[4] Forrest Norrod, AMD EPYC

[@revolutionofone8541]

This is one of the best and most candid interviews Forrest Norrod has given in several years. You asked great questions as always but I was most impressed with how open and thorough Mr. Norrod was in his answers. Not the normal, tight lipped, or company scripted answers. Maybe you should title these videos Conversations with Ian? That is much more accurate in my opinion. He definitely feels comfortable and relaxed conversing with you. One of your best interviews. Love the technical depth and longer length too. So much to cover. Thanks!

[@TechTechPotato]

I've interviewed Forrest a few times and he always comes across as relaxed and open. But I do phrase my questions a bit differently so I get different answers than expected

[@solidreactor]

Thank you Norrod for sharing these interesting insights of Epic and AMD. Hopefully Dr Cutress and you can have another interview toward next release. Great interview Cutress! :)

[@AndreTVOC]

i dont know how you can get this interviews, but one thing is for sure your doing a good job, thanks for everything Ian, and thanks Forrest for sharing with us your knowledge

[@TechTechPotato]

I've been senior CPU editor for AnandTech.com for 10 years ;) I've Interviewed Forrest before, and Lisa, etc.

[@cj09beira]

usually interviews with these big companies endup with little to no new info or insight into anything, really glad this wasn't the case, this is probably one of the best interviews in quite some time.

[@jakefromhackmud9457]

Amazing! Another great interview Ian!!!! Thank you!

[@OverlordActual]

Alright! Great interview. Always like listening to him speak at events.

[@M00_be-r]

8:25 I m glad you asked this question about Milan rollout and demand and not skipped/left this question just to be polite cheers Ian

[@MaD_fX]

Excellent interview, thanks Ian!!

[@gabrielecarbone8235]

Greeting from Milan, Italy! Too bad you could not make it here this time :(

[@Speak_Out_and_Remove_All_Doubt]

'Epyc Instinct' is such a great coming together of product names!

[@LittleMan2225]

Hey, quick tip regarding your thumbnails, don’t include text on the bottom right, it gets covered by the “video length” label on mobile (try visiting your channel on mobile and seeing this videos thumbnail and you will understand it)

[@TechTechPotato]

If I did that, people would complain it looks like a thumbnail from a different famous interviewer. Plus, in this format, you can see the person being interviewed and their name. My name is inconsequential.

[@aev-g8c]

@@TechTechPotato For a moment I actually mistook this thumbnail to be a video by a certain famous interviewer / podcaster. I don't think those sorts of details differentiate the overall look of the thumbnail enough to make a munch difference.

[@eazen]

@@TechTechPotato you should do what you want and not think about others. You're hampering your own creativity otherwise.

[@mentalplayground]

Great interview. Thank You

[@HidroPig]

awesome interview!

[@asitisrequiredasitisrequir3411]

Is Trento like an APU but supercharged?

[@TechTechPotato]

More like a Milan with a unique IO die

[@jaynorwood2]

Extended Speed Mode is a CCIX feature, right? Is their coherency also CCIX symmetric?

[@Speak_Out_and_Remove_All_Doubt]

I found what Forrest said about Oak Ridge's Frontier setup very interesting/confusing. I was under the impression that Frontier was going to use Milan + Instinct MI100 in a 1 + 4 design. But given that the Instinct MI100 was announced back in Nov 2020 and Forrest said it would use an as yet unreleased Instinct chip then clearly it can't be using the MI100 after all. Well, unless Forrest was getting confused with Lawrence Livermore's El Capitan which has been officially confirmed to be using a next generation Instinct GPU.

[@fotmheki]

In this tour of Italy I will ask about Florence, because Ponte Vecchio is there :P Nice interview!

[@Speak_Out_and_Remove_All_Doubt]

Hey Ian, do you ever watch Ready Steady Cook? One of the chefs on there, Jeremy Pang, reminds me so much of you, voice, sentence structure, intonation, etc.

[@TechTechPotato]

I don't, but I just looked up some videos. I can certainly see the correlation!

[@peterjansen4826]

Hi Ian. Here my big question for mister Norrod. For Epyc (and Threadripper but even more so for Epyc, I assume) AMD cherrypicks the best dies. Which impact does this have on the quality of the silicon for regular desktop-users? Would we get CPU's which get the same performance at lower power if we would get those cherrypicked dies on the 5800X/3900X...? Would we get a higher performance? Do the dies which perform best at lower power (servers) also have a higher peak performance or not? I understood that that is not necessarily the case but I am not an expert. Second question: how much more performance would be possible if the I/O die would be made by TSMC ('7 nm')?

[@TechTechPotato]

1) I believe AMD's chiplets are so good with bins that there are plenty of good dies for both. TSMC's defect rate is only 0.09 per cm2. 2) I tried to ask that, with respect to the IO power wall, but it was hard to get an answer. Innovation would appear to be coming, though exactly how much isn't known yet.

[@NCXBOXx]

​@@TechTechPotato He did say that they're "going to continue to drive process very hard" regarding the uncore and that's making me somewhat hopeful that we're going to get a non globalfoundries process for the IOD soon enough \o/

[@cj09beira]

for the difference between bad and good dies, your answer lies in the 5900x and 3900x, when people over clock them there is usually quite a huge difference between the first ccd's max clocks and the second's max clocks, i dont remember exactly but its some thing like ~4.6+ all core in the main one the the other one is ~4.3 all core, so what one could expect to happen if they gave us a cherry picked 5900x with 2 great ccds would be around the same single core clocks, but higher all core clocks as the second ccd would no longer be quite a bit slower

[@peterjansen4826]

@@TechTechPotato It makes sense that Forrest and other high executives from AMD are tightlipped about 2. They don't want to throw shade at GF which is a partner, be it a forced partnership. Yes, you did try to get the second question answered. Unfortunately.

[@peterjansen4826]

@@cj09beira My impression of the 3900X (I don't have one, I have a 3700X), it seems to have one good chiplet and one bad chiplet (still all cores at the same clock frequency as the 3700X for a full multithreaded load of course). That way it can still get a good single/dual core boost. I have seen multiple KZbinrs point this out and they showed some monitoring-data which seems to confirm it.

[@wandererdragon]

My guess is that Trento family combines Zen 3 cores with PCIe5 capable IO die, which enables the use of CXL and thus creates "coherent" system.

[@davidbetancourt4028]

Is it just me, but does Ian seem like he's got a powerlifter build? Would be pretty cool if that's a hobby, heh. Random thought. Nothing but love! Great interview!

[@slyblood85]

Forrest seems like a nice guy to work for

[@tanmayd9873]

I didn't know John Goodman joined AMD

[@BigEightiesNewWave]

He didn't.

[@Speak_Out_and_Remove_All_Doubt]

I wonder why AMD were so happy to lay out their roadmap for 3 generations at the start of Epyc but then since that have become really tight lipped about upcoming stuff? We have known about 1 extra step, Genoa, officially for ages but absolutely nothing about what's after this. Maybe it's Turin but my point is AMD were so happy and proud to show off that they had a clear roadmap all planned out but are they saying now that they really aren't sure what their plans are anymore?

[@RyTrapp0]

Road maps aren't and have never been released out of "pride" and "happiness" - this is stupid, they're just road maps Hint hint, we're a year deep into a global pandemic that has literally affected everything, especially the tech industry - it probably has something to do with that...

[@tommihommi1]

Once Intel has something on the horizon that can compete with Rome and Milan, there will be new announcements from AMD. That three generation roadmap was made under the assumption that Ice Lake is launched in 2019 to compete with Rome, AMD had to promise something solid that can keep up with the innovation in the industry. You don't need to announce details about what's in your pipeline, when the competition doesn't have anything announced that can hold a candle to your own currently released products.

[@Speak_Out_and_Remove_All_Doubt]

@Unabashed Hedonist All the more reason you want to show customers you have a clear and very strong roadmap ahead???

[@Speak_Out_and_Remove_All_Doubt]

@@tommihommi1 Yeah, I am wondering how much AMD is now 'doing an intel' and slowing things down, sandbagging performance etc as like you say who are they competing with?

[@tommihommi1]

@@Speak_Out_and_Remove_All_Doubt they don't need to slow things down, there's just no point in throwing out risky promises. With all the exascale supercomputers based on upcoming AMD generations, that's really all we need.

[@salmiakki5638]

Was locking EPYC chips to the vendor's chassis and motherboard really necessary for "secure computing"? How big is the actual "threat level" for hpc/cloud etc? I feel like this move was not really needed, and the only thing it does is kill the second hand market and turn perfectly capable chips into e-waste way before time. Secondly: can I have a ticket for the Genova event? :)

[@tommihommi1]

All the big customers asked for the feature, so AMD implemented it. That's all the justification they need

[@TechTechPotato]

Part of me thinks he misunderstood the question - locking processors down to specific systems vs just customized parts for that specific customer, perhaps he thought I meant the latter, e.g. the 7V13 for Azure etc.

