AMD's Ryzen CPUs (Ryzen/TR/Epyc) & Vega/Polaris/Navi GPUs

Sux people would even think to hijack a video.

 

I run into them all the time, and get caught up by them too, and it's sometimes hard to find the original.

Oddly enough the monetary reason to do it has passed; youtube has set new rules to get ad revenue, yet people are still doing them.

Hopefully these hijackers will realize their youtube ad income has dried up and stop

 

Links;

I think this was done, but just in case;
http://wccftech.com/amd-epyc-7000-series-server-cpu-family-specifications-price-performance-leak/

 

Intel is scared enough to talk Cascade Lake. They say 14nm+, but prior articles discussed 14nm+ on skylake-SP already, like Kaby, meaning this would be 14nm++, while also suggesting cannonlake-sp is scrapped. Considering 10nm cannot compete with 14nm++, I'm saying more 10nm woes, so going with the strongest process to compete with AMD at 7nm next year. This article also suggests Epyc will be one of the first lines with the 7nm (like I predicted).
http://wccftech.com/intel-cascade-lake-xeon-scalable-platform-optane-dimm-2018/

Sent from my SM-G900P using Tapatalk

 

Could we please change the title to add RX4xx / RX5xx GPU's?
(Thanks for adding Polaris to the title )

We don't really have an RX 4xx/5xx thread, and they go with Ryzen best, since that's what many will consider a perfect fit right now for their AMD build, at least until Vega releases - which could be many months.

Asus ROG Strix G702ZC with Ryzen 7 1700 8-core CPU and a Radeon RX580 GPU



There are a number of the Ryzen 3/5/7 CPU's with 65w TDP that will work in a laptop. Of course the missing piece is a GPU over the RX580 in performance to compete with the 1070 / 1080, otherwise we are going to have CPU performance heavy laptops.

(Aorus) Apologizes for the RX 480 => Aorus RX 580 XTR: THE DO-OVER!


Sapphire Pulse RX 570 and RX 580


Special Q+A: Ed from Sapphire on the Pulse RX570 RX580 Graphics Cards

 

SPECint_rate_base2006 is for people comparing workloads such as may be faced by a cloud provider... and ignores almost everything else except integer performance of a CPU. It is also based off a 1997 platform...

In other words; most likely nobody here that knows what they're comparing for a home/small business setup.


Passmark tests while not perfect, do allow comparisons between platforms and if just the CPU is changed; follows the % change of the 'scores' pretty closely, ime. Again; not perfect... but usually gives a good guess of what should be expected between two systems.

Anyone running a cloud server that needs this much more SPEC performance is more than likely not running any of the (Intel) CPU's listed in that table either.

Yeah, comparing the large boost in integer performance from a 1997 platform base is interesting for about five seconds.

But when todays workstation workloads are involved, or; server workloads where efficiency is just as important as performance; not so much.

 

Links;
http://segmentnext.com/2017/06/19/amd-threadripper-cpu-pictured/

 

We've had that exact same conversation here in these forums and the article seems to have it backwards; Intel is giving info - AMD; not so much.

Either way, after all the posturing and prancing by both companies... the bottom line is what they will actually deliver.

We'll be here to judge that for ourselves when the time and the products arrive (hopefully sooner than later).

 

After reading more about the process, IBM (who designed it in the agreement with IBM, GloFo, and Samsung) focused on processing power for 7nm, then balance to power efficiency for 5nm. Samsung's process is more power efficient, which is the 14nm process AMD licensed. AMD turned down a licensing deal of Samsung's 7nm, instead licensing IBM's, which may be repeated on 5nm. So, I don't know the balance that they will choose, but it will be interesting to watch. Here is some information on the Gate-All-Around of the new 5nm.
https://arstechnica.co.uk/gadgets/2017/06/ibm-5nm-chip/
https://www-03.ibm.com/press/us/en/pressrelease/52531.wss

So, this is dependent on EUV lithography being ready and Samsung and GloFo finding a way to mass produce the GAA transistors quickly, as it leaves behind finFET and adds a 4th "gate". This means we could see a finFET 5nm first, followed by a GAA. But, IBM's 7nm was first made in 2015 and will see production in 2018. If they can figure out how to build the 5nm in roughly the same amount of time, we will likely see either the finFET or GAA in 2020-21 (or at least have more news on it by then). But considering that is 40% transistor performance or 75% energy efficiency (over 10nm), that may be the better place to take the energy savings than going to 7nm.

