How AMD is Fighting NVIDIA with RDNA3 - Chiplet Engineering Explained
HTML-код
- Опубликовано: 7 июл 2024
- Sponsor: Thermaltake Core P3 Pro on Amazon geni.us/6Q6M
We're talking with one of AMD's leading engineers (Sam Naffziger) about RDNA3 GPU design and challenges with going to chiplets for a GPU. We only had about 15 minutes to cover the highlights, so we focused mostly on the differences between chiplets on Zen CPUs and RDNA3 GPUs. This gives a unique insight into the architecture of AMD's next attempt at another "Zen moment," this time with its graphics division.
Watch our video about the cool hardware feature on the AMD RX 7900 XTX reference video card: • More Video Cards Need ...
Until 11/23/22, code 'THISISFINE' will give you 10% off the GN store! The best way to support our work is through our store: store.gamersnexus.net/
Like our content? Please consider becoming our Patron to support us: / gamersnexus
TIMESTAMPS
00:00 - Learning About AMD RDNA3 GPUs
01:48 - AMD Zen to AMD RDNA
06:45 - Chiplets Changed AMD
10:54 - GPU vs. CPU Design Requirements
12:04 - Microscope Shot of RDNA3
13:05 - Fanout Routing
14:15 - Maximizing the Die Area
** Please like, comment, and subscribe for more! **
Links to Amazon and Newegg are typically monetized on our channel (affiliate links) and may return a commission of sales to us from the retailer. This is unrelated to the product manufacturer. Any advertisements or sponsorships are disclosed within the video ("this video is brought to you by") and above the fold in the description. We do not ever produce paid content or "sponsored content" (meaning that the content is our idea and is not funded externally aside from whatever ad placement is in the beginