 points, where the AMD EPYC system did worse and scored 35492 points. The reason for this higher result is unknown, however, it shows that Ice Lake-SP has some potential.


View at TechPowerUp Main Site

 

[W1zzard]

If that Ice Lake really was running at 1.5 GHz, vs a 3.4 GHz EPYC, with 15% more cores, then it's quite a huge achievement

 

[dj-electric]

If that Ice Lake really was running at 1.5 GHz, vs a 3.4 GHz EPYC, with 15% more cores, then it's quite a huge achievement

I would not believe that's the case. I'm thinking there's some generic base clock reading.
Its more than likely to see that ICL-SP chips running closer to 3Ghz at load.

With all that taken into count, at least we have comperable results at the high end

 

[Vayra86]

Maybe there is potential, but geekbench and nondescript bench situations + Intel makes me highly suspicious. We've seen this before.

I also am puzzled how we keep thinking Intel will pull some magical rabbit out of the hat even now, after countless announcements of them not reaching targets. And after a decade of confirmation that the low hanging fruit in CPU performance has been picked. Wishful thinking? I think its time to accept the blue Giant has tripped and has been falling since Skylake. Its just a very long way down. So far, every single good looking Intel bench result had a major caveat or was simply an outright lie. Company officials also have histories of being dishonest and keeping vital info from us until they simply can't anymore.

Bottom line. If Intel had anything, we would have seen it and it would have been clearer.

 

[xman2007]

So a 23% or so higher single core score for amd and 8 more physical cores in total yet Intel apparently beats the multicore score by quite a margin, seems legit..

 

[Dredi]

If that Ice Lake really was running at 1.5 GHz, vs a 3.4 GHz EPYC, with 15% more cores, then it's quite a huge achievement

Anything is possible if you double the TDP.

Geekbench does not report actual frequencies achieved during the test, but rather something generic that the CPU/motherboard tell it if asked.

 

[Assimilator]

@W1zzard please can you ban news articles about Geekbench? It is quite simply the most worthless so-called "benchmark" to ever exist. It was designed for ARM smartphones FFS, why do people think it will be any use for gauging x86 or x64 performance?

 

[AnarchoPrimitiv]

How come nobody has mentioned that the dual Ice lake setup is probably more than DOUBLE the cost of the single CPU Epyc Rome Setup. Although we don't know pricing yet, this is Intel we're talking about, and based on literally everything they've done up to this point, I think it'd be safe to assume that two Ice Lake 28 core CPUs are substantially more than a single Eypc 64 core, heck I'm willing to bet that a SINGLE 28 core Ice lake CPU costs more than the single Eypc 64 core CPU.

This is why, in my personal opinion, comparisons between hardware should always account for price first and formost. Nothing exists in a vacuum, and hardware doesn't exist in a world free from the realities of cost, and this metric is especially important in the enterprise market. Besides, the comparison I'm more interested in is Ice lake vs Epyc Zen3 (Milan?, I believe) as that's what the true competition will be based upon for the next year.

How come nobody has mentioned that the daul Ice lake setup is probably more than DOUBLE the cost of the single CPU Epyc Rome Setup. Although we don't know pricing yet, this is Intel we're talking about, and based on literally everything they've done up to this point, I think it'd be safe to assume that two Ice Lake 28 core CPUs are substantially more than a single Eypc 64 core, heck I'm willing to bet that a SINGLE 28 core Ice lake CPU costs more than the single Eypc 64 core CPU.

This is why, in my personal opinion, comparisons between hardware should always account for price first and formost. Nothing exists in a vacuum, and hardware doesn't exist in a world free from the realities of cost, and this metric is especially important in the enterprise market. Besides, the comparison I'm more interested in is Ice lake vs Epyc Zen3 (Milan?, I believe) as that's what the true competition will be based upon for the next year.

 

[bug]

Extremely relevant, because that's what these CPUs usually run: Geekbench.

We already know Golden Cove has potential. Until we don't have an actual incarnation (i.e. not a development part) we can't predict where it will end up.

 

[Cyler]

Without being 100% sure, isn't geekbench also counts avx512 performance? If so, that might explain the drop in GHz and the disparity of the tests, keeping also in mind what Linus Torvalds said about avx.

 

[W1zzard]

should always account for

but we don't have pricing

Intel are masters in maximizing revenue, so if they can sell half the CPUs at twice the price, they'll happily do it, (and I guess they should if I gave them my money by being a shareholder?). If people buy more EPYCs then Intel will lower their prices

 

[Tomorrow]

Ice Lake will be a failiure. Low clocked 10nm that is soon to be eclipsed by 3rd gen EPYC Milan anyway even if it manages to briefly overcore Rome in some aspects of performance (most likely AVX512).

 

[Aelussa]

Correction in the article: Ice Lake is Sunny Cove, not Golden Cove. Golden Cove is what's going to be Alder Lake.

 

[bug]

Ice Lake will be a failiure. Low clocked 10nm that is soon to be eclipsed by 3rd gen EPYC Milan anyway even if it manages to briefly overcore Rome in some aspects of performance (most likely AVX512).