 

" Intel’s Frank Soqui said in an interview that these CPUs were planned from the start and these high-end CPUs were going to come out whether AMD announced Threadripper or not. Here is what he had to say:

"Independent of what our competitors do we will put out extreme edition processors, We’ll push performance, we’ll push threads, we’ll push cores, we’re pushing memory performance. This is not a response to the competition, this is a response to the end user.""

And, the "end user" wants 16 Core AMD ThreadRipper, so we responded

(maybe these posts should be in the TR thread?)

 

Links;
https://www.nextplatform.com/2017/06/19/amd-winds-one-two-compute-punch-servers/

 

I read it, but the author is biased toward Intel and obviously did not see the latency and performance issues on the X series which are likely to be present on the Xeons.

Edit: but still an article worth a look...

Sent from my SM-G900P using Tapatalk

 

My issue is it is presenting opinions rather than straight facts, but who am I to say.

Edit a Link;
https://www.overclock3d.net/news/gp...ega_silicon_to_their_aib_partners_this_week/1

 

TANWare, thanks for the chuckle (complaining about opinions, then posting a link to a 'rumor').

 

It follows suit, as I am the poster of the link to the article that is based on opinions. My point being I am not to say the relevance, so here is another. Glad you enjoyed though.

 

Ok, I am now confused. Epyc releases today so where is all the info? I would have thought someone would have had the chip but was under embargo etc. but thee is nothing out there. At least I would have hoped for some official benchmarks from AMD to compare them the Xeon offerings. Disappointing again to say the very least.

 

http://www.servers-maintenance.com/...-processors-to-the-worlds-bestselling-server/
It is also still early.

Edit:
Also there is a chance they will do a launch event.

Sent from my SM-G900P using Tapatalk

 

https://twitter.com/AMD/status/877167059448233984
https://twitter.com/AMDServer

 

AMD Bringing B2 Stepping to Ryzen Summit Ridge for Fixing Hardware Bugs
http://www.mobipicker.com/amd-bringing-b2-stepping-ryzen-summit-ridge-fixing-hardware-bugs/

"AMD disrupted the processor market earlier this year with the launch of Ryzen processors based on an entirely new architecture designed by the company over the years. While everything was going good, AMD figured out a few hardware-level issues that can’t be simply fixed by an AGESA update.

Canard PC Hardware reported that AMD is readying a new B2 stepping for the 14nm Summit Ridge eight-core CPU silicon. The stepping is expected to fix a lot of hardware issues that AMD hasn’t been able to fix via a BIOS/AGESA update. AMD has pushed a lot of AGESA updates in the past few months but a few issues are still left unresolved."

 

Now I am more concerned, no live stream and they want you to tune back after markets close for information. Same with video playback afterwards. Maybe I am just being paranoid.

 

I think you are. Think of earnings reports and similar that come out after markets close. Good or bad, it is released about the same time every time. It also is to let some get off of work to watch, etc. You do it in the middle of a weekday, you would have no one watching, even if they wanted to!

Read that, but here is why it is problematic: canard pc was talking B2 for epyc and TR. They didn't say for Ryzen. Second, Epyc and TR use way more I/O and memory, so the exact improvements or fixes claimed coincide with exactly what is advertised on those two lines. Third, they try to say the B2 stepping is Pinnacle Ridge, which is Ruzen 14nm++ refresh, that is said at its earliest to be Q1 2018. So a lot of extrapolation off of a little tweet. I'm guessing it is a regular revision that fixed some things, not how they started it in the article.

Edit: Sorry, some of that applies to a different article I read this morning (like pinnacle ridge). The rest is the same, and if you call TR ryzen, then it is technically true, although who knows if it will be used on the R7 and R5.

