IPC comparison with Edram

[Justinus]

thank you, but you do not need to change anything as your results show exactly what I have said. RAM copy speed is better at my setup, read write is similar, and latency is much worse.

your L3 speeds are also better. is this due to your cpu have 2mb more of L3 cache, or Haswell/Devil have anyway higher L3 speeds than 5775C (no matter the cache size). I don't know what exactly determine that speed. if you know more about it, please tell me.

yes latency of your RAM is roughly equivalent to eDRAM on my chip, but as I already told, that latency vary between low 40 and 55ns. is this due to some benchmark problem or it really vary, don't know.

take a look at link I provided from Techpowerup forum, on page 25, there is 4790K with DDR 2800Mhz 11-13-13-31-T1. it have as expected, better speeds and latency (very low, under 40) compared to your setup. but using very fast RAM, I fear, will limit overclock potential.

I see also guys with Skylake with very fast DDR4, have very good RAM results. read speed is little lower or similar like eDRAM, write is much better, copy is similar od little better.

eDRAM still make sense if latency is about 40ns, but if it vary to 55ns than it make no sense, if you ask me

I have here on paper results of some testings, so I will write everything I have at 4.3 speed, to compare to your 4770K.

at 4.3 core speed, it make CPU-Z single core result 2082, and average result of multi core 8570. Uncore ratio don't seem to affect CPU-Z results.

at 4.3 core / 3.3 uncore - Cinebench C15 - average 904
at 4.3 core / 4.0 uncore - Cinebench C15 - average 918 (take into account, as I have said before, Uncore ratio affect L4 speed much).

Passmark 8 test results with 4.3 core / 4.0 uncore :

12521 overall (results below are best scored, so it maybe does not match perfectly with overall score, because I made more than one tests)
20880 integer
9152 floating point math
54.2 prime
59 SSE
16202 compression
2297 encryption
903 physics
9401 sorting
2484 single core

prime and physics are very affected by Uncore ratio, others are mostly affected by raw Core speed. so probably this 5775C have better prime and physics scores compared to other processors, because of L4 presence.

I downclocked my cache to 4.0 and changed my memory timings so this should be as close as we're gonna get:

4.3 core/4.0 cache DDR3 2400 10-12-12-30

Cinebench R15:
883

Passmark:
12127 Overall
20998 Integer
9173 Floating Point
41.6 Prime Numbers
58.9 Extended Instructions
16125 Compression
2260 Encryption
772 Physics
9353 Sorting
2587 Single Threaded

These are from a single test, however. If I had more time to run multiples and post best-of scores I would, maybe this weekend.

Interesting, it would be fun to see a Broadwell chip without eDRAM compared to both of these to see if the improvements in prime numbers and physics are due to the slight architecture change or if they are from the benefit of the eDRAM itself. Interesting that there are some scores that were higher on my 4770K, probably due to the increased ram performance I would think.

 

[Concerned Citizen]

I'm looking for an IPC comparison for Intel CPUs with eDRAM and without. Especially I'm interested in the performance improvement from eDRAM in skylake compared to broadwell. eDRAM in skylake is supposed to be improved because that eDRAM cache in broadwell could only acts as a victim cache. I'm curious if there are any relevant performance differences in practice or if the victim cache is just fine.
AFAIR Zen also has a victim cache and that could be an important piece of information that would show if we should be concerned about that victim cache.

Waste of time.
Get one of those 950 M.2 ssds and call it good.
Also: Decently binned (3k+) ddr4, OC your cache and you're good to go!
When the ssd is almost as fast as the RAM,it does not matter that much.
950;There you go.

 

[Justinus]

Waste of time.
Get one of those 950 M.2 ssds and call it good.
Also: Decently binned (3k+) ddr4, OC your cache and you're good to go!
When the ssd is almost as fast as the RAM,it does not matter that much.
950;There you go.

