0

I've tried searching to answer this question, but I've found it difficult to get concrete information, so thought I'd ask here. Sorry for the complex title, but the essence of my question is this: in the Xeon E5 V3 family of CPUs, the 4S SKUs often have more L3 cache than the 1S/2S SKUs; does this provide a material benefit when using 4S SKUs in single/dual CPU configuration?

For instance, consider the following 2 broadly similar chips: E5-2620 v3 and E5-4655 v3. They're both 6 core/12 thread, both have 3.2GHz max turbo clocks, identical L1 cache (6x32 KB 8-way set associative instruction/data), identical L2 cache (6x256 KB 8-way set associative), but the E5-4655 v3 has a higher base clock (2.9GHz vs 2.4GHz) and double the L3 cache (30 MB 20-way set associative shared cache vs 15 MB 20-way set associative shared cache). Excepting the base clock difference, in varied workloads (including Windows/Linux/FreeBSD virtualisation, video transcoding, databases, archive compression/extraction etc), would it be reasonable to see a difference between these CPUs in a single/dual CPU configuration due to the L3 cache difference?

I ask because I've currently got a Dell R730 server with two E5-4655 v3 CPUs, and I'm considering a CPU upgrade in the not too distant future to a single CPU (I don't really need the PCIe lanes or extra memory capacity offered by the 2nd CPU and I'd like to cut power consumption and potential performance penalties inherent in a 2S system). Prices for used 2S CPUs are typically quite a bit lower than those for 4S CPUs with similar specs (core/thread count, base/boost clocks), but I wonder what impact halving the L3 cache would have. Keep in mind that if I did consolidate 2 CPUs into 1, I wouldn't be buying the E5-2620 v3 mentioned above (that was just for illustrative purposes); I'd be getting a >= 12 core part (as yet undecided).

I understand this question may not be simple to provide a concrete answer to, but would appreciate any input or relevant benchmark data. Thanks.

2 Answers 2

0

I understand this question may not be simple to provide a concrete answer to, but would appreciate any input or relevant benchmark data.

Depends, especially since you described a varied and diverse workload. No way to know specifics short of you testing your workload. Benchmarks can define relative performance of one thing, but performance is more complex than that.

In general, of course newer microarchitecture, more cores, higher clocks, and more cache are better. But you have further constraints such as power and cost.

The Haswells you mention are several generations old now. Often that old isn't worth the power consumption, and raises concerns about the length of support left. Try one generation behind state of the art if you are looking for used or discounted.

Four socket capable system just for one socket and its cache seems dubious so long as Intel maintains a highly segmented product line with premium features. Not the most perverse incentive on CPU economics, might make sense. But its rather contorted just because more cache looks tempting, need some data to support a decision like that.

Also consider if you are willing to evaluate other vendors. AMD EPYC has a simple one socket story. For example, the 16 core EPYC 7302P has 128 MB L3, same as the 2 socket capable EPYC 7302. Not the used Xeon you asked for, but it may provide price competition. AMD definitely wants to sell single socket systems.

1
  • Yeah, I wish I could afford to move to an EPYC platform, but this is just for homelab use, so I can't justify that at the moment. Thanks for your response though.
    – amp88
    Mar 11, 2020 at 5:24
0

The premise of the question - that Xeon E5 4xxx always have more L3 cache than Xeon E5 2xxx - is not universally true. Take, as example, the Xeon 2690v3: it has 30 MB L3 cache, albeit being a 2S processor.

That said, the impact of halving the L3 cache on a Broadwell-based CPU is going to be totally dependent of the workload your application is putting on the processor: transcoding or other FPU heavy tasks will be almost unaffected, while virtualization and databases can show a 5-15% hit (at same clock/cores).

However, the key reason why 4S processors often have more L3 cache per core is that they use more complex NUMA layout, where the added cache means lower latency accessing main memory. Moreover, 4S systems can use a significant portion of the LLC (L3, in this case) for directory/snoop/coherency, "wasting" more cache than a 2S system (but this is strongly CPU and system architecture dependent).

For a future update with used parts, simply buy what will provide you the best band for the buck. If going for a large L3 CPU is expensive, simply avoid it (maybe investing in more clock/cores).

1
  • Yeah, the E5-2690 v3 is a good counterpoint. I've been considering it, but it's still a bit out of my price range for the moment (this is just for my homelab, so I don't have a real budget, as such). I don't really know much about the "directory/snoop/coherency" issues you mentioned, so I'll have to go and do some more research on those topics on my own. Thanks for the pointers.
    – amp88
    Mar 11, 2020 at 5:22

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .