Xeon E5-2600 v3-v4 base and turbo speeds

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

joek

New Member
Mar 20, 2016
27
12
3
104
In search of a CPU with goood performance at lower active core counts, but with enough cores to spread around for a virtualized environment (and still be affordable) led me to a closer look at the turbo speeds. One thing led to another (the "search" got a bit out of hand), and I ended up with the eye chart below.

My takeaway is that base clock speeds are no longer a good indicator, or at least a sufficient indicator, when comparing CPU's. The turbo speeds and capability of these CPU's must be considered to obtain a complete picture and and tells a more interesting story. In particular, the E5-2600 v4 appears to have significantly more headroom than v3. The v4's turbo-binning is also a bit more granular than V3 (not shown in the chart).

p.s. Apologies in advance for any errors; it's a bit of a slog to collect-collate-verify information from various sources. Corrections-suggestions appreciated; thanks in advance.

Primary sources:
[1] Intel® Processor Pricing, Intel, 31-Mar-2016.
[2] Intel® Xeon® Processor E5 v3 Product Family, Processor Specification Update, Intel, 10-Feb-2016.
[3] Introducing the Intel® Xeon® Processor E5-2600 v4 Product Family, Intel, 31-Mar-2016.
[4] Intel ARK web site, Intel
Secondary sources:
[5] Turbo Boost 2.0 Frequency Bin upside by SKU, Patrick Kennedy, www.serverthehome.net, 1-Apr-2016.
[6] Intel Xeon E5 2600 v4 Broadwell-EP unmasked, Tarinder Sandhu, www.hexus.net, 31-Mar-2016.

Nominal precedence of sources for information in table (if it wasn't in the first source, try the second, if it wasn't there...)
- Technical: [2], [3], [4], [5], [6]
- Price: [1], [3], [4]

edit: Add highlights to more easily distinguish v3 and v4 models.
edit2: Correct pricing (screwed up merging data from several sources). Primary source is Intel official prices list; italics indicates source is Intel press material or ARK.
edit3: Add sources.
edit4: Add D-series. NOTE: No turbo-boost details are available; assumption is that @max active cores, the per-core frequency and total clocks is directly related to the base frequency. Not necessarily a valid assumption (as can be seen from other processors), so take with a grain of salt (until-if detailed turbo-boost information is available for these CPUs).


e5-2500-chart4.PNG
 
Last edited:

unwind-protect

Active Member
Mar 7, 2016
415
156
43
Boston
I always found that one of the biggest differences between the generations of i7s is how flexible the turbo mode is and how fast it kicks in.

The step was particularly notable when running things like shellscripts on Sandy Bridge and then Ivy Bridge. Ivy bridge runs away big time if turbo is on, even for same base and same turbo multiplier. The scripts are so fuzzy in how they take CPU, never contiguously, but they are in the end CPU limited. So just keeping going in turbo. Ivy did that a lot better.
 

joek

New Member
Mar 20, 2016
27
12
3
104
@unwind-protect -- Yeah, that is something that started me looking closer at this. At first glance, the E5-2630-v4 with a drop of 0.2GHz base from 2.4GHz (v3) to 2.2Ghz (v4) was a bit off-putting, even with two more cores. Pretty much the same with the 2630L, and that 1.8GHz base for the 2630L was a bit scary.

But if you look at what those CPU's do when they are actually working--with only a few cores or all cores loaded--the picture changes. The v4 significantly exceeds the v3 across the board at the same TDP and with more cores. I'll freely admit that all my investigations have been paper studies--which don't necessarily translate into real-world results--but IHMO there is a lot of unexplored territory here, and the v4 turbo improvements have not received the attention they deserve.

For someone like me--who wants desktop-class performance when using low-core-capable apps or good single-core performance, but who still wants plenty of cores to spread around for virtualization/consolidation when needed--the 2630-v4 or 2630L-v4 looks great, and a great step up from v3. (And at $600-700 is within reason.) AFAICT, to get the same performance with a v3 CPU would about double the price in the low-to-mid-range.

At 55W TDP with +2 cores and significantly better turbo, better expansion, mobo selection, etc... the 2630L-v4 makes me think thrice about a D-1541 (which I had my heart set on a few weeks ago). Unfortunately, still no details on the D-series turbo capabilities, so it still might surprise us. (edit: but it appears to have a 2.7GHz turbo cap, which likely puts it well below the 2630L-v4.)

edit: Hint/request... If anyone has D-series turbo-bin details, I would love to incorporate them into the table.
 
Last edited:

joek

New Member
Mar 20, 2016
27
12
3
104
In my continuing Academic-Eye-Chart-Studies-Probably-of-Limited-Real-World-Relevance-or-Interest-But the-Best-I-Have-to-Work-With-Because-I-Can't-Afford-to-Buy-and-Test-All-These-Systems (TM)...

I added Xeon-D to the chart. However, one major caveat (as noted in the edits): No detailed turbo-boost/bin information is publically available for Xeon-D. The chart/ranking assumes that the maximum per-core frequency is the base frequency when all cores are active. That is not necessarily a valid assumption.

