Dissecting AMD Zen Architecture - Interview with David Kanter

[@mentalplayground]

Tech Talk with no nonsense. Thank You so much.

[@unacomn]

It's stuff like this why I follow you folks. Heck, It's the reason I'll be sponsoring you on Patreon soon.

[@mycophilic]

Awesome seeing David Kanter! He always has so much to contribute to discussions. Get him on again, please!

[@Pholostan]

Interesting talk indeed. You need to do this again in the future :-)

[@Horus9339]

It is so good to hear a man that can explain what he knows in an easily digestible manner. Well done for getting this fellow involved. Please bring him back soon. Thank you guys.

[@DiamantisGR]

This was soooo interesting. Please do this again.

[@fireonawire]

Damn, I understood almost none of the technical stuff that was being discussed but don't you ever dumb down the content, it's what makes this channel great.

[@lelouchvibritannia4495]

at school today we watched a video of dissecting a heart i liked it but dissecting a processor?! count me in

[@peterjansen4826]

I am glad to see that Josh is also there. :)

[@YizkiM]

I'm so looking forward to this March! Please do another discussion!

[@ManuelMijaresCh]

This is why I love PCPer, nobody does content like you guys, thank you.

[@ylu5384]

Tech journalism at it's best. Thanks guys!

[@Matrix1Gamer]

Great Video! It's good to see people who get straight to the point. Most Podcast, it takes speakers forever to just talk about the subject in their descriptions.

[@mikey7326]

Come on team red.

[@ChildOL]

In-depth technical discussions on various computer tech is in short supply. Thanks for these videos

[@peterjansen4826]

Interesting conversation, I am looking forward to the post launch analysis. I bookmarked the Real World Tech forum, I will check it out later.

[@jellowiggler]

Yup, David needs to be a guest on a regular basis. Please!

[@hotroof]

This was the most interesting videos I've watched in a long time. Nice job pcper. David's knowledge and insight was amazing to listen to. Not to mention, his delivery was understandable and educational. Great interview.

[@DanielFlores-kg3tw]

Nice interview, looking foreword to more info on zen.

[@jkaustok]

This was a really great video. I love hearing about the underlying architecture and its implications. You should really have this guy on again. It made me wish I could subscribe to this channel twice.

[@jetpil0t]

Certainly if you have been a hardware enthusiast for more than 5 to 10 years, it's significantly more interesting to start taking a deep dive into processor architecture and detailed discussion, rather than simply evaluating processors based on their commercial positioning. As far as I know this is one of if not the only channel that goes this far into the technical particulars, especially since it is such a niche community that even have this level of interest, full props to PCPer for hosting these types of discussions. It might not rake in the views, but IMO it's all the commendable that PCPer feels its important to cater for a truly enthusiast community, especially on KZbin.

[@papal1ef]

Nice to see this good content.

[@helloansuman]

can David explain the block diagram of the zen and how they work wholly? that will be more beneficial. it's a kinda discussion video.

[@MrHarney]

its a miracle Josh is not blue anymore

[@zaq_d]

Thanks guys, great content as per usual. Not sure sure about everyone but I for one am very excited to see some competing high end parts back in play.

[@victor2410]

Is there going to be a follow up panel like discussion like this on Ryzen now that it has launched? Or was there already one and I missed it?

[@Walterus_]

Great quality, awesome guests!

[@Consoul909]

great idea, Josh...and Ryan actually took your advice?!??

[@joshwalrath9518]

Shocking, I know.

[@vertonical]

Great video! Although I don't quite agree with Kanter about Zen's IPC lying between Sandy Bridge and Haswell, think it will be higher than Broadwell, closer to Skylake/Kaby Lake as indicated by the very recent leaked 3D Mark scores where the 1700x outperforms 6900k on single thread performance with a slight overclock.

[@asdw1983]

Make AMD great again :o)

[@icecold19833]

What a great discussion!

[@MatthewSuffidy]

I was curious about what changed since the bulldozer that was better and how it handled problems like upscalability and heat control. It sounds like they rethought it, shank the process and added a few optimizations.

[@SilverforceX]

The funny thing regarding AVX/2 is that GPUs do it massively better, inherently GPUs are designed for those operations and there's many more compute units for vastly better perf and perf/w. This is why a lot of HPCs have moved to GPU accelerators instead of relying on racks of thousands of Xeons. So having weaker AVX2 on a CPU is not actually a disadvantage for all HPC, it's a wrong conclusion. It may be a disadvantage for some until they move to GPU clusters.

[@DAVIDGREGORYKERR]

I am running a system based on a FM2+ motherboard with a AMD A10-7800k with 16GB of GDDR3 RAM and I was wondering if I should remove the addin graphics card that someone stupidly installed and reinstate the blanking plate and remove the piece of card that is blocking the HDMI socket that connects directly to the FM2+ socket so that the APU will run cool? BTW I see no need for sound cards or graphics cards when there are perfectly good graphics and sound cards integrated into the AMD A10-XXXX series of APUs.

[@coxyofnewp]

That a bold statement saying - Apple's 2 core is the best phone hardware, Shame the don't make the software make that the truth.. As for Zen - I'm just thankful that it''s going to make prices drop all round. As Intel are really just not pushing the hardware now,as they don't need too - due to lack of competition..

[@delatroy]

coxy ofnewp going to be a good year to buy coffee lake :) finally looking like a good time to upgrade from the 980x gulftown

[@alexvillalobos7137]

coxy ofnewp Apple started making their own phone CPUs now

[@Razyre]

Since when is Apple's software bad...? It's pretty great, though I definitely think stability has taken a dive since Jobs left the building.

[@railshot888]

PCPer is also the same source that said Zen's IPC is slightly behind Haswell a few month back when the first Zen engineering sample benchmark was found on Ashes website.

[@SomethingtoappeaseGoogle-1024]

On a mobile device, you don't need a beefy cpu because all the graphically intense mobile games are fucking shit and the majority of people choose to run simple free apps anyway.

[@landwolf00]

Excellent video. Extremely informative for something I will most likely be buying within 30 days (depending on availability).

[@The_Noticer.]

Hey it's mr. morden from babylon 5. Cool.

[@BUnicutZza]

Good interview, see you next time guys :)

[@prycenewberg3976]

I'm not subscribing until I see a part 2 made after the release of Zen!! :) lol Seriously, this was good and I would love a part 2 with this guy.

[@robertpearson8546]

When computers went from sequential fetch execute to simultaneous execute and fetch next instruction the memory buss utilization went from about 50% to 95%. The maximum utilization is 100%. That means that no matter what you do inside the CPU, the best improvement you could have is a 5% improvement. What is the number for modern architectures?

[@guidocurcio6903]

Looking forward to more of this! I loved it

[@Furens9]

Thank you guys for this video. This was the most interesting tech video I have ever watched.

[@RepsUp100]

Great interview

[@PappaLitto]

All hail the nerdiest nerd

[@Great_America]

AND Ryzen actually improved IPC by 58% according to the leaked benchmarks. Not sharing a floating point calculation per two cores alone would account for the majority of this kind of performance increase. There are other enhancements as well. Needless to say, I can't wait for all those real 'hands on' hardware reviews in March!

[@kensmith5694]

The quantum transistor is digital. Other than that everything is analog. Logic gates are amplifiers with a non-linear transfer function. Once place I wonder if they optimize is on the order of doing the pushes and pops on function entry and exit. For example: EntryPoint: push ax push bx mov bx,#3 Seeing this come down the pipeline, they could opt to do the write for the push bx before the write for the push ax. This way, they would have bx free to use just a bit sooner. The same sorts of things could happen with the pop instructions if the processor already knows what will happen right after the return.

[@Kurukx]

Super interesting talk...Enjoyed

[@vitaobatera]

Important question for APUs: Since AMD is jumping into DDR4 we may see an improvement in iGPU performance, but wouldn't that performance Delta be more than double if they had chosen to go the 4-channel path instead of simply going dual? Was the dual channel idea only a money saver focused on Non-APUs SKUs ??

[@daweitao2668]

Don't you need 4 RAM sticks to make use of quad channel? Seems like a lot for someone who I assume is a budget gamer hoping to run games off an iGPU. And I think even for intel quad channel support is only on the -E series as well, so the real high end stuff.

[@joshwalrath9518]

AMD will get a boost from going to 14nm, having larger caches, and being able to throw more transistors at the GPU portion. We have heard rumors of HBM based APUs, but we do not have concrete info as to actual existence and availability. But when they will be eventually released, yes they will be significantly faster than last generation parts.