Sent from my SM-G900P using Tapatalk

 

A good read, thanks for posting. I can see either increased performance or increased efficiency being attractive to different audiences. Personally I think 75% efficiency increase increase sounds preferable.

 

TANWare, I'm concerned too. Is it possible that the B2 stepping hmscott brought to our attention is needed for Epyc too?

Great that they caught it before Epyc shipped (if that is the case) - but, how long to get new stock levels up to where they want/need them?

I guessed this (new platform - unknown problems - hardware fix 'required') over 6 months ago. Sad to see it turned out true. I thought that after 4 months of actual products in users hands, AMD would have ducked that bullet on their new platform...

In any case: Nobody buys on price or performance. Reliability, dependability, availability and compatibility of the platform with your current workflows/workloads is paramount (and by a large margin). If your existence depends on these tools...

In any case; if I was AMD and have Epyc stock; I'd sell them as is. Not every workflow will be affected greatly (enough) by the benefits the B2 stepping will bring (I hope), and for those that fit that model can get their hands on them as they were expecting. (i.e. I hope launch day is today for Epyc).

 

There are several scenarios I can think of, some good, some not so. One I could think of is the 1920 and 1950x being originally ready silicon that was tabled under a revision of the Ryzen core that had issues with the interconnect of the two or more CPU's (not just 2 CCX's), this would have caused a revamp of TR and Epyc and because of known issues they were never intended for release. This would explain why they started TR at 1955 and the benchmark numbers for the 1950 not being much better than a clocked 1800x.

Again PURE wild speculation!!!!!!!!!

Edit; Also this would explain why Intel only had up to a 12 core planned for release. If at Ryzen launch the 1920 and 1950x were the planned releases for TR then Intel only needed the 12 core to compete. Now if AMD pulled the two skews and planned release of these fixed skews along with a fixed Epyc, that would have thrown a wrench in the Intel plans.

 

http://www.tweaktown.com/news/56891/amds-12c-24t-16c-32t-cpus-called-threadripper/index.html
March 29th:
AMD's next-gen ThreadRipper details so far:
Zen HEDT CPU's are called Threadripper! (confirmed)
Each CPU will include 64 PCI-E Lanes! (confirmed)
It includes 4 CCX's. (confirmed)
Lower SKU(Probably 12/24) 140W TDP, Higher SKU (Probably 16/32) 180W TDP. (speculative, but likely; some have 150/60)
Socket will be an SP3 LGA (modified version of this socket, but mostly confirmed)
Platform's name will probably be X399 (confirmed)
Chips will be B2 revisions. (Confirmed)
32MB L3 Cache ES's are 3,3 or 3,4GHz base and 3,7GHz Boost (confirmed)
It is aimed for Retail SKU to have 3,6 Base/4GHz Boost (confirmed that this was the aim; expected is 3.5 base, 3.9 boost)
ES's that are in the wild have 2500 CB R15. (speculative)
Infinity Fabric can have a bandwidth up to 100GB/S (unconfirmed, but hopeful)
Announcement; COMPUTEX at Taiwan, sales will start after 2-3 weeks following COMPUTEX. (False; announcement was true, but it was EPYC going on sale 2-3 weeks after computex)

Read more: http://www.tweaktown.com/news/56891/amds-12c-24t-16c-32t-cpus-called-threadripper/index.html

So, considering B2 was known about since MARCH, we are looking at plenty of time to stack inventory (literally 2.5+ months). This is not a last minute revision, but a planned revision. Further, if the 100gbps is true, we are talking about a nice kick in bandwidth! This would make sense considering it is mentioning memory and I/O tweaks in the revision. So ignore anyone saying be worried when they don't say the same about Intel with the current crap coming from them (BIASED)! (taking it that is who you were responding to).

 

Not giving either of them a pass at this point until I get more information about the end products.

 

That is fair! I have no problem with that so long as BOTH are evenly given a critical eye! *Tip of the hat*

 

It will ewill be reluctant criticism in both cases. I honestly want both Intel and AMD to bring their top game and offer good choices, rather than one tanking and the other dominating.