The dangers of comparisons based on a simplistic base frequency X cores calculation are illustrated by other E5-2600 CPU's. Specifically, note the "GHz per active cores" and "@max cores" frequency vs. the base frequency for various processors in the chart. There is often a significant (200-800MHz) variation vs. the base frequency. Which IMHO says that using base frequency when comparing these processors is a canard, may lead to erroneous conclusions, and should be avoided. It understates what these processors are capable of when asked to do actual work.

That said, the 2630L v4 increasingly appears to stand out as a star in this lineup--especially for those interested in a performant CPU which can push >3GHz with <5 active cores, but can still push 2.5GHz with all (10) cores active--and at a price accessible to the home builder. It holds its own against the Xeon-D, but if you need/want the integrated 10GBe offered by the Xeon-D, it's a closer race.
 
Last edited:

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
Also - an important (or hugely important) factor in this is that all of those turbo speeds go out the window with AVX execution. You also do not always get full turbo on every core either so there are other workload TDP/ power considerations that go into this. Changing clocks still does have a negative impact on performance.
 

joek

New Member
Mar 20, 2016
27
12
3
104
Also - an important (or hugely important) factor in this is that all of those turbo speeds go out the window with AVX execution. You also do not always get full turbo on every core either so there are other workload TDP/ power considerations that go into this. Changing clocks still does have a negative impact on performance.
Good point. I started looking at AVX but unfortunately have no information for v4 and D. All I can find is that with v4 AVX turbo no longer penalizes all cores--only those which are using AVX (not sure about D). However, for most v3 models, that penalty is nominal if all cores are loaded.

The chart below shows the AVX penalty when all cores are heavily loaded for various v3 CPU's. In short, if you are running a heavy mixed AVX/non-AVX workload on most v3 CPU's, the penalty is low-to-zero (the 2603 and 2698 being outliers); if you are running lightly loaded mixed AVX/non-AVX workloads, the non-AVX workload is going to suffer disproportionately.

(Edit: That lightly mixed workload penalty is, as Patrick suggests, more than this chart indicates. The penalty will depend on workload. I can run numbers for specific workloads to determine the AVX penalty if anyone is interested.)

If you have information on the v4 or D AVX turbo-bins that would be much appreciated, as we could then do a better comparison.

edit: clarify this is E5-26xx-v3 only; add data bar detail.

e5-2600-v3-avx-penalty1.PNG
 
Last edited:
  • Like
Reactions: Ramos

joek

New Member
Mar 20, 2016
27
12
3
104
Doing a bit more digging on AVX turbo...

Looks like the worst thing you can do is intermittently mix AVX/non-AVX workloads across many cores. For v3 you are going to pay an AVX penalty across all cores; for v4 you are going to pay the penalty only for those cores using AVX. In any case, looks likes the worst mix is when you have an intermittent AVX/non-AVX mix, as AVX forces a down-clock for up to 1ms (for the entire package/cores for v3, for a single core for v4).

In short, if you are making heavy use of AVX on a v4 CPU--and don't want to penalize non-AVX workloads--you want to isolate their use to as few cores as possible. Regardless of whether you are using v3 or v4 CPU's, you want to group or bundle AVX use so you incur that 1ms down-clock penalty as infrequently as possible (and once you've incurred it, make the most use of it).

AFACIT, that means: If at all possible, don't do a few AVX instructions followed by a few non-AVX instructions; do all your AVX-intensive work in as large a batch as possible, followed by whatever else (e.g., IO). Hope that makes sense. Then again, this is all largely academic-paper-based, and I have no workloads which are AVX-intensive (or need for such), but would be interested in any data. Thanks in advance.
 

RolloZ170

Well-Known Member
Apr 24, 2016
5,159
1,549
113
View attachment 2045[/QUOTE]

I see a fault in the 2630Lv4 Turbo data
the turbo is 2,9 Ghz for 2 cores not 3,4
i think its a missintertretation or mixing of the AVX turbo bins/base frq.
AVX base is 1,3
nonAVX base is 1,8
AVXturbo 2core +16 1,3 + 16 = 2,9 (1,8 +16 = 3,4 <- fault)
Look here:
Intel-Broadwell-EP-Xeon-E5-2600-V4_AVX.png Intel-Broadwell-EP-Xeon-E5-2600-V4_Non_AVX (1).gif
 
Last edited:
  • Like
Reactions: JanCerny and joek

Matthew Copeland

New Member
Mar 23, 2016
7
0
1
52
Thank you, thank you, thank you!

I am building a workstation and I was using the Turbo Boost data published for the E5-2600 v3 chips as a dirty estimate.

I want to be able to use the workstation for high end gaming. The v3 data for the 2630 and 2640 were right on the edge for me as far as clock speed. With the significant clock reduction in v4, I had decided against them. v4 also downgraded the 2623 clock, and I was convinced that I was going to have to get the 2637.

Now I will be spending 400 dollars less, getting 12 more cores (dual processors), a much better multi-threaded application capability, and a respectable 3.5 GHz when i bring the workstation into a gaming state.