You do understand the 950 Pro is at 2500MB/s read 1500MB/s write and your typical dual channel DDR4 setup is at 35000-40000MB/s read and write? This is more than an order of magnitude in difference. That SSD isn't even within a mile of RAM yet.

If the eDRAM was actually about 1/2 the bandwidth of the L3 cache but still maintained only 2-3X the latency of the L3 cache, it would really be something special. 75000-100000MB/s with 30ns latency could bring considerable performance improvements over typical dual channel DDR4.

 

[tex_willer]

I downclocked my cache to 4.0 and changed my memory timings so this should be as close as we're gonna get:

4.3 core/4.0 cache DDR3 2400 10-12-12-30

Cinebench R15:
883

Passmark:
12127 Overall
20998 Integer
9173 Floating Point
41.6 Prime Numbers
58.9 Extended Instructions
16125 Compression
2260 Encryption
772 Physics
9353 Sorting
2587 Single Threaded

These are from a single test, however. If I had more time to run multiples and post best-of scores I would, maybe this weekend.

Interesting, it would be fun to see a Broadwell chip without eDRAM compared to both of these to see if the improvements in prime numbers and physics are due to the slight architecture change or if they are from the benefit of the eDRAM itself. Interesting that there are some scores that were higher on my 4770K, probably due to the increased ram performance I would think.

most probably that more test will get you slightly better results. 5775C should have better IPC because it is newer, so it is strange that older CPU same clocked get better results. maybe something is bottleneck my sistem or I just didn't have luck with CPU performance (there is some difference at same CPU models with same speeds).

I'm not sure you get better results (not including Prime and Physics) because of RAM. first, you had put RAM settings exactly as mine. second, I'm not sure does RAM even matter in this test, because it should be just CPU test. you can test RAM as separated test in Passmark.

and that also should explain big difference in Prime and Physics results. because of L4 which count in CPU test. if RAM is used in that test, than L4 impact on result would be much smaller.
also another evidence that Prime and Physics results are heavily affected by L4 is mentioned before in my first posts, where I explained about reducing eDRAM speed. when you reduce eDRAM speed, Price and Physics results are MUCH worse.

absence of L4, which still obviously perform better than DDR3 2400 C10, is reason why you get lower Cinebench C15 score compared to my results.

 

[tex_willer]

You do understand the 950 Pro is at 2500MB/s read 1500MB/s write and your typical dual channel DDR4 setup is at 35000-40000MB/s read and write? This is more than an order of magnitude in difference. That SSD isn't even within a mile of RAM yet.

If the eDRAM was actually about 1/2 the bandwidth of the L3 cache but still maintained only 2-3X the latency of the L3 cache, it would really be something special. 75000-100000MB/s with 30ns latency could bring considerable performance improvements over typical dual channel DDR4.

I have that SSD, Samsung PRO 950 512gb. it really perform great. but as storage drive, not as RAM

no doubt that eDRAM, altought it is not so great, still benefit my setup. only thing which worries me and make confusion is that vary of L4 Aida64 test latency. why sometimes I get low 40ns, sometimes 55ns?

I would like if that guys in this thread, which also owns 5775C, can join us in discussion, so we can make more tests of L4 speed and latency and Uncore impact on it.

and it is still big question, why any altering of eDRAM ratio in BIOS, have such bad impact on L4 speed, in my setup.

 

[Concerned Citizen]

You do understand the 950 Pro is at 2500MB/s read 1500MB/s write and your typical dual channel DDR4 setup is at 35000-40000MB/s read and write? This is more than an order of magnitude in difference. That SSD isn't even within a mile of RAM yet.

If the eDRAM was actually about 1/2 the bandwidth of the L3 cache but still maintained only 2-3X the latency of the L3 cache, it would really be something special. 75000-100000MB/s with 30ns latency could bring considerable performance improvements over typical dual channel DDR4.

I do understand that.EDRAM ain't gettin' that.Next?