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

I can agree with that, to a degree. But we heard that same line about Intel a year ago, and look what a year changes. Now, unlike Intel, as you astutely pointed out, Nvidia does not have the process node and foundry troubles that Intel has. Anyone that says they do doesn't know what is going on. They are doing 1180, I believe, on 12nm, meaning they have their 7nm design in the wings, plus, even though I believe them to maybe have some difficulty on the interconnect data fabric (they have one, just it is more related to data interconnects on server nodes rather than on-PCB interconnects, not to say they do not have the knowledge, money, and expertise to create one or license one), the multi-die cards being worked on. They also could do a TSV on 12nm, which would increase the performance, but decrease the frequency, most likely.

All of what they have is fairly predictable. Instead, what I was getting at was the specs of the 1180 and lackluster increase in gaming performance of their Volta architecture. That suggests a potential stagnation of sorts. Do I think they have more than Intel? Sure, because Intel cannot move forward strongly with a **** 10nm process and horrible yields. But, when the 50% performance increase between the 1080 and 1180 primarily comes from higher bandwidth memory and increasing the die size by 27% and matching the specs on the 1080 Ti on stream processors (40% increase), all while doing a die shrink from 16nm to 12nm, there is the chance that it comes from lack of competition, but I have stronger suspicions it is a stagnation. Now, that may change with an architecture waiting in the wings (see above on the long list I have of either known or potential techs they have waiting). But, what they are doing with Volta is lackluster at best, which is why I'd rather wait for AMD at 7nm, Nvidia at 7nm, or wait for multi-die coming soon. With that said, my point is they may be moving back to more incremental, smaller advancements, as the 50% number they gave is artificial due to the memory speeds, the cuda core counts, and increasing die size which cuts against it being a large improvement over Pascal (if you throw more processors at it, it does more, same with memory bandwidth, whoop-de-doo).

Now, I could make similar comments about AMD with the 400 to 500 series, certain complaints on GCN, etc. Instead, what I am trying to say here is there is an opportunity for AMD, if super-SIMD works well enough, along with other changes that seem to be worked on, and the working with the zen team to implement certain things, for AMD to take back more market share. AMD has 7nm Vega for commercial, likely a 12nm Vega for the 600 series this fall (or further tweaks on Polaris on 12nm), and then Navi next year on 7nm. There is a chance they are gearing up for multi-die cards. With 12nm being so short-lived, if used, the TSV seems less likely than just implementing multi-die. But, it is the weakness I'm getting on Nvidia's side on analysis of what they are doing with the consumer card, based on Volta, which is weaker than I would expect, that gives an opening. People also constantly forget that Intel practically had the CPU market cornered to such a degree that any gain by AMD is huge and great, especially reaching double digits. The graphics card size, although the high end has been captured by Nvidia, still had more competition. Yes, Vega being late and not released with Polaris was a folly, with a lot of that being with HBM2 issues, etc. But that hurt them when going for that space. Now, they have a shot, but it will be a much larger slog with them.

This is where you have to balance how long until software optimizes for the use of higher core counts versus how long you expect your build to last. You can upgrade graphic cards more frequently with the rest remaining and able to perform. If you are the type to do that, then it may change the equation, but point taken.

That gets to the point I made about the life span expected for the build and upgrading graphics versus CPU, Ram, MB, etc.

Definitely agree it is revolting.

Definitely in agreement here!

It's been awhile since we've had this much competition. Really liking where it is headed.

 

AMD Needs More 7nm Capacity Than GlobalFoundries Can Provide:
https://www.extremetech.com/computi...7nm-capacity-than-globalfoundries-can-provide

So, not only will they be splitting the GPU's and CPU's between TSMC and Glofo... if I understood this article accurately, TSMC might also have to produce some Zen 2's too if Glofo can't meet increased demand?

 

I don't see how Intel could pressure AMD to improve clocks given the limitations of the GLOFO manuf. process being designed for low clocks and mobile parts... 12nmLP did lower the voltages and helped raised the bar a bit - which improved things for a Ryzen refresh... but had they went with the high performance 12nmLP from IBM AND took advantage of 15% higher chip density... Zen+ would have taken more time (likely), be a bit more expensive... but it might have also clocked much higher and overall been a better refresh.
If AMD can clock higher without breaking certain TDP limits (provided the manuf. process allows them), then they will (Intel on the other hand got sloppy intentionally and milked 4c/8th CPU's for a long time, but also had the advantage of a better manuf. process that's suited for high clocks and efficiency).