Ice Lake will be good (in one form or another, possibly after it moves to 7nm). However Ice Lake is currently a failure. Because you can't get it anywhere.

 

[Tomorrow]

Well then it won't be Ice Lake anymore. The successor should be Sapphire Rapids if im not mistaken.

 

[bug]

Well then it won't be Ice Lake anymore. The successor should be Sapphire Rapids if im not mistaken.

It's not a new architecture. Sapphire Rapids is to Ice Lake like Zen2 is to Zen. Basically the same thing.

 

[Renald]

How come nobody has mentioned that the dual Ice lake setup is probably more than DOUBLE the cost of the single CPU Epyc Rome Setup. Although we don't know pricing yet, this is Intel we're talking about, and based on literally everything they've done up to this point, I think it'd be safe to assume that two Ice Lake 28 core CPUs are substantially more than a single Eypc 64 core, heck I'm willing to bet that a SINGLE 28 core Ice lake CPU costs more than the single Eypc 64 core CPU.

This is why, in my personal opinion, comparisons between hardware should always account for price first and formost. Nothing exists in a vacuum, and hardware doesn't exist in a world free from the realities of cost, and this metric is especially important in the enterprise market. Besides, the comparison I'm more interested in is Ice lake vs Epyc Zen3 (Milan?, I believe) as that's what the true competition will be based upon for the next year.

How come nobody has mentioned that the daul Ice lake setup is probably more than DOUBLE the cost of the single CPU Epyc Rome Setup. Although we don't know pricing yet, this is Intel we're talking about, and based on literally everything they've done up to this point, I think it'd be safe to assume that two Ice Lake 28 core CPUs are substantially more than a single Eypc 64 core, heck I'm willing to bet that a SINGLE 28 core Ice lake CPU costs more than the single Eypc 64 core CPU.

This is why, in my personal opinion, comparisons between hardware should always account for price first and formost. Nothing exists in a vacuum, and hardware doesn't exist in a world free from the realities of cost, and this metric is especially important in the enterprise market. Besides, the comparison I'm more interested in is Ice lake vs Epyc Zen3 (Milan?, I believe) as that's what the true competition will be based upon for the next year.

you can also add that both configuration are totally different :
windows 10 vs Windows Server 2019
single socket vs dual socket
128GB RAM vs 512GB RAM

It's clearly not fair in the configuration.
So it's not a comparison of two CPU but two systems. And one is hell lot more costly than the other (like 512GB RAM ECC+dual socket MB, not even counting CPU...)

 

[Tomorrow]

It's not a new architecture. Sapphire Rapids is to Ice Lake like Zen2 is to Zen. Basically the same thing.

Not an entirely new one no. But it should use newer arch that has better IPC. Ofcourse it will be held back by the lackluster 10nm node.

 

[efikkan]

I wouldn't put any faith in scores from Geekbench, the only interesting fact is the fact that it's being tested.

We've seen impressive scores of Ice Lake-SP from back in February, but the benchmarks could be reading the clocks wrong.

Ice Lake will be good (in one form or another, possibly after it moves to 7nm). However Ice Lake is currently a failure. Because you can't get it anywhere.

Well, unless you're only talking about Ice Lake-SP, Ice Lake-U has had good availability since early this year.

It's not a new architecture. Sapphire Rapids is to Ice Lake like Zen2 is to Zen. Basically the same thing.

Sapphire Rapids is a new architecture on a new server platform, not a refresh of Sunny Cove.

 

[TheoneandonlyMrK]

@W1zzard please can you ban news articles about Geekbench? It is quite simply the most worthless so-called "benchmark" to ever exist. It was designed for ARM smartphones FFS, why do people think it will be any use for gauging x86 or x64 performance?

Intel's use of it Will come back to bite them on the ass.
It's irrelevant yet it's arm's, multi core X1 arm chips are on the way I would imagine, the likes of Which could make Intel shudder and possibly AMD

TBF Fujitsu changed my mind on some possibilities with their new A64FX chip.

 

[InVasMani]

If that Ice Lake really was running at 1.5 GHz, vs a 3.4 GHz EPYC, with 15% more cores, then it's quite a huge achievement

The Intel chip is 3.19GHz boost and that's a real insanely high frequency to begin with a lot of cores, but not a severely high clock rate so power is probably moderately more tolerable and manageable as a whole and able to sustain that boost across all or most cores decently.

So a 23% or so higher single core score for amd and 8 more physical cores in total yet Intel apparently beats the multicore score by quite a margin, seems legit..

That was sort of my first knee jerk reaction reading it as well, but then I got thinking these could be a bit like the other new hybrid chips bigLITTLE chips Intel made. It's hard to say for certain, but if they are indeed like those a bit with a reworking of hyperthreading into more of a per-core performance super-threading take on HT as a whole the actual combined multi-core performance could end up better perhaps perhaps even at a slightly lower clock frequency especially if they've done some other IPC improvements outside that. Also if these chips have higher scale memory compatibility support and/or a broader memory channel as well that account for a fair amount of difference too especially in the multi-core scenario where it would come into play to a much larger degree.