Ryzen vs i7 (Mainstream); Threadripper vs i9 (HEDT); X299 vs X399/TRX40; Xeon vs Epyc

wait what?

 

The chiller usually costs more than the individual cuts of Beef inside

 

Plus, the cooler will be used beyond the next CPU. It has 1650W cooling capability, meaning it can handle the hottest beast of a CPU AND multiple GPUs AND the VRM keeping nice and chilly AND anything else I want. Thinking of getting a full cover block, if they are available, to cover VRM, chipset, and the CPU, plus dual GPUs to start (maybe more later). That takes a beast! So, instead of buying multiple 4x120 or 3x140 blocks (my case can handle 3 of either of those), I can just buy the chiller and CPU for about the price of an Intel 16 core, then it is just the water blocks, dropping the cost of coverage while increasing performance. You amortize those costs over the lifespan of the cooler, 5-10 years, depending, while likely pushing it less than 90% cooling capacity under constant strain (we know the heat on Intel when pulling 400W+, estimating 550+ on the 16 core, AMD will be at least 250W+ under full, then you have the heat of the GPUs pushed so far) to help keep it good (and set it above the condensation point for the relative humidity, but still insulate the lines and properly insulate the board to prevent a condensation incident which would be disastrous). I'll be benching at the cooler set around 5C, so it is the same setup you have to have for sub-zero, but....

Intel 16-core = $1700
AMD 16-core= $850
Chiller = $900

But I explained the full rationale already, so...

Edit: even thinking of trying fujipoly for this material like a thermal pad, but puddy, and using that instead of eraser for the areas where it can make contact with the cold plates, thereby providing a little extra for any components covered by a full block. Not sure yet, though.

 

This is an awesome video for those wanting to get a general idea of behind the scenes of TR and Epyc compared to Intels take on how to get things done.

 

http://hexus.net/tech/news/cpu/107725-reconfigurable-jenga-cpu-cache-memory-system-proposed/
Interesting concept. If someone wants to dig in and see if usable on existing hardware and if this is a potential for software based optimizations which could benefit both Intel and AMD (seems like it is possible for changing the scheduler), I'd be appreciative. Don't have time to dig into the PDF this morning.

 

Thanks TANWare. This is only relevant for historical value today though (unless you're actually running Broadwell EP).

I can't wait for the Skylake SP comparison (soon; as promised in the video).

 

Well it releases today or tomorrow, so we'll know more over the week!

Sent from my SM-G900P using Tapatalk

 

As far as Intel's side yes but the TR Epyc side will remain as this video shows. It explains allot on all those PCI-e lanes etc.. And now we know that getting that high speed ram enhances just about everything with the infinity fabric and why and how.

 

(Emphasis mine).

Isn't that just another way to say 'algorithm'?

This does look promising for a server type (i.e. 24/7/365 85%+ utilization) workflow. At 100ms 'refresh' rates, I'm not convinced it would be much more useful than what the CPU can do internally (and better, I'd guess) for normal consumer/workstation type workflows. After all; the software still needs to be run on top of everything else too...

When looking at some of the design aspects of Jenga; can't say it is suitable for any type of 'normal' computing most on this forum would be interested in (1GB Cache w/36C... for example).

I quickly read/skimmed the PDF twice - but didn't find anything definitive on the 'hardware requirements'. My understanding now is that this will require a change in CPU's (at least their cache architecture ) for Jenga to be effective in any real world scenario.

This does look interesting and will definitely be the future. But in the end, it is simply an extension of previous work on the subject. Small steps=Big gains.

 

Yeah; the video did bring all the different nuances together in a single presentation for us. But the details; I had already parsed from previous reading(s) of info presented so far.

Still want to know what the hit will be for IF for off die accesses...

 

That's how I read the article, but was looking at how this is what was being discussed when Ryzen came out and the inter-CCX latency was discovered. This is basically a cache awareness scheduler, whose latency will improve with time, with a properly tuned predictive branching algorithm to be implemented in software (scheduler) pre-hardware optimization and control. If viewed in that manner, I could see improvements for Ryzen's system, with better utilization of the victim L3 cache, as well as benefits for both AMD and Intel on utilization of the 512KB and 1MB L2 caches moving forward, before the hardware is designed to better utilize it. Wanted to explain where the middle might be before hardware designed to fully utilize it.

Sent from my SM-G900P using Tapatalk

 

I can't see a software option making hardware better than it is (inherently), overall. For certain workloads; yeah - careful tuning can get those results, of course; for almost any specific/narrow workload. But to me a CPU is a general purpose type system; I want it good/great in all my (varied) workloads. Hence why I hinted that server type workloads would be better optimized for/with Jenga implementation in mind.