7nm should address the clocks as it will be suited for high performance (on CPU's and GPU's) and efficiency at the same time.

 

I'll dig into this more with you a little later when I get time to respond in full. Definitely something I'd like to spitball with you.

 

@hmscott - figured you want this article also, as it is Rick Merritt's piece published yesterday (he usually has really good work). I'm reading it now and will discuss more in a bit. https://www.eetimes.com/document.asp?doc_id=1333326&page_number=1

Also, you are correct @Deks , but are missing one thing of note, Ryzen 2000 is temp limited in large ways. Yes, the more dense process would have helped on the voltage management, to a degree, and hit higher clocks, likely, but due to the extra density, the heat may have impacted going higher than what they are getting from the lower process. Also, remember that LP got changed from low power to lead process moving forward, so going to be a headache keeping straight which is which with 7nm. So, due to heat concerns, and likely yields which are easier to hit on lower densities, those may have played a part in the decision on Samsung for 14nm and 12nm. Being able to start early with the IBM process for 7nm means you can work around some things, but they also are making it similar to TSMC's transistors, so I have more to digest on this issue (this has large implications beyond the CPU and GPU market, which also means that TSMC and GloFo will be increasing their capacities and working closer to capacity, along with other companies able to test and go between them, which may be part of TSMC's push to overtake Samsung, which I think took number one recently, and GloFo looking for not just them being at capacity, but being able to take overflow from TSMC once EUV goes online if the EUV variants are also doing the same thing as the current 7nm). This literally makes me need to think how this could effect my timelines (including what the 5nm process at TSMC and it being ready first, which might be why GloFo is doing the 3nm pushup and wanting to jump 5nm, which plays into whether this is a one off situation or if they are going to work to keep Samsung and Intel at bay below 5nm, which refers back to IBM's patents related to GAA (gate all around)). My head is churning through the possibilities.... I'll find the article on which fab is number 1 and number 2 in a little bit, but think that Samsung just took number one not too long ago.

Edit: OK, TSMC has been larger than Samsung since 2014. Samsung put into motion a plan to take it back, which included the EUV first, in the past year.
http://www.businesskorea.co.kr/news/articleView.html?idxno=21486
https://www.extremetech.com/computi...s-triple-foundry-market-share-just-five-years

 

Considering that we were quoted 40% perf. increase or 60% power reduction over Ryzen1 on 7nm node, I think the 7nm process is indeed designed for high performing parts and efficiency this time around.
The 12nmLP is of course Leading Process, yes, but, AMD from what I understand changed at the last moment and decided to go with a low power 12nm design (similar to 14nmLPP) as opposed to the high perfoming one based on IBM as it was initially touted.
At any rate, since we have some quotes on what the 7nm process offers, I think we can say with reasonable certainty that AMD won't be limited to low frequencies anymore because Glofo FINALLY managed to get a better performing node with 7nm based on IBM (probably why AMD avoided 12nm high perfoming process - to save costs and to prepare for7nm better).

 

40% sounds more reasonable if you go by a TR runs 3.4 GHz full core and 40% would be 4.75 GHz full core. Now at 4.0 I see CB R15 scores of about 3,500 so now we would look at scores of 4150 or so. Now low core could be 5.0 or even 5.2 GHz, that would raise gaming IPC up substantially. Again I am waiting to see the silicon.

 

lol ok so that would be 3.4 Ghz base clock compared with 4.75 Ghz BASE CLOCK. not gonna happen....taking all core boost of 3.7 GHz we should then expect a new all core boost of 5.2 Ghz? over 16/24/32 cores? well in the end the 40% boost will not only be based on clocks but also architecture, improvements in RAM, as well as core interconnect, etc....

 

If at stock you run CB R15 you will see all cores settle at 3.4 GHz. this is why I say over stock we should see 4.75 GHz.. Partial core usage of course gets higher core clocks. The new power profiles for ramping cores should keep them even higher.



Edit; of course even at stock clock today a 24 core TR is 50% more cores so up to a 50% improvement. I think they are saying the 7nm yields 40% improvement over 14nm or 60% improvement on power. Zen-2 changes may yield even further. but like 12nm (or Zen+) I do not think it will be earth shattering.