I can't tell you how much time this weekend I have spent analyzing this. It really should not be this hard. I am fortunate to have found this post.

Thanks Again!
 

RolloZ170

Well-Known Member
Apr 24, 2016
5,159
1,549
113
Thank you, thank you, thank you!

I am building a workstation and I was using the Turbo Boost data published for the E5-2600 v3 chips as a dirty estimate.

I want to be able to use the workstation for high end gaming. The v3 data for the 2630 and 2640 were right on the edge for me as far as clock speed. With the significant clock reduction in v4, I had decided against them. v4 also downgraded the 2623 clock, and I was convinced that I was going to have to get the 2637.

Now I will be spending 400 dollars less, getting 12 more cores (dual processors), a much better multi-threaded application capability, and a respectable 3.5 GHz when i bring the workstation into a gaming state.

I can't tell you how much time this weekend I have spent analyzing this. It really should not be this hard. I am fortunate to have found this post.

Thanks Again!
remember the v4 IPC is about 15% higher than the v3
Xeon-e5-v3-spec-turbo.gif Intel-Broadwell-EP-Xeon-E5-2600-V4_AVX.pngIntel-Broadwell-EP-Xeon-E5-2600-V4_Non_AVX.png
 

Matthew Copeland

New Member
Mar 23, 2016
7
0
1
52
remember the v4 IPC is about 15% higher than the v3
View attachment 2978 View attachment 2979View attachment 2980
The cherry on top :).

I am a little confused by your charts. The OPs chart showed the 2640 as having 3.8 GHz with 2 cores active. Your chart chows a 2.4 GHz base with a 1 GHz boost at 2 cores which would be 3.4GHz.

IS the OP showing the effective v4 clock cycle in terms of the v3 processor by using the IPC gain you refer to?
 
Last edited:

RolloZ170

Well-Known Member
Apr 24, 2016
5,159
1,549
113
The cherry on top :).

I am a little confused by your charts. The OPs chart showed the 2640 as having 3.8 GHz with 2 cores active. Your chart chows a 2.4 GHz base with a 1 GHz boost at 2 cores which would be 3.4GHz.

Which is correct?
My is from official INTEL docs.
OP one is very faulty
i.e.
E5-2630Lv4 can do 2 cores 2,9ghz NOT 3.4ghz
maybe he uses the nonAVX base clock and added the AVX bin's on that.
AVX base is 1,3ghz 2 cores +16 bins = 2,9 Ghz OK
nonAVX base 1,8ghz 2 cores +11 bins = 2,9Ghz OK
nonAVX base 1,8ghz 2 cores +16 bins = 3,4ghz <<<<<<<<<<<<< FAULTY
 

Matthew Copeland

New Member
Mar 23, 2016
7
0
1
52
My is from official INTEL docs.
OP one is very faulty
i.e.
E5-2630Lv4 can do 2 cores 2,9ghz NOT 3.4ghz
maybe he uses the nonAVX base clock and added the AVX bin's on that.
AVX base is 1,3ghz 2 cores +16 bins = 2,9 Ghz OK
nonAVX base 1,8ghz 2 cores +11 bins = 2,9Ghz OK
nonAVX base 1,8ghz 2 cores +16 bins = 3,4ghz <<<<<<<<<<<<< FAULTY
I see. Still great numbers in the official docs.

ARe those docs available on Intel's website? I can't seem to find them.
 

RolloZ170

Well-Known Member
Apr 24, 2016
5,159
1,549
113
Last edited:

trumee

Member
Jan 31, 2016
222
12
18
54
Now I will be spending 400 dollars less, getting 12 more cores (dual processors), a much better multi-threaded application capability, and a respectable 3.5 GHz when i bring the workstation into a gaming state.
What did you go with?

Also - an important (or hugely important) factor in this is that all of those turbo speeds go out the window with AVX execution. You also do not always get full turbo on every core either so there are other workload TDP/ power considerations that go into this. Changing clocks still does have a negative impact on performance.
Are you suggesting that with AVX workload the base frequency is more important than the full turbo on each core? What sort of applications use AVX instructions: Handbrake, Scientific computing?
 
Last edited:

EsbenB

New Member
Jan 9, 2017
1
0
1
40
There are some differences between the official "Non-AVX Turbo Boost 2.0 Frequency Bin upside by SKU" table and the one compiled in first post. What caught my eye is the The E5-2667v4 which is listed as 4,1 GHz, while the upside list only shows up to 3,5-3,6 GHz. Which is the authoritative source?
 

RolloZ170

Well-Known Member
Apr 24, 2016
5,159
1,549
113
There are some differences between the official "Non-AVX Turbo Boost 2.0 Frequency Bin upside by SKU" table and the one compiled in first post. What caught my eye is the The E5-2667v4 which is listed as 4,1 GHz, while the upside list only shows up to 3,5-3,6 GHz. Which is the authoritative source?
the table in the first post is mostly wrong. u have to ignore it :(
my tables are copied from Intel documents.