If I'm correct that cache architectures need to be redesigned at the chip level to work properly with Jenga; we (as consumers/workstation class users) may never see it in the next 20+ years.

If it is simply a software solution that can be run on any CPU/Cache architecture 'as-is' - where can we download the demo?

 

I don't think it would be used as is, but I do think it could be adapted. We haven't seen huge advances in schedulers in a decade or two. And I do agree, once on the CPU, let it do what it was designed to do. But every program still gets put through software optimizations and a scheduler. They found out chipping it into smaller chunks and changing the way it sent info to Ryzen effected the L3 refresh and allowed better utilization of L2. We've already found a couple instances where software needs optimized for Intel's new L2 size.

So what I'm suggesting is M$ and compiler companies, including Intel, contact and work with them to do a modified version made for a cache aware implementation of a scheduler to better utilize the hardware. Will we see the same performance jump? Probably not. But if even 5% can be gained and it creates a better scheduler moving forward that is more adaptive to the new architectures in the works, I'm all for it. We have to remember, Intel is already working on the icore replacement (2021-22 expected time frame, potentially).

Does that make more sense? Talking about modifying a new tool to be used for a similar task while waiting for other changes to better utilize the tool as originally envisioned.

Sent from my SM-G900P using Tapatalk

 

I can see MS implementing this as you envision...

5% jump? Yeah!!! Could we get it for the fall 'creators update' please, MS?

If this (Jenga) can be coded into a few short and sweet lines of machine code and put into the newest Intel icore's and AMD zen's of the future; we're golden. Keeping this as 'software' though seems very limiting to me, long term.

Software allows us to build, flesh out and perfect a new 'idea'/'process'. Making it run 'invisibly', 'optimally' (via machine code) and in hardware is the pinnacle that should be strived to be reached. Least power, fastest and invisible to the O/S it is run on...

Yeah; we want the same thing(s) - where performance is concerned. Above is just my take of achieving it, 'optimally'.

 

I agree. But, another research group just created a 3D chip with RRAM integrated on top of the CPU. Scheduling is but the start. This can be used for HBM on die and for integrated memory types on the CPU in the future, going to those as another layer of cache, while further utilizing it on traditional types of cache, all based on hardware in the future, like you suggest. So I definitely agree with where you are going with this long term. I just also really like the idea of utilizing it short term, and a scheduler can help to flesh it out in the meanwhile before we get the hardware to more fully utilize Jenga as envisioned.

edit: http://hexus.net/tech/news/cpu/107659-scientists-build-3d-computer-chip-fuses-logic-memory/
https://phys.org/news/2017-01-scientists-memory-chips-processors-tasks.html

 

Thanks for the links! (I think I've already read one previously...).

The research is great (and it reminds me a lot of Optane in certain aspects...) but my watch has more computational power than the 2M transistor and 1MB RAM of the prototype in that first link.

Will be interesting to see where it goes...

I also like the conflicting information though about how compatible it is with current tech:

While reading about future tech is enjoyable, the present is where I work/live and breathe in. If/when these advances come to us; I'll process and incorporate them into my everyday life as needed (then).

 

I think the conflation between the two's statements is one is thinking about size and what is possible today and the other is likely talking about the planned 1TB RRAM planned for the near future. I'm taking it putting the full RRAM stack would act to insulate heat (same reason we don't have full 3D memory yet, heat dispersion) which would destroy the underlying silicon transistors. But I do agree, incongruous!

I also love analyzing data and predicting patterns for adoption timelines. I don't see RRAM picking up speed yet, considering it still isn't ready for commercial production from the papers in 2014 on it having 1TB on the size of a postage stamp. But, at 5nm, we're looking at gallium nitride based transistors, so with the change might come attacking. But, by 2024-25, they are trying to have out light based ram, which if achieved, and the right transposers and transceivers are integrated on MBs, we'd see another large change, potentially needing something like nano-vacuum tubes or grapheen based near THz to push data. That jump is possible in 10-20 years.

Sent from my SM-G900P using Tapatalk

 

Ln2, baby;
https://hothardware.com/news/intel-core-i9-7900x-skylake-x-overclocks-to-6ghz

 

we need a closed loop LN cooling that wont cause condensation hahahha, always 8ghz, all day everyday

 

They have it. It is a dry chamber with access holes, the blocks connect with hoses to the pressurized dewar, the exhaust hoses vent to atmosphere, and the blocks have heaters in them to help hold the temp at the precise temp set. It is controlled with a tablet. You should check it out! $6-$10K, not including the dewar, the ln2, and computer components should get you started.