 

What people don't realize is that AMD must succeed or Intel may get broken up. Period. Full stop. AMD definitely brought their a game, using a handful of talent instead of hundreds of people, thereby allowing revisions and changes to move quicker this round. They separated divisions (graphics from cpu), then instituted leap frog R&D (two teams working on subsequent designs, one after the other).

Now, we'll see how those decisions wash out in the products already available or coming soon. (Don't take the above description to mean that is preferable, but reporting decision).

Sent from my SM-G900P using Tapatalk

 

I think right before release of Ryzen7 they knew that a B2 was needed. I do not think it was last minute but if B1 were good enough that would have been 1920 and 1950x released at close to the same time as Ryzen 7 and only the 12 core would have been required to compete. That sounds about right for the 2500 R15 score.

 

Well, I read a rumor that the ES on Ryzen allowed for 1:1 DR instead of SR speed on infinity fabric. Of course, take that with a grain of salt as the Ryzen 7 released with B1 uses SR ram speed. But, if the rumor of 100gbps Infinity Fabric on TR is true, then they would have to use 1:1 DR of the ram AND banding of the channels to gear the speed in aggregate. Same with EPYC if the 176 or 178gbps rumor is true. We won't know for awhile though.

There is also the fact Ryzen uses 16 PCIe lanes whereas each die of stepping B2 uses 32 PCIe lanes (64 on TR, 128 on EPYC). Also, they were working on faster ram support up until release on Ryzen, with the major problems not being fixed until Dec/Jan according to some vendor rumors, which contributed to MB manufacturers saying they were unimpressed with Ryzen ESs. This contributed to the brush off. If you notice, the rumor about B2 is improving the I/O and memory support (which would be beyond the AGESA update). If this included getting the Infinity Fabric to run at DR, then we are looking at a very different beast than Ryzen 7. But, as noted, that is all rumor and conjecture, ranging over the past 6 or 7 months...

 

http://www.anandtech.com/show/11551...7000-series-cpus-launched-and-epyc-analysis/4

Sent from my SM-G900P using Tapatalk

 

https://arstechnica.com/information...mds-zen-steps-into-the-server-room-with-epyc/
https://hothardware.com/reviews/amd-epyc-7000-series-processor-and-platform-details
http://www.pcworld.com/article/3201...chip-to-take-on-intel-in-the-data-center.html
https://www.hardocp.com/news/2017/06/20/amd_reveals_epyc_7000_server_cpus
https://nl.hardware.info/nieuws/525...ijgt-concurrentie-in-meest-lucratieve-segment

http://www.anandtech.com/show/11551...7000-series-cpus-launched-and-epyc-analysis/4

"The connectivity here is set at a bidirectional 42.6 GB/sec per link, at around an average energy of ~2 pJ per bit (or 0.672W per link, 0.336W per die per link, totaling 4.032W for the chip). It is worth noting that Intel’s eDRAM for Broadwell was set as a 50 GB/s bidirectional link, so in essence moving off die in EPYC has a slightly slower bandwidth than Crystalwell. With a total of six links within the silicon, that provides a total of 2 terabits per second of data movement, although AMD didn’t state what the bottlenecks or latency values were."

Edit:
"Socket-to-socket communication is designed at the die level, rather than going through a singular interface. One die in each processor is linked to the same die in the other processor, meaning that for the worst-case scenario data has to make two hops to reach a core or memory controller on the other side of the system. Each link has a bidirectional 37.9 GB/s bandwidth, which is only slightly less than the intra-socket communication bandwidth, although we would expect socket-to-socket to have a slightly higher latency based on distance. AMD has not shared latency numbers at this time."

Edit 2:
"It is worth noting that the 42.6 GB/s die-to-die bandwidth is identical to the dual-channel memory bandwidth quoted per die:

http://images.anandtech.com/doci/11...or_press_and_analysts_06_19_2017-page-077.jpg

Time will tell if these become bottlenecks. Latency numbers please, I’d love to fill in that table above."