 

40% also means that say Ryzen 1600x clocked at 3.6GhZ (base across all cores) would go up to 5GhZ base... and when fully stressed across all cores, it would likely go up to 5.2-5.3GhZ
Ryzen 1700 for example would go from 3GhZ base, to about 4.2 GhZ base across all cores.
Mind you, when Ryzen 1700 is fully stressed, frequencies are at 3.2GhZ across all cores... that means, that say Ryzen 3700 (7nm) would clock at about 4.48GhZ when all cores are stressed (under same 65W TDP).

This of course doesn't say how much of a turbo clock would increase on a single core... but, if the 400 MhZ increase stays accurate for most Zen's, then, I suspect it could go up to 5.4GhZ on 3600x, and 4.8GhZ on 3700 (though, 1700 had 700 MhZ boost on a single core, so that could mean that 3700 'might' boost to 5.2 GhZ on a single core).

But, we don't know how this new silicon will ultimately behave in relation to 14nmLPP... if anything, the boosts would likely translate to what I described, or, for 8 core Zen2's, it could go higher (since 4GhZ was the 'cut off' point usually for Ryzen 1 due to 14nmLPP limitations - the same limitations wouldn't apply to 7nm).

 

huh, really? here i thought 3.4 was just the base clock. @ajc9988 what do you say?

 

This is the problem with the 1st gen boost in that other than low load it ramps back to base frequency quickly. The new boost is much better. In fact I say use all the overclock headroom up they can. I would much rather have the higher IPC under low core load but still see the max core speed under full load.

 

I haven't used this chip on stock much. In fact, I don't think I've ever seen it run that low (but I don't let it sit stock). All core boost should be 3.7GHz, so 3.4 is more of a throttle event, I believe. Then new boost is better in so many ways.

With that said, I do agree that we may see a mid- to high-4GHz clock, then boost speeds at 5GHz or so.

What peeves me is this that I came across talking about new features and controls for second gen Ryzen (Zen+):

Edit: what peeves me about it is that I wanted the features, like per-core frequency control, more control over core-disabling, etc., on my current chip. I'd also like to eventually be able to do a per core voltage tuning. TBH, I'd be happy once I can stress test, assign the top core, second core, etc., for the multi-core boost, and tune voltages for each, thereby attempting to get the lowest energy and heat for the highest return in performance. Taking the anal retentive thing even further.

I also wouldn't mind instruction set offsets with the variable frequency per core, etc. Basically, allowing the user to tune the hell out of the product. But, that is the nerd in me.

Also, non-consolidated temp monitors to allow for seeing the per core heat on these would be nice (even if having a temp sensor on two cores on one side of a CCX and one temp sensor for the other two adjacent or something like that so bad paste jobs are more recognizable or it can be seen if a core runs hotter than nearby, etc.).

 

@hmscott @jaybee83 @Deks - so, after reviewing the articles and getting some rest, this is what those articles mean to me:

AMD, due to increased market share and demand, had enough leverage with GloFo that they could force GloFo to design the transistor pitch, etc., in a similar way to TSMC's 7nm dimensions. This is abnormal because fabs like to lock in their designs and make going to another fab a bad idea because it requires an expensive redesign. So, you dance with the one you brought, so to speak. Well, with IBM's 7nm license, and getting huge ASIC demand for 7nm, all of IBM's business on their chips, and AMD's CPUs, GPUs, APUs, and embedded products, GloFo is going to be running near or at capacity. Because of this, and not being able to fully meet AMD's needs, they designed those elements to allow quick switching on AMD's part on which lines will be produced there or at TSMC (while noting that capped capacity is usually written as a part of the wafer supply agreement between GF and AMD). So, GF doesn't care so long as it is raking in the cash and it takes care of one of their largest customers. This makes the event seem more like a one off situation for a single node, as we do not know how EUV will effect these same decisions in a year. So, let's assume that it is only next year's products that will be treated like this for sure (during a major growth year due to market grab while Intel flounders with 10nm). Beyond that, they may return to designing for certain products to be made at one or the other fab, like good god fearing fabless semiconductor companies have done for generations! LMAO!

When we examine, generally, the process nodes between the two companies, GF claims a just barely denser product than TSMC, but they should be roughly the same. The question then comes what the difference in wafers, process, and other factors will have on the products in question. Most likely, they will keep it focused to lines of products coming from a specific fab rather than skus being produced at both (which would be a nightmare for the silicon lottery on trying to find the batch numbers that go with certain runs to get the best chip for your money).