Sent from my SM-G900P using Tapatalk

 

Great! Great! (really...).

But this reminds me of NOS in 'street' cars circa 2000... Great for up to 10 seconds of 'fame'.

 

http://www.lnlcooling.com/shop
If you guys want it, you can setup an LN2 rig and just have to keep that Dewar filled! ( @Papusan - you might like this!)

Edit: the purge case makes it so that condensation isn't an issue.

Edit 2: going with the street racing analogy, then this is the Bugatti! Get a 50 Gallon Dewar.

 

i wonder how mobile that would be.. dayum

and wtf i check website all out of stock

 

With a 10 liter dewar (about 2.5 gallons), you could do a little moving. If you had a 50 gallon on wheels, good luck!

Sent from my SM-G900P using Tapatalk

 

might as well get a portable AC LOL. or pay for a custom made peltier internal cooler for internal air flow while the hot side faces outward the case with fan blowing away from it.

 

The competitive laptop overclockers, like johnksss and Mr. Fox use portable ACs. Lol. And the peltier comment made me think of this manufacturer:
http://www.digitalstorm.com/cryo-tec.asp

You can do a bit better with a single stage phase change unit, but still impressive considering they only use a triple radiator to cool (and can use a double and cool the hot side).



Sent from my SM-G900P using Tapatalk

 

Intel's Answer To AMD Epyc!

AMD EPYC 7601 vs Intel Xeons in Cinebench R15

 

TBH though we need to compare the 7601 against newer Xeon's to be fair.

 

intel had a peltier laptop cooler type back in the days too bad we never get to see it go public. i'd have grab one right off the bat

 

looks like skylake new xeon poewr consumption is off the chart as well. broadwell was the good gen donno what intel did there to increase few % in IPC but power consumption jumps like 30-40%+

 

See:
http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade

See:
http://www.tomshardware.com/reviews/intel-xeon-platinum-8176-scalable-cpu,5120.html


Spoiler: Intel still the one to beat.

 

No one won. It is too complex to compare, given how little time they had.

With that said, I still find it a joke that you pay twice, sometimes, even more for Intel's top dollar line, the price to gets only a marginal amount better performance, but hey, that is how it works.

 

Yes and no. Absolute performance (like 6%), AVX in HPC, single threaded tasks, and certain specific other use cases, you are absolutely correct.

On price point, AMD wins, as well as winning in certain specialized tasks. So it will depend on software optimizations or special software, etc., as well as tasks, but Intel is going to be hit hard. Especially looking at those energy savings!

Sent from my SM-G900P using Tapatalk

 

Is it me or is that Anandtech article also not making a point of how cheap EPYC is in relation to Intel while still offering 90% of performance and even outperforming it or holding on its own depending on the tasks at hand?

 

That is exactly what they are doing.

They specifically mentioned that 1 week is not enough to test server CPU's. Such CPU's require rigorous testing under crazy workloads, to most scenarios play out.

I enjoy their stance, it is a neutral one, until they can have a better review done.

 

Might be the whole IVR instead of FIVR

Exactly, especially considering AMD had less than 1% of the market share before EPYC. If estimated at $16B, and AMD takes 20%, they just made $3.2B. If they take more than that, well... So, AMD is set to pay back the large debt, turn a profit in a matter of years, AND put in the R&D to successfully execute 7nm. AMD isn't beating Intel, but they are landing some monster blows on Goliath, something everyone loves to see.

 

Core i7-7800X vs. 7700K, 6 or 4-Cores for Gaming?
Plus Ryzen comparisons, 7800X is more $ for worse performance than Ryzen in comparison. Direct Ryzen 1700 scores to be added to the gaming charts...will update when they show up. Pretty dismal results for the X299 & i7-7800X

 

i need 2 rigs, 8 cores or 10 cores @5ghz, and the 2nd one being 18 cores @4GHZ. missed siliconlottery now all sold out have to wait.

 

What X299 motherboard(s) are you going to use?

 

lol no.

they compared 4k cpu with top of the line 8k one

 

ASUS Rampage VI Apex VRM check - the savior of X299?
der8auer finds a "good" one

 

Yeah; from a pure (overall) performance (today and long term, on these platforms...) standard: Intel is in the lead.

AMD has the chance to compete, but right now price is the only thing it competes on. Really; a $4K CPU vs. an $8K or more CPU doesn't mean anything when the platform is in the $250K range (complete, with hardware and software and...).

Who has 'won' here is the potential buyers of such setups. And hopefully they'll continue to do so in the future (near and far) and with the trickle down effect; we may benefit too.

