News AMD EPYC Siena Has Similar Performance to Intel Sapphire Rapids at a Lower Power Draw

[Admin]

New Linux benchmarks from Phoronix reveal that AMD's latest Zen 4c equipped EPYC Siena chips can be as fast as Intel's Sapphire Rapids Xeon chips in many server workloads.

AMD EPYC Siena Has Similar Performance to Intel Sapphire Rapids at a Lower Power Draw : Read more

 

[bit_user]

Zen 4c is just as fast as Golden Cove.

Let's not be too reductive. If you dig into the results (and overlook the derivative efficiency & value metrics), there are a few decisive wins for the Xeons, such as in Tensorflow, OpenRadioss, OpenSSL/AES, and a couple of the database benchmarks. However, it's clear that the 8-channel memory configuration is what's really behind a couple of those wins, given how much worse the 6-channel Xeon did.

It's a more complex story than to just say Zen 4C is equal to Golden Cove. In single-threaded tests, it's definitely not. However, when you pack so many of them into a CPU, then power efficiency becomes a greater liability for Golden Cove and it can't stretch its legs (i.e. run at high enough frequencies to pull ahead).

 

[jthill]

So, if we overlook all of the many epyc wins golden cove pulls ahead, yes, that's the way to look at it.

Only slather a lot of words on it so people get distracted from what we're carefully whistling our way past.

 

[bit_user]

So, if we overlook all of the many epyc wins golden cove pulls ahead, yes, that's the way to look at it.

I guess you're talking to me, but I have to guess because your comment doesn't sound like you took the time to read what I wrote. I never said anything about overlooking the majority of cases where the EPYC 8324P beat the Xeon 6421N. I was trying to point out that there were a diversity of results. In some cases, the EPYC won by a large margin. In a handful of others the Xeon even managed a decent margin for its part. Of course, the majority were fairly close, with the EPYC mostly in the lead.

One more thing I'll note is that these Phoronix benchmark articles have some graphs mixed in where the performance is normalized by cost or by power. If you're not careful to look at only the graphs that actually show performance, you can come away with a slightly skewed impression of the performance picture, particularly if the cheaper & more-efficient part is also very competitive on the performance front - such as in this case!

I certainly don't deny the balance of the tests, which showed a decisive win for the EPYC 8324P ("Siena"). What bothered me was the reductive conclusion that Zen 4C is equivalent to Golden Cove. It's really not. We know this, because Golden Cove-powered CPUs typically beat Zen 4-powered ones on single-threaded and lightly-threaded benchmarks. Zen 4C has the exact same microarchitecture as Zen 4, but with half the L3 cache. Also, it doesn't clock as high. So, there's no way Zen 4C is even as fast as regular Zen 4 - and Zen 4 wasn't equivalent to Golden Cove.

Only slather a lot of words on it so people get distracted from what we're carefully whistling our way past.

Again, I have to wonder whether you're really referring to my post, because that's really not a lot of words.

CPUs are complex machines. Their performance and behavior has a lot of nuances. To reach a robust conclusion, you have to work through the data and form a model which explains the nuances. What I was trying to do is to bring people along on that journey, using words. Honestly, it's not such a complex picture, but I think one easily befitting of the analysis I afforded it.

 

[JayNor]

SPR has the tiled matrix acceleration and the possibility of HBM.
EMR launching before year end, adding cxl 2.0 memory pools.
Intel already launched SPR chips with vRAN layer 1 acceleration and extended temperature operating range.

Sierra Forest and Granite Rapids will both add MCR DIMM support and full cxl 2.0.

 

[bit_user]

SPR has the tiled matrix acceleration and the possibility of HBM.

I assume that's why even the 6-channel configuration scored wins on TensorFlow/ResNet-50.

Intel already launched SPR chips with vRAN layer 1 acceleration and extended temperature operating range.

Does Phoronix Test Suite have a benchmark correlating with vRAN performance? If not, maybe Intel should contribute one.

Sierra Forest and Granite Rapids will both add MCR DIMM support and full cxl 2.0.

That's not for at least 6 months, though.

 

[bit_user]

BTW, Aaron really should've shown the power usage graph, as well. It's pretty amazing that the 32-core/64-thread EPYC 8324P (at stock settings) averaged just 89.8 W, with a max of only 136 W. That's just 64.5% of what the Xeon Gold 6421N used @ stock settings & 8-channel memory.

 

[TerryLaze]

BTW, Aaron really should've shown the power usage graph, as well. It's pretty amazing that the 32-core/64-thread EPYC 8324P (at stock settings) averaged just 89.8 W, with a max of only 136 W. That's just 64.5% of what the Xeon Gold 6421N used @ stock settings & 8-channel memory.

If you want to take power as an consideration you would have to compare against ARM servers because that's what customers that need power efficiency go for.
Anybody still sticking with x86 in server does so because they have to, it's all arm and GPUs now.

 

[bit_user]

If you want to take power as an consideration you would have to compare against ARM servers because that's what customers that need power efficiency go for.

Fair point. I don't know much about the CSP market, but I could believe ARM gets a lot of play there.

Anybody still sticking with x86 in server does so because they have to, it's all arm and GPUs now.

In the CSP market segment targeted by Siena, I think they're more likely to use DPUs, if anything. I think Xilinx has some products in that sector.