But, when I saw 5GHz on HPC products, I imagine that is the boost clock. Now, EPYC has two rumored skus, starship and rome. Starship is the 48 core variant that spurred the rumors of 12 core coming to the mainstream (even I've speculated on how this could be accomplished). If a 48 core can hit 5GHz in a roughly equivalent power envelope, then things change in a hurry on the server side (with two of those creating a beast of 96 cores, 192 threads, massive L3, and massive memory bandwidth). If that can even reach that high with that core count, then, depending on core counts for HEDT and mainstream, you wind up with products that can boost to speeds or OC to levels similar to those seen by coffee lake chips. That, while at Intel's density for a node that Intel cannot achieve yields on, suggests that AMD will take back a lot of share and that demand may be very significant, especially in the server sector.

It also means that graphic cards may be made at different vendors depending on sku and potentially even with who handles certain tech (like one graphic card memory controller at one fab and another at the other fab). But, this speaks more to where AMD has already come from and is going rather than just the Fabs decisions. It also may be a time for GF to flirt with TSMC clients for future expansion. What harm is there to see what the chip made at a competitor performs like, right? That could, as some have pointed out, impact the distancing of AMD from GF over this next couple years.

Now, this also gives a better way to test the assertion that TSMC is always better than GF. That will be fun.

 

https://download.amd.com/documents/ryzen-master-quick-reference-guide.pdf
Here is the document on what is new...

 

loving those new in-depth tuning features, this is how its always shouldve been totally with you on the nerding out @ajc9988 would lead me to go back and integrating differing multis (like on intel stock cpu config) as opposed to equalizing all core frequencies in order to better tune in overclocks.

 

AMD Readies Big Hardware Updates For Computex, Powerful Fenghuang 15FF APU Spotted:

https://hothardware.com/news/amd-teases-computex-unveil

Looks like that these results indicate similar/better performance levels to desktop RX 580 and GTX 1060 (stock clocks).

Nice... though, it might have been better for AMD to clock the GPU maybe slightly lower (1100 MhZ?) instead and increase the base clocks on the CPU... that way, gpu performance wouldn't be affected drastically (likely) and it would release more headroom for the CPU instead.

And while 3GhZ on Zen architecture is nothing to sneeze at, I wonder if lower base and boost CPU frequencies would make people cry 'AMD's CPU is still too slow'.
Intel does have a bit of an advantage with turbo boosting higher on a single core... hence the disparity in Physics score... but does 3d Mark not take into account all cores, and just focuses on 1 core for Physics?
Plus, if 3dMark is inaccurately registering max. turbo clock for the CPU, it would also negatively affect the Physics score.

Then again, it also depends on the power envelope...

 

3DMart Physics test uses all available cores/threads. SystemInfo doesn't always do a great job of accurately detecting clock speeds, but it doesn't affect the score.

 

I'll be interested to see if the other hardware is a prototype 7nm GPU or EPYC chip. TR would make sense to try to take steam out of Intel's sails for the Cascade-X and Coffee/Whiskey announcements that are guaranteed to be there. Also, I expect that what Der8auer was discussing on being excited about Computex meant overclocking with LHe again this year, going all out on benches. Then again, maybe it is an AMD rather than Intel OC session, but traditionally it is Intel at this expo.

 

Ryzen 7 2700X Review - Part 4: VRM - X470 Vs B350 Vs B350 mITX
Hardware Numb3rs
Published on Jun 3, 2018
Fourth part of the review, I've tested the VRM of the Gigabyte X470 Gaming 7, AB350M Gaming 3 and AB350N Gaming WiFi (mITX)

Spoiler: Details...

EK Water Blocks Blog Article:
https://www.ekwb.com/blog/what-is-mos...

Components used:

AMD Ryzen 2700X
Gigabyte Aorus Gaming 7 X470
Gigabyte AB350N Gaming WiFi
Gigabyte AB350M Gaming 3
G.Skill Flare X 3200 C14 2x8GB
Samsung 960 EVO 512GB
Sandisk Ultra II 480GB
Seasonic Prime Gold 850W
Seasonic Platinum Fanless 520W
Seasonic S12II 420W
Full EK Custom Loop
EVGA GTX1080

ASUS Prime X470-Pro IN THE BOX(case) VRM temperatures
Actually Hardcore Overclocking
Published on Jun 3, 2018

 

But you have to admit that the difference in physics scores are 42%... which is not realistic considering that the CPU parts are both 4c/8th... Intel has 200 mhz higher base though, and it turbo boosts to 4.2GhZ on 1 core (but not on all of them - on all, the boost would be closer to the baseline... maybe 3.4GhZ).

So, what's the deal?

 

All-core turbo on the 8809G is 3.9GHz (+30%), plus a 10% IPC increase (1.3*1.1=1.43) and you get your 43% difference. Also that physics score is low for the 8809G. NotebookCheck got 12168: https://www.notebookcheck.net/Intel-Core-i7-8809G-SoC.275832.0.html

 

Isn't IPC differential closer to 5% or less when comparing Ryzen+ and latest Intel parts on same clocks?
Plus, according to 3d MArk, the system with the new APU had 6GB RAM. So, no active dual-channel (Which would affect Ryzen's Infinity Fabric).
Also, why is intel's part able to clock so high on all cores?
Nevermind... the 14nm++ process is designed for high performance and efficiency. Glofo's 12nmLP process is still designed for low clocks and mobile parts, so the iGP is taking up majority of the overall TDP likely.

But, in fairness, these would be preliminary results which the software doesn't appear to be reading properly either... and the AMD system doesn't have an SSD on it.

 

i7-8809G is Kaby Lake-G, so 14nm+ not 14nm++. 6GB isn't a possible configuration for DDR4 as there are no 2GB or 6GB SODIMMs, so that's SystemInfo reading it incorrectly.

 

AMD B450 + New 2nd Gen Threadripper Board Spotted!
Hardware Unboxed
Published on Jun 4, 2018

 

i dont see a reason why you shouldnt be able to run a 2 GB and a 4 GB stick in dual channel mode or single channel with two sticks.

 

Are there any 2GB DDR4 SODIMMs?

 

haha i actually had to check that myself because I had no idea. but i found exactly TWO offers most likely based on the same identical SODIMM stick by Crucial:

https://geizhals.de/?cat=ramddr3&v=...bl1_id=30&xf=1454_2048~5828_DDR4~5831_SO-DIMM

 

There does appear to be 2GB so-dimm stick:
https://www.amazon.co.uk/Crucial-CT2G4SFS624A-PC4-19200-SODIMM-260-Pin/dp/B01NBMUN7C

Its not uncommon for AMD APU's to be paired with worst possible RAM configurations.
Plus, if the software is reading hardware incorrectly, including the frequencies... don't know how much stock can we put into those results.

 

Threadripper 2 Equipped Board + AMD's B450!
Optimum Tech
Published on Jun 5, 2018


MSI B450 Tomahawk, B450-A Pro & X399 Creation - Computex 2018
Tech Showdown
Published on Jun 5, 2018
For my first ever Computex video I decided to check out the new MSI B450 Motherboards and the new super high end X399 Creation motherboard.


Phase-Change Cooler & Novec Cooling | Der8auer Phase-Shift Cooler
Gamers Nexus
Published on Jun 5, 2018
Der8auer's newest invention is a phase-change cooler made in combination with Caseking. The cooler uses a Novec-like fluid to achieve its cooling performance. Caseking's Computex booth featured the new Der8auer Phase Change Cooler and a 3M Novec-cooled system (using Novec 7100). The phase-change cooler uses a Novec-like fluid -- it's not actually Novec -- to cool the CPU, evaporating into a gas and condensing later. This is all at Computex 2018.

 

I really hope AMD this time can offer a lot more than 16 cores for next (coming) Ryzen Threadripper... 24 cores ain't enough. Probably not even 28.
http://forum.notebookreview.com/threads/intel’s-core-x-i9-and-i7-series-x299-xeon-1p-2p.804776/page-27#post-10740625

 

hmmmm, intel showing off such a 28 core monster can only mean that AMD indeed plans to up the core count on the TR2 chips. could mean that they would either release a 24 core TR2 based on semi-defective, but high clocking EPYC skus or theyd just bin already existing EPYCs on 12nm and clock them as high as possible with 32 cores.