If/when I'm looking at platforms such as these the final cost is a consideration, but it is not a determining factor. The performance/productivity over time (i.e. the expected lifecycle of the hardware) is what I'd be much more interested in. Not in saving a few thousand dollars while simultaneously leaving performance/productivity on the table that may amount to less than 1% of the budget for this aspect of my business.

I applaud AMD for giving those that are price sensitive a choice. But if absolute/overall performance is the 'need'; Intel is who ultimately delivers at this for this round, once again.

From the AnandTech article:

From the Tomshardware article:

 

So, a couple caveats. First, this is the first time that Intel's E7 equivalent offerings could be used in a 2P rig, as many locked down the firmware before. 4P and 8P are a minority in the server sector. So this is the first time it was fair to compare the flagship Intel with the 2P offerings from AMD (and this is the first time in a long while AMD showed up ready to fight).

Second, although I agree that the HPC done by exchanges and analysts in finance pay the premium regardless, when it comes to research, many have a limited budget. This is where we implement your formula, figuring out energy use, platform density (meaning CPUs per cubic foot for the rack), etc. When that is applied, you start setting a shift. If the research only uses one type of processing, there are many areas Intel beats AMD, but for overall research varied workloads, AND truly showed up.

Third, the CCX critiques are BS for a couple reasons. A) it gives way more bandwidth and testing by Anandtech showed when you blast a server on integers so the scheduler doesn't matter, AMD wins and it is precisely the throughput of the IF that gives the 40% advantage there. B) because of cache utilization, both by AMD and Intel, chopping it into smaller sections helps both, meaning the bit about latency on the ram isn't as fair as after software optimizations, you may not see the hit shown. Finally, C) it isn't the CCX at issue, it is actually the IF. That is what connects the CCXs. So, the argument is about the scheduler and latency his when employing IF on inter-CCX communications, which is about the same latency as that seen on Intel mesh. That gets into a couple issues we've discussed before.

But, when looking at price point (Intel's $4K processor versus AMDs), they concede Intel's is slower. That means there is a large segment buying in that range that will consider AMD.

But, taking market share is a long-haul game. Also, unless test beds were used before launch, like with AMDs partners, we won't see many deployments until 2018, which is the nature of the Server market.

Also, 100W savings per rack isn't de minimus either. In a server farm, that adds up to HUGE savings. So, that will change some minds enough to go with AMD, but only after thorough testing and paying for the software changes needed to fully utilize it.

IF2 may be ready by 2019. You also have the server chips at 7nm that will increase speed and performance a fair amount while utilizing the same racks (server market won't have much use with DDR5 until price dips, etc. In many cases, although some will adopt it as doubling the speed helps, moreso with AMD than Intel due to IF).

Sent from my SM-G900P using Tapatalk

 

not sure but prob ASUS. heard they got some of the best boards and bios. well i still gotta figure out how pcie lane split works. i need 8x8x8x4 if 8 cores, 10 cores and above isnt a problem.

 

Tom's article sounds like it ws done by Intel themselves. Anand, well I doubt it could have been more biased. Use some large databases that fit in RAM as a first step.

 

They both are biased, but at least Anand had hedging language, admitted at price point AMD wins, and regardless of their interpretation, the numbers show what they show. Anyone that understands the tests used, compiler, etc. can look beyond what they said and see the value directly. Also, purchasers don't rely solely on these reviews. They look at the numbers, then order a test bed. After reading internally, having longer than one week access, and performing their own optimizations, they make the purchase decision for replacement of a section or the upgrade of the floor, which is many, MANY racks. So, the presentation means nothing, nor their recommendations, when the numbers show clear uses where AMD makes more sense, and the energy savings are incredible! Also, after weighting their disclaimers, etc., it really shows that they used scenarios in certain tests to show Intel in the best light. Such as the compiler for Intel.

Sent from my SM-G900P using Tapatalk

 

4GHz boost, 3062 cinebench 15 score.
1920X=$800, 1950X=$1000
@TANWare, @hmscott, @Papusan, @jaug1337, @rageset, @tilleroftheearth, @ole!!!, @ghegde, @triturbo

It's official!!!

Sent from my SM-G900P using Tapatalk

 

AMD is not taking Intel's attacks lightly.

This means war! - on another note, 4GHz on the Threadripper is insane, and the pricing is out of this world...

 

Found something interesting as one site mentioned the 8160 becoming the 7980x;
https://imgur.com/hys8oGg

 

Wouldn't surprise me. Do that and add a little on clocks...

Sent from my SM-G900P using Tapatalk