"As part of the launch today, AMD is announcing partners working with them to optimize the platform for various workloads. Sources say that this includes all the major cloud providers, as well as all the major OEMs. We saw several demo systems at the launch event with partners as well,such as HPE and Dell."

 

I don't think this belongs in this thread?

 

It is indicative of the potential performance of infinity fabric used on the RYZEN TR!

Sent from my SM-G900P using Tapatalk

 

No, no... just no.

IF is not what makes a processor or a platform great - at best; it makes it okay. But it can make it worse, as AMD has shown us so far.

 

Considering this showed directly reaching DR of ram, not SR like the R7, it speaks directly to the speed of interconnect between dies. Considering, also, that this speed is double the prior, you may see latency in ns reduced considerably for communications between CCXs. That may bring it to under Intel's mesh, but we don't yet know that for sure.

Sent from my SM-G900P using Tapatalk

 

While it may not directly reduce latency it does increase bandwidth and this may be the ticket. Also there could be other improvements that have changed. we w3ill have to see.

 

What is DR and SR?

You're still thinking that Intel and AMD's 'fabrics' are similar. They're not. For one; Intel is all (mostly?) on die; Epyc; not so much.

 

The latency per cycle should be roughly the same, slightly higher communicating between the dies. But, with faster speeds, it will reduce it on time, not on the latency relative to cycle. This was seen in the testing of faster ram with the R5 1600X. Latency was 140ns at 2133MHz ran, with infinity fabric at 1066. Latency dropped to 111ns with 3200MHz, which is IF at 1600MHz. So, it dropped about 30ns by increasing the speed by about 50%. This would double that if it is running at the same transfer rate as the DR ram (but it should be noted, this was not fully linear and there may be a plateau on reduction of latency).

Sent from my SM-G900P using Tapatalk

 

Dual rate, normally stated as the speed of the Double Density Ram. Single rate is to refer to half the rate of the ram.

Edit:Also, Intel's is very different from AMDs. I never said nor thought they were the same. Although, it should be noted that Intel's mesh did benefit to a degree on latency with faster ram speeds.

Sent from my SM-G900P using Tapatalk

 

I need to also mention about 40ns of that time is related to CCX communication, which means that is the lowest any latency can be. So with infinity fabric running at 1600mhz, we saw a reduction of 30ns out of the 100ns time above that 40ns. So don't expect huge reductions past this, as I said, a potential plateau.

Sent from my SM-G900P using Tapatalk

Edit:


After working out the math, you would, if perfectly linear, expect to see no latency at around 2800MHz, which is impossible. Now, you could expect at SR to (or IF at 1333) to see 120-122ns latency, approximately.

The first step goes down 11ns, the last one goes down 4ns, so I would estimate an 8ns and 6ns steps in between the next couple steps to fill in the 2666 (1333 IF) speed. That suggest that you see a drop that approaches a plateau. It's been awhile since I worked out the math, but you have a 20% reduction, then 25%, 33%, X%, etc. So if anyone wants to work that out for me (I'm being lazy), please do so...

 

Links;
http://www.gamersnexus.net/news-pc/2962-hw-news-threadripper-release-date-intel-7900x-hot-xbox-one-x

 

Please delete - this means please delete this post as it isn't needed, the response was posted in another thread - is that clear enough? Thanks.

 

Threadripper = Ryzen

 

Little cheeky tonight, are we?

Sent from my SM-G900P using Tapatalk

 

http://www.tweaktown.com/news/58113/amd-epyc-32c-64t-flagship-cpu-costs-4200-monster-perf/index.html
Look at that pricing for the 8 and 16 core chips!

Sent from my SM-G900P using Tapatalk

 

The reason it is not a vs. link is while both CPU's are in the link there is no comparison of the items within the link.

 

Actually, it was a standard please delete this post request that got blown out of proportion. I thought he had accidentally posted it, and was simply giving him a heads up. If he did as I expected, read it and follow the link to my reply, and then delete my post as requested, it would have been over.

Your comment probably didn't help

How about all these posts get deleted and we forget it, ok?

 

No! We must enshrine misunderstandings FOR-EV-ER!!!

Sent from my SM-G900P using Tapatalk