BUT (and a BIG BUT(T) at that! )

can u imagine how FREAKISHLY expensive such a 28 core chip would be? lets take the jump from last gen as an example:

6950X when introduced had an MSRP of around 1700 USD. thats 170 USD per core. in comes the 7980XE at around 2000 USD, which equates to 111 USD per core. thats a price decrease of around 35%. so if intel continues this with the 28 core chip, we will be looking at around 72-73 USD per core = 2033 USD.

WELP! i didnt expect this! turns out they could actually offer this 28 core monster at the same MSRP as the current 7980XE... dayum! but would they do that? hmmm.....

 

I'm 100% sure they will charge quite stiff prices if AMD can't compete with next Ryzen Threadripper. But even without competition from AMD this one won't be cheap

 

responded in the compare thread between AMD and Intel

 

guys, u know... im actually getting sweaty hands and my heart is racing. why u ask? because i am reading computex news LOL (thats what i call being a PC hardware tech nerd )

32 core TR2, nextgen Nvidia GPUs end of july, 28 core intel monstrosity oced to 5 ghz, 32 GB sodimms with Samsung chips, nextgen Noctua D-type cpu air cooler.... OH GAWD! please someone donate me 10k so i can built a proper system this year!!!!!

 

Wait for next year if going TR3!!! This year we might get PCIe 4.0, but with the delay in seeing boards with it AND with PCIe 5.0 being published next year, you might have a little more future proof board. Plus, that 32 core count could grow. Also hoping for an AM4+ and TR4+ which supports either additional PCIe lanes or memory channels at those levels. Granted, dreams at the moment, but a man can dream!

 

haha i couldnt even afford a laptop upgrade atm, let alone a brand spanking new HEDT desktop system still, excited AF about the tech progress imagine computex without AMD having come back into the performance CPU biz! we would MAYBE have gotten the 7920X out of Intel, but thats about it....

 

I'm right there with you!! Wow!!

 

AMD Reveals Threadripper 2 : Up to 32 Cores, 250W, X399 Refresh
by Ian Cutress on June 5, 2018 11:05 PM EST
https://www.anandtech.com/show/12906/amd-reveals-threadripper-2-up-to-32-cores-250w-x399-refresh

https://www.reddit.com/r/Amd/comments/8oxal4/amd_reveals_threadripper_2_up_to_32_cores_250w/

32 Core Threadripper 2, AMD Drops the Mic on Intel
Hardware Unboxed
Published on Jun 6, 2018


AMD Ryzen Threadripper 2nd Generation First Look at Computex 2018
Engadget
Published on Jun 5, 2018


Threadripper 2 Motherboard & Fake DDR4 | Gigabyte at Computex
Gamers Nexus
Published on Jun 4, 2018
Gigabyte's headquarters at Computex 2018 was primarily focused on a new X399 Threadripper 2 motherboard, fake DDR4, and a new B450 motherboard.


Noctua Prototype Fans and Coolers | Computex 2018
Gamers Nexus
Published on Jun 5, 2018
Noctua's Computex 2018 booth was dotted with new products, including prototypes for upcoming fans, black-out D15 coolers, and a next-gen U12S.

 

AMD's Jim Anderson discusses 32-core Threadripper 2
PCWorldVideos
Published on Jun 6, 2018
AMD's Senior Vice President Jim Anderson discusses the new 32-core Threadripper 2 with Gordon Mah Ung. Threadripper 2 will be modeled on the second-generation Ryzen processor, based on 12nm, and the Zen+ architecture. It will use the same TR4 socket allowing customers to shift from one Threadripper generation to the next.


AMD announces 2nd Gen Threadripper, VEGA 7nm and EPYC 2nd Gen | Computex 2018
iGadgetPro
Published on Jun 5, 2018
AMD presents new processors 2nd Gen Threadripper with 14 nm architecture, the new GPU VEGA with 7 nm architecture, Ryzen and Epyc Gen 2 at Computex 2018 Event on June 5, 2018. All-new processors from AMD are amazing! 2nd Gen Threadripper now has 32 cores with 64 threads! That’s amazing!
This is the compact version of AMD’s press conference at Computex 2018 on June 5, 2018.

 

Full AMD Computex 2018 coverage, starts at 09:58, sound starts at 13:07. ThreadRipper coverage starts at 01:10:27...

AMD at Computex 2018
AMD
Streamed live 7 hours ago
Watch AMD’s press conference at Computex 2018 in Taipei, showcasing high performance product leadership and innovation.

 

Go to 8:20 on the second video, where he says infinity fabric is leveraged for multi-die! I think I just about finished with that statement! That got me excited. It practically guarantees what I said about Navi. Two to four Navi dies, the HBCC and shared HBM2 is really going to help get that game going more, just saying.

Edit: He referred to IF to connect GPUs. So far, IF is in PCB and AMD hasn't used a cable to connect Xfire in years, instead switching to PCIe. The die shown only has a single die with likely 32GB of HBM2 (which some reported to be running at 1.2GHz on the HBM2). Good stuff.

 

Yeah, it looks like Navi (the consumer version) might be the first using IF to connect multiple GPU dies (it also might explain why they won't be releasing it until next year)... or at the very latest, the GPU's using EUV litography (7nm+) will have a multi-GPU-die design with IF.

Vega on 7nm (the one being sampled right now for AI) still seems like a big monolithic die... and according to what was said in the article, they are expecting 1.35 times performance increase for half the power draw.
1.35 times performance increase... what does that translate to?
35% performance increase over Vega with half the power draw (150W)?

35% increase in performance is consistent when compared to 16nm TSMC.
Vega 56 and 64 weren't made on 16nm though... they were made on a 14nmLPP process not suited for high frequencies or desktop parts.
So, I'm thinking 35% performance increase is a given, yes, but to reach 150W on this new (high performance) process, I think AMD will have more clock headroom (or at least, they should).

 

So, for IF, the guy that presented for AMD said Vega 7 with IF this year. I am taking the company insider at his word given at one of the largest presentations of the year for the company. He said he wanted to introduce the Vega 7nm instinct (around 8:05), then described it, then including IF to connect GPUs as a distinguishing factor. That means specifically 7nm Vega multi-die, potentially, this year. Even though for commercial, it makes sense because this was Vega 7nm potentially used as the pipe cleaner at GF, although they may do limited runs there and TSMC for the rest, don't really know. So, also testing multi-die on the more expensive, higher margin products makes sense as well. Public perception is as important as the performance, because some wouldn't even consider AMD due to the name and how far they fell, so to speak. Navi was always going to be next year, though. EUV likely will not go into effect for Navi. Reason I say this is that EUV, due to less masking, and less patterning, can take other techniques to develop. Even with Fabs expected to pick up volume EUV between Q1 and Q3 next year, depending, I would err on the side of caution, instead saying Navi would be 7nm without EUV, just because we also do not know if the EUV transistor pitch and patterning techniques will match between GF and TSMC on the EUV 7nm, just on the 7nm without the EUV.

As to Vega 7nm, they kept transistor count. This means pure die shrink, making it the size of 242 sq. mm., I believe, which is like 20% bigger than the Ryzen die size. You can call it monolithic, but it really isn't anymore, or at least not when you cut the size in half. Nvidia, as a contrast, increased the cuda cores by 40%, matching the prior gens Ti card, and increased the die size by 27%, matching the die size of the 1080 Ti, rather than the 1080, for its 1180. That is a monolithic die.

Now, the Vega 64 is about that of the 1080, and the 1080 Ti is about 25% more powerful, with the 1180 about 10-15% over the Ti, supposedly. This means the Radeon Instinct should be better than the 1080 Ti on a chip that is around half that die's size. Now, it seems they got 50% power reduction & 35% performance increase, something I was not expecting. I know with the power reduction, it reduces heat, which can allow better clocks, and that Vega runs HOT (this is beyond dispute at this point that Vega is a furnace), but I would not have expected both. If this does bear out, I will need to revise my beliefs on AMDs ability on multi-die chips coming up. I also agree with the clock headroom. But, I need to take a moment to point out that 10+% of this IS coming from the faster HBM2. Normal HBM2 is clocked in the 800s or up to 950, whereas this HBM2, IIRC, is 1.2GHz. It is practically 30+% increase in speed, plus it has the full four stacks with 32GB and massive bandwidth. So, keep this in mind that not all of the benefit in performance is coming from the die shrink alone, as some is related to the memory used. This really makes it only around Ti levels (without the memory speed) on the core performance, while using less wattage. Not that bad, but when viewed from this angle, it makes a lot more sense for what we are seeing.

So there is some benefit here from the shrink, AMD took almost all benefit to reduce voltage, but in doing so took off the heat bottleneck, most likely, that allowed for some additional performance (if you look at TSMC's comparison of 16nm and 10nm to 7nm, I think it was, I think I said it was 35% and around 55% on energy; which, if true and correct, places about 5% on the process for performance, and 10+% on memory speed and bandwidth. If they also figured out how to better control voltages and not pump them full well beyond what they need (as seen on every recent AMD product so that more qualified for being incorporated onto PCB for the dies), then that would put a little less taken on the energy savings and the correlating increase in the transistor performance. Either way, not a bad showing. But, that is what my thoughts are on it.

 

So, both TR2 24c / 32c will draw more power than the lower core count TR1 in absolute terms, but less per core (perhaps), and Intel 28c is just a power sucking (600w+?) thermal demon.

MSI X399 MEG Board for 300W of Threadripper 2 | Computex 2018
Gamers Nexus
Published on Jun 6, 2018
The MSI X399 MEG Creation motherboard is built with a 19-phase VRM -- configured as a doubled 8-phase + 2 VSOC phases, for Threadripper 2 accommodation. Threadripper 2 CPUs will be power hungry at 250-300W, and so motherboards are being designed around for TR2 CPUs on the TR4 socket. MSI's X399 MEG Creation is built with IR 3555 power stages for a 16+3 VRM design (doubled 8-phase), using an IR35201 and IR35204 for the PWM controllers. Learn more in the video!


28 Cores of Bulls#!t - Intel's "5GHz" Parlor Trick
Gamers Nexus
Published on Jun 6, 2018
We talk about Intel's 28-core, 5GHz CPU unveil at Computex, which we found disingenuous and misleading. Intel used a Xeon processor for its "new" desktop demo, likely the 8176 Xeon CPU, then pretended that the CPU was a brand new desktop part. The company also pushed the narrative of a 5GHz overclock on 28 cores, but neglected to note that they were using a chiller to achieve these temperatures. It's an old CPU -- rebranded as a new one, at that -- and it's overclocked with an unnamed chiller that operated below -10 degrees Celsius.


Intel's Surprise 28 Core CPU - WE RAN IT!
Linus Tech Tips
Published on Jun 7, 2018


Be Quiet! Dark Rock Pro Threadripper 2 Cooler | Computex 2018
Gamers Nexus
Published on Jun 6, 2018
Be Quiet! showcased the Dark Rock Pro TR4 cooler at Computex 2018, a new cooler suitable for Threadripper 2 by way of large coldplate. Be Quiet!'s showcase this year included focus on a refreshed series of cases -- the Silent Base and Pure Base 601 and 801 cases and the Dark Base 900 rev 2. Aside from this, the showcase highlighted just the new Shadow Wings 2 fans, positioned between the Silent Wings and Pure Wings fans.

 

haha i LOVED that video of GN on the 28 core BS by intel

 

2nd-Gen Threadripper...32 CORES!?
TechLinked
Published on Jun 6, 2018


32 Core Threadripper, 7nm Gaming GPUs ARE Coming!
Gamer Meld
Published on Jun 6, 2018
AMD brings it all with a 32 core HEDT, Threadripper CPU! Plus, an update in Intel's 28 core CPU, AMD's first 7nm GPU and Vega 56 Nano! Stay tuned...


He must have missed AMD's 7nm announcement for Epyc.

FRYZEN Threadripper Cooler
Optimum Tech
Published on Jun 6, 2018


Thermoelectric Liquid AIO Cooler - Cooler Master Prototype!
Optimum Tech
Published on Jun 5, 2018

 

Asrock B450 Gaming-ITX/AC & Radeon RX Vega 56 Graphics Card
Hardware Unboxed
Published on Jun 7, 2018

 

Well this is interesting: Jen-Hsun Huang and Lisa Su are second cousins

7:52 in this video for ppl who understand Mandarin

https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=en&ie=UTF-8&u=https://www.cw.com.tw/article/article.action?id=5061946&edit-text=&act=url

 

"Jen-Hsun Huang immigrated to the United States at age 10. In high school, he discovered the magic of computers and after studying computer science and chip design at Oregon State University, where he met Lori, he moved to Silicon Valley to work at AMD and then LSI. "

Jen-Hsun Huang, NVIDIA co-founder, invests in the next generation of Stanford engineers
Jen-Hsun Huang (MS ’92 Electrical Engineering) discusses his company and breakfast habits.
By Stanford Engineering staff, October 1, 2010
https://engineering.stanford.edu/ne...er-invests-next-generation-stanford-engineers