[@salmiakki5638]

@@tommihommi1 i guess you are right. How can the end customers be fine with this move from the vendors? The resell value is not insignificant and even equipment from the biggest players (google, facebook) usually find it's way on the second market. Now they buy into a product that has 0 value after is separated from his chassis

[@tommihommi1]

@@salmiakki5638 the end customers have also asked for it, it's an important part of the move towards trusted supply chains.

[@NicolaiSyvertsen]

@@salmiakki5638 Why is this a problem if you can just get the entire server second hand? I think this will just force the liquidators to sell complete servers instead of parting them out to ease shipment. That is how I understand this locking of the processor to the platform.

[@ceasarentila1539]

What if you call these "ian-terviews' lol.

[@odogg6899]

potatoes. Algorithm comment

[@Speak_Out_and_Remove_All_Doubt]

I wonder if AMD might start to 'do an intel' now they have such a lead on the competition, ie slowing down chip roll outs, sandbag performance etc as the only real competition AMD has right now is supplying enough of what they manufacture. Intel doesn't have anything close to a 64 core Milan so why would AMD rush out Genoa, let alone release time scales and details on Turin? I wonder if the reason we have heard so little about Warhol is because it's dependant on how good Alder Lake is and when it's real availability will be? If Intel's AL does manage to pull out a big lead over Zen 3 and so therefore Zen 3+ (in gaming and sub core workloads at least) when can/will AMD ditch Warhol and bring out Raphael early instead? After all, a load of 5nm wafers just became available as Apple is moving to 3nm.

[@1idd0kun]

They need Zen3+ on desktop because they can't get 5nm wafers soon enough for Zen 4. That's why Zen3+ is only desktop, while Epyc will go from Milan (Zen 3) to Genoa (Zen 4) and launch early next year. Basically, they're prioritizing server over desktop which makes a lot of sense considering their limited wafer supply.

[@Speak_Out_and_Remove_All_Doubt]

@@1idd0kun I wouldn't be so sure about 5nm wafer supply, AMD has had a lot this year and this recently increase a lot too with Apple moving to 3nm. We don't currently have any 5nm AMD parts so it's interesting what AMD is doing with all this capacity. I would suspect some kind of closed door NDA bespoke chips but it's hard to say.

[@1idd0kun]

@@Speak_Out_and_Remove_All_Doubt Since AMD did confirm Genoa is coming early next year, it means that Zen 4 will land first in server rather than desktop. For a while, they'll have zen 3+ parts in desktop while server is on Zen 4. This is kind of a first. Generally desktop and server gets the new generation more or less at the same time. So it's not like AMD is slowing down, they're just changing their priorities. And 5nm supply being limited has to be one of the reasons. May not be the only one though.

[@Speak_Out_and_Remove_All_Doubt]

@@1idd0kun I am not sure AMD have definitively stated Genoa will be coming 'early next year', tbh I would be surprised if it's even the 1H. But you are making the assumption that Zen 3+ isn't 5nm and also that AMD won't bring out Racheal before Genoa. We have no official time tables of decisions for any of these things, not to mention I don't believe AMD has officially referred to Zen 3+ even once so it would be no skin off their nose to not release a Zen 3 refresh. I simply don't believe we have enough information to make any assumptions at this point. For example, we don't know that 5nm availability will be worse than 7nm by the end of this year. Don't forget that AMD will still be needing huge amounts of 7nm capacity for the PS5 and Xbox chips, plus all their RDNA 2 cards, then factoring in Apple moving over to 3nm, I believe all of this suggests that 5nm would actually have a better availability than 7nm in Q4 2021 / Q1 2022 for Zen 4 chips.

[@1idd0kun]

@@Speak_Out_and_Remove_All_Doubt I'm pretty sure Forrest Nord mentioned Genoa is coming out around this time next year in the interview. He mentioned it casually, so it's not exactly an official confirmation, but he wouldn't have said so if he wasn't sure. Besides, we know Genoa has to be ready early 2022 for the Capitan super computer to be deployed on schedule. And about Zen 3+, it was leaked from pretty good sources that it will be fabbed on 6nm, which is basically 7nm+.

[@danieljones6778]

first