EPYC 7642/7742 vs. Xeon Platinum 8280 Performance With Intel-Recommended Benchmarks

[phoronix]

Phoronix: EPYC 7642/7742 vs. Xeon Platinun 8280 Performance With Intel-Recommended Benchmarks

Here are the latest numbers in our ongoing AMD EPYC 7002 "Rome" series benchmarking. This time around for some curiosity over the weekend is a look at the EPYC 7642 and EPYC 7742 up against the Xeon Platinum 8280 all in 2P configurations while running benchmarks publicly recommended by Intel in one of their whitepapers. Even going by Intel-recommended tests, the EPYC 7642/7742 deliver significantly better performance and cost savings over the comparable Xeon Platinum 8280.

http://www.phoronix.com/vr.php?view=28423

[ddriver]

Oh, how the mighty have fallen...

[davidbepo]

man this is salt in the wounds

[schmidtbag]

I'm curious how a single 7742 would've compared. That's more of an apples to apples comparison vs the pair of 8280s (obviously not in terms of cost). Sure, it still has more cores, but fewer memory channels.

[yeeeeman]

Nice article Michael!
On one side, it is clear that Intel is trying to sell something inferior for more money and given how many connections Intel has in the industry, that won't be a problem.
Still, it is amazing to see how much life Skylake uArch from 2014 has into it after all these years and even though they are clearly overtaken by AMD offerings, they are still good products.
I do wonder what Intel would have had now if things went well on the manufacturing side of things. We would have had Ice Lake SP in 2017 with supposedly 38 cores (that would go against EPYC 7601 and it would have beaten it completely) and in 2019 we would have had something like Sapphire Rapids which could also use chiplets, PCIe Gen 5, DDR5 built on 7nm.
So yeah, AMD was very lucky to have great products in the pipeline and have a competitor with big issues on the manufacturing side of things.

[valici]

It's honestly remarkable how AMD can compete with both Intel an Nvidia having lower RnD than both of them. And as confidence in AMD products increase it will only get better.
The only area AMD has kind of neglected is mobile which is a huge market. It needs better mobile CPU and GPU because it's loosing a big portion of the market.

[profoundWHALE]

Big oof

[c2h5oh]

I do wonder what Intel would have had now if things went well on the manufacturing side of things.

Better, but likely not better enough - let me explain:

AMDs major advantage is the modular (chiplet) design that makes the individual components of the CPU small.
Let's assume manufacturing defect density is around 0.5 / cm^2:

• yield for a single chiplet that is just over 10 x 7 mm is about 70% and you get around 560 of them from a single 300mm wafer - that's enough for 140 32-core CPUs
• yield for a monolithic 28-core Xeon die with it's massive 32 x 29 mm is about 4.5% or just 2 28-core CPUs from a 300mm wafer.

Even if Intel was on 10nm and 10nm was a straight shrink from 14nm and reduced transistor width and height by 28.6% ( (14-10)/14 = 28.6% ) assuming same defect density Intel would have a 15% yield on the 28-core CPU and just 18 of those CPUs from a 300mm wafer.

Next AMD can bin individual chiplets - it's easier to find a 8-core chiplet that is stable at low power than it is to find a 28-core chip.
The ones that don't meet server CPU requirements AMD puts in desktop parts. The ones that aren't good enough at Intel have to be either scrapped or sold as lower, cheaper SKUs.

Die yield calculator if you want to play around: https://caly-technologies.com/die-yield-calculator/

Edit: I'm aware that Intel defect density on 14nm has to be lower than 0.5/cm^2 for manufacturing to make sense - I just used the number rumoured a couple months ago for TSMC 7nm process AMD uses

[reavertm]

In before Intel recommends only avx512-heavy benchmarks for "real world" scenario tests.

[Teggs]

So, according to Intel's official guidance, EPYC processors beat the shit out of Xeons.

Got it.