[@MarceloTezza]

Victor Custódio Kind of contradictory to buy 4 ram sticks for an APU on desktop, still on laptops i think it could make some sense, not sure if quad channels are more power hungry but could be an alternative to a discrete gpu. Anyway, APUs with HBM are comming, will probably solve that lack of bandwidth.

[@BrophyMichael]

@Josh Walrath - The coming Zen (Ryzen) will be 'Summit Ridge' and will support DDR4, After this will be 'Raven Ridge' which AMD have confirmed will support HBM.

[@MarceloTezza]

lzhlzh 2x4 is less so its cheaper, who buys APUs? People who want cheaper products.

[@bitbloop]

real nerds all know kanter is the true mvp in the game

[@makegaminggreatagain3907]

@4:48 "The bulldozer steamroller excavator had a few pretty fundamental problems" Well it only really had one and that is 65% of software still being released in 2017 is only programmed and optimised to make use of one core.

[@jasonlisonbee]

I hear people say things like that and want to say that most people run more than one program that at times demand the CPU simultaneously.

[@Thundermonk99]

Will echo the other commenters in saying that I really enjoyed listening to this. This is the type of stuff that goes a bit beyond what you will read on lots of different outlets.

[@MrApplewine]

Are there plans for AMD zen to provide complete mini computers in the sub $150 range which are cheaper or more powerful than the Intel Atom solutions which we have this year like the Biostar Racing P1?

[@bobthebuzztard]

MrApplewine we'll have to wait for a new mobile/desktop APU's

[@MrApplewine]

Ok, I'd like to see something like the ECS Liva Z, but maybe cheaper or more powerful. Hardware h.265 decoding definitely like Apollo Lake and also AV1 hardware decoding would be very good. www.ecs.com.tw/ECSWebSite/Product/Product_Overview.aspx?CategoryID=12&TypeID=130&DetailID=1729&DetailName=Feature&MenuID=209&LanID=0

[@OneManArmy1]

didn't understand anything until 36min where he said it was good. 😂

[@FrumpyPumpkin]

I MISSED DAVID!!!!!!!!!!!! DAVID 4 LIFE!!!!!

[@stevenprichard7864]

This is awesome! Tons of info

[@godssoldier597]

Excellent video!!!

[@creepinwhileyousleepin]

there's so much vocabulary i don't know here, but its very interesting.

[@thomasduhamel6766]

great video guys

[@SIC66SIC66]

Funny how I saw floating point benchmarks that show Ryzen beating Intel by a big margin.

[@TeslaRifle]

Will Ryzen is priced in Zenny in the Japanese market?

[@1967KID]

thank you for the info keep up the good work

[@jjforcebreaker]

Great material.

[@martyrogers8629]

This guy can teach me way more than any college class i ever took (i only took a few but you get what i mean)

[@Ott4vi4nu5]

this is the kind of stuff i subbed for !

[@blackfireburn]

this is why i subscribed keep it up

[@hentosama]

YES!!! 60 FPS HD JOSH

[@SIC66SIC66]

I am only 15 minutes in, and Josh said as many words as I hoped, none. lol

[@Etheoma]

No... Well ok I don't complain too much about the 5 - 7% IPC improvements it's the fact that not only has a 4 core 8 thread CPU stayed at the same price for 5 years it's actually went up in price even though the cost to manufacturer has went down. That is what I am angry about.

[@matttw7]

love content like this

[@StephenRimingtonrimtech58]

Awsome tech guys and crikey great tech info etc

[@alexandresoaresdasilva1966]

perceptron has been studied for more than 50 years in academia ;)

[@balbinderful]

It was a nice discussion, but entirely irrelevant too the entire objective Zen has for AMD future as a global technology company. AMD has 9% of the market share for consumer CPU sales; these people have bought these AMD CPUs because of one of two reasons: a) it was the best within my budget, compared too Intel; b) brand loyalty at mid and low range for AMD CPUs. Consequently, AMD simply has to chop out the potential aspects that at it’s customer base dose not want or use. Yet, removing them leads too unit production costs falling and allows them to deliver products at lower and mid price point for their existing customers. Smaller die size, means higher yields per wafer, which equals cheaper CPUs (tick that box of existing customers satisfied). Getting rig of fringe benefits, to focus on the mid-range usage and deliver identical or close performance compared to Intel tied to mid-range typical utilizations (tick that box of existing customers satisfied). The most expensive processor from last gen AMD offered to brand loyal users was the FX9590, which cost $200. Therefore, AMD is working within tight constraints on how much it can offer above this price point.

[@Great_America]

You guys rock! Great highly technical discussions. This makes Jayz look like a kindergartner in comparison ;-)

[@OldMan_PJ]

Wait, he complimented Intel's integrated graphics? This guy is definitely not a gamer.

[@jetpil0t]

Integrated graphics is quite useful for gamers, if you need to run diagnostics, run BIOS hacks, force remove drivers or RMA a GPU, iGPUs are a godsend. Previously you would need to have a second GPU on hand, just in case, and Windows 10 support for legacy GPUs makes this even more relevant. Obviously you can't game well on an iGPU, but that's not really the point of iGPUs in the first place. Being able to power a 4K display over displayport with an iGPU is actually quite impressive and immensely useful for a lot of people, especially in mobile computing.

[@TheGuruStud]

USELESS! That's what old GPUs are for. Any real man has those laying around.

[@SIC66SIC66]

Pretty sure Intel and AMD have been talking for a while about Intel licencing AMD GPU tech for use in Intel CPU's.

[@carlsagan3065]

These millennials need to respect AMD. At one time they were the top innovator.

[@Mutation666]

More content like this

[@mrtwinky2007]

0:23 Says everything about why you shouldn't listen to Ryan Shrout about anything related to "real world performance" at intel

[@AngelofVerdant]

Loved it

[@ichabaudcraine2923]

Awesome, awesome video. This is why pcper is bae. PS. Is someone getting spanked in the next room?!

[@ezforsaken]

Since when Ryen doesn't have eyebrows?

[@BenWillock]

Did... did we just get called dumb?

[@KevinSmith-qi5yn]

I can explain Zen's micro-architecture. The die size is so small it goes into the realm of god and thus brings about miracles.

[@norcalnative5582]

AMD is cheap enough with the next big releases for me to jump back in and try them out , Intel is ahead now but the price for their processers is too much for my budget. I'm going to wait to see if ryzen can atleast make amd a viable subsitute for the price

[@WesSites]

Not sure if the comment, "apple makes the best phone products on the market" was sarcasm or not...

[@Cooe.]

I love David, but he just got a TON of stuff wrong lol. Zen having worse simultaneous multi-threading than Intel? xD If only.

[@Maelthras]

Good god, this guy is supposed to be a cpu design expert, but his microphone is from the 90's. AMD bulldozer was not designed by engineers, it was an automated cpu design, the following designs of the architecture were touched up. Bulldozer was 20% larger than it would have been if it had been designed by actual engineers and there would have been no way they would have missed the design faults like bad prediction paths and cache misses that literally destroyed any chance that it would be competitive in any market. Ryzen on the other hand was made by jim keller, he created the amd k7 processors which were very competitive at the time, even had a hand in creating x86 and x64 instruction sets themselves. I am really waiting to see AMD dive into the server market with zen, where they will be having processors with up to 32 cores/64 threads with a competitive chip that has high ipc. Currently the most cores intel has packed into a cpu is 22 cores.

[@joshwalrath9518]

A little bit more complex than that. AMD does rely more on automated place and route as compared to Intel, but that does not mean that Bulldozer was entirely automated. Many of the more important structures (ALU/AGU pipes) were placed by hand to get the most performance/efficiency out of, but I have heard some nebulous stats that AMD is about 75% automated while Intel is in the 10% to 15% range. AMD still does not have the sheer numbers of engineers to do hand placement of transistors and structures that Intel does, so they continue to use automated processes in design. Keller is an awfully smart guy, but he is not the only smart person at AMD. They seem to have found a pretty good balance with Zen to give them good performance and good efficiency even in the face of Intel's seemingly unlimited amounts of money and manpower.

