HN Explorer

Intel P Core vs. E Core actual advantage?

2 points by highfrequency ↗ HN
The marketing has been that Intel's E Core line (eg Sierra Forest) has more cores with more power efficiency, whereas the P Core line (eg Granite Rapids) has much higher per-core clock speed. This would make sense, but doesn't seem to be reflected in the actual numbers. Sierra Forest 6780E has 144 cores with 2.2 GHz Base Frequency at 330W TDP [0], whereas the top of the line Granite Rapids 6980P is expected to have 128 cores at only 2 GHz Base Frequency at 500W TDP [1].

From this, it seems like Granite Rapids is being rolled out later, with fewer cores, lower clock speeds, and more power. Am I missing something? Why would anyone go for Granite Rapids over Sierra Forest? (It does look like the L3 cache is bigger for the P Cores, perhaps that is the main selling point...)

[0] https://www.phoronix.com/review/intel-xeon-6780e-6766e [1] https://www.tomshardware.com/pc-components/cpus/5-new-intel-grante-rapids-chips-discovered-with-up-to-128-cores-500w-tdp

5 comments

[ 0.24 ms ] story [ 20.2 ms ] thread
theandrewbailey
The architecture of P cores and E cores are different. A 2 GHz P core isn't necessarily slower than a 2.2 GHz E core. It's like comparing a Pentium 4 with a Core 2 using only clockspeed: it can't be done.

You point out a 144 E core Xeon and a 128 P core Xeon. The 128 core chip has a much higher TDP, so it's possible that it's way faster than the 144 core one. It's not burning electricity because it's fun: doing things faster tends to cost more power.

There's also the big question of what kind of workload you're intending to run. Some are heavily threaded, so the E cores might be faster. Others are not as threaded or are sensitive to processing latency, so the P cores might be faster.

highfrequency
Still, we need some way to compare the per-core performance if not the clock speed. I guess you’re suggesting that IPC is significantly higher in the P Cores? Is there evidence for this? Besides hyperthreading (which always seemed like mostly a gimmick), some portions of AVX-512 (which most people never use), and a bigger cache - in what concrete ways are the P Cores much faster?
theandrewbailey
The only way to compare them is to run benchmarks on them and compare the results. This would also pick up on other things that can impact performance, like memory, I/O, and other software optimizations.

Yes, I'm suggesting that p-cores have more IPC. Each e-core occupies a much smaller area on the CPU die and uses far fewer transistors, and are claimed to have Skylake (~10 year old CPU architecture) levels of IPC. There are some very real drawbacks to build a core that uses less power than others.

Also: https://www.techpowerup.com/review/intel-core-i9-12900k-e-co...

Throwawayhahzoh
You also need to look at cache real estate. Granite Rapids has 16KB more L1 data cache per core, and the instruction cache is a large 16 way, I could not find that for the Sierra Forest L1 I cache.

4 Sierra Forest cores form a cluster that shares 4MB of L2 cache, each Granite Rapids core has its own 2MB L2 cache.

Sierra Forest has 3MB L3 cache per cluster, for a total of 108MB for the 144 core parts which is a smaller total than the L2 caches. No firm word on L3 shared caches for Granite Rapids, but its predecessor Emerald Rapids has 5MB per core. At that level, the 6980P would have a total of 640MB, but rumor at the beginning of the year is saying "up to 480MB."

So much more space is being dedicated to Granite Rapids caches, and the end of Dennard scaling a couple of decades ago means they'll draw plenty of power at rest. As theandrewbailey says, you really have to benchmark what you plan to run on these CPUs, but many loads will benefit from the bigger cache hierarchy even if some of these bigger ones are slower as they tend to be.

brudgers
Actual advantages and disadvantages depend on actual use...what I mean is test your actual workloads and applications over your actual network. Develop meaningful metrics and measure them.

But if that seems like too much work, it probably doesn't matter and either will probably be more than fast enough. Engineering isn't theory and doesn't happen on paper. Good luck.