[@Maelthras]

Actually the automation was much higher for bulldozer because instead amd pushed its engineers towards apus and hsa which they used in the ps4 and xbox one, those are literally the only thing that kept amd afloat while bulldozer under performed in every aspect. Everyone was thinking amd would be sold off or bought out but the cash revenue from the consoles and the apus kept that from happening, jim keller is the leading expert in the microprocessor field. Worked for apple and amd for many years and was the lead engineer behind ryzen.

[@SuperG316]

I mis something. 1st The prime problem of AMD Buldozer high-end is 32nm vs 14nm. Handicap. It are small cores sharing FPU , but are real big real estate on 32nm Die. Eats a lot of power and limits clock scaling compared to 14nm. Zen and intel Core Are Big Full cores on 14nm which are small on 14nm Die . 2. Bulldozer architecture may be bad. but how bad it should be compared to something similar on 32nm or estimate what nextgen from Bulldoze family on 14nm could replace AMD 9 Series. 6GHZ Turbo and 125Watts? A Excavator on 14nm small 8core 4 FPU chip Architecture is problem its one point in the stack of problem. Stack 1 32nm vs 14nm 2 Architecture IPC 3 TDP 4 cloks 5 number of cores etc ------------------ Performance al those work to together. Buldozer at 14nm Higher kloks to compensate lower IPC. Or because 2core/1FPU module could be == iNtel 1C/2HT 16c/8FPU vs 8c/16t What are the tought on that?

[@joshwalrath9518]

Look back at Sandy Bridge vs. Bulldozer arch... similar process tech and speed. Still a large difference between the two chips in terms of performance. Yes, process tech plays a big part, but AMD stalled right out of the gate with the architecture and only now that they have cleansheeted something new are they going to be competitive. Excavator on 14nm would still have an IPC deficit vs. Intel's latest.

[@SuperG316]

With the First Bulldozer AMD did not reach clock target. AMD latest top end what the can do , a 9 series 5gHZ 200+watts monster on 32nm. Imagine what that will do on much more optimised but shrinked achitecture. ULV TDP on 5GHZ or boost klok as high to compensate IPC I just wonder.

[@dakrawnikg4480]

FX 9590 was released with a $1000 MSRP... You can now buy it for $200. Bulldozers' handicap was AMD.

[@sebastienberger2890]

I think it's was because the disponibility was to low to offer at the time the 200-300$ price.

[@dakrawnikg4480]

Seb Gontrand LMAO. Because of the what?!

[@TheKruxed]

Dissecting Zen via the medium of talking about Intel and Nvidia for 2/3 of the video lol... The chat is fine, just thought I'd learn something other than the obvious stated at 51 mins

[@MARCOBOSS123]

bro that mic was laud

[@jmsantos1983]

amd rizen gpu?

[@net200777]

I just want to know I was thing... why the hell does Zen only have 16 PCI lanes??????¿??????

[@joshwalrath9518]

I think the rumors are that it has 24 PCI-E lanes. 16 for graphics that can be split into 2 x 8. 4 lanes for M.2 and S-Express, and then four more that attaches to the Southbridge.

[@net200777]

Josh Walrath so my next question is.. why? I haven't don't my research as you might have guessed but.. if it's true that Zen is going to head to head with Intel in the 6900k.. then I believe the 6900k supports what is it 48 lanes?

[@joshwalrath9518]

Yes, but AMD is again aiming at a different price point and trying to make the chip more marketable. So it really does not want to have 2000+ pins on a CPU that won't reach $1K all the while keeping motherboards relatively inexpensive. I believe they have 1331 pins which allows more PCI-E lanes than what FM2+ gave with 906 pins. Intel is able to give all that PCI-E support because these are essentially server level parts with a price tag to match.

[@noway8662]

Platform cost. Broadwell-E is more of a server or super high end type platform with corresponding features and price. So Ryzen is not really going "head to head" with Broadwell-E, but instead doing its own thing and slotting itself between Intel enthusiast and mainstream lines and offering good price/performance ratio while doing that if what we have seen in leaks and speculations turn out to be true. R7 1700 ($319) or R7 1800X ($499) + X370 ($149-199$) + 16gb DDR4 ($120) = $588-$638 / $768-$818 i7-6900K ($1030) + X99 ($199-$400) + 16gb DDR4 ($120) = $1350-$1550 10-core 6950X adds further $700 to that price for 2 extra cores. Naples CPU:s and motherboards for those are what you want if you are looking for more server-type features or large amounts of PCI-E lanes. Those will launch during Q2 according to current schedule from AMD.

[@oysterhead5150]

My 8350 is still pretty fast, especially for the price.

[@666papcio]

Go Go AMD

[@ngkhanmein]

"u know" i don't know!

[@chomper720]

8-threads should been the average 3 years ago!!!

[@johanjac1]

Intel and AMD is holding the public back they have the tec its a money monopoly game between the two companies they have cpu's and gpu's that is 3 time faster than the stuff that is coming out with Zen, Ryzen or i7's its not close even what is out in their "Black Box Rooms"

[@dakrawnikg4480]

AMD has been in the red for YEARS, and you think they are sandbagging? Wow...

[@johanjac1]

It's no conspiracy you tell me, we are getting the right stuff from their labs and newest tec ?? Give me a brake Ye Right !!,so they dont have better tec in their Arsenal !!

[@seongwonchoi]

ryzen? or zen?

[@stayfrost04]

Zen = Actual micro-architecture Ryzen = Desktop chips based on Zen Napples = Server chips based on Zen

[@robertpearson8546]

When you talk about IPC or the older MIPS (Meaningless Information Propagates by Salesmen) you have a meaningless statement until you actually define what an "instruction" is. Is it just furious activity in silicon or does it actually do a computation? MIPS is "It is a tale Told by an idiot, full of sound and fury, Signifying nothing." On one 68000 one instruction can be an 32-bit integer multiplication. In the SPARC chip is can be 1/33rd of a 32-bit integer multiplication. I once designed a system to test SCRs. When I looked at microprocessors, I first picked a simple semantic - the computer must turn on a light bulb. The F8 took 14 instructions. The Intel 8085 took 7 instructions. The 6800 took 3. Adding an addressable latch cut the 6800 instructions to 1. All to perform the same act. Direct Execution architectures have the best performance since the machine language is a high level language. Compare the WD LSI-11 to the WD P-Machine. It is the same hardware. They just changed the microcode from implementing the LSI-11 architecture to implementing the Pascal p-code. The throughput tripled. The instructions/second changed only because the architecture changed and the meaning of each instruction changed. Compare the Novix to the 68000. The Novix has 5000 gates on a gate-array chip and is clocked at 2MHz. The 68000 has about 68000 gates clocked at 10MHz. The Novix has over three times the throughput. The Novix is a stack machine implementing the Forth interpreter in hardware. The 68000 is a variant of the 1950s von Neuman architecture. I asked Charles Moore if he would make the Novix chip as a custom chip. He said no. I had to agree with him. The resulting CPU would cost about $7.50 (only 5000 gates) and would require 2ns memory chips (typical chips were 250ns). Again the instructions/second changed because the meanings of the instructions changed. IPC only reflects the amount of activity the silicon is doing but has nothing to do with the actual throughput.

[@pcper]

MIPS is a company, and the acronym stands for Microprocessor without Interlocked Pipeline Stages, and not your clever "Meaningless Information Propagates by Salesmen". Their architecture was RISC.

[@moorishbrutha4363]

*PC Gamers:* Isn't this the same dude who was lying about Nvidia's Async Compute capabilities back when the Beta of *Ashes of the Singularity* came out during Maxwell lineup...... _I guarantee you all that he isn't even a gamer to begin with_. He was talking all of this stuff about Nvidia will have Async Compute capabilities for PASCAL not telling people that Nvidia's pipelines are mostly full so there wouldn't be any use for Async Compute for Nvidia's cards to begin with. *Gamers:* Be cautious of alot of these dudes on KZbin who aren't real gamers.

[@dakrawnikg4480]

"Isn't this the same dude who was lying about Nvidia's Async Compute capabilities..." Why don't you find out first?

[@moorishbrutha4363]

*You really should learn what a rhetorical question is before typing....you really should.*

[@dakrawnikg4480]

lol all that was rhetorical? cool story.

[@moorishbrutha4363]

God, how much are you being paid for all of this?

[@dakrawnikg4480]

lol Paid for all of what? My three sentences?

[@seanhornibrook]

51:45 HAHAHAHA -- seriously - AMD, Intel, Nvidia, basically everybody - please PLEASE unfuck your naming schemes... :/ Please come back and talk again in a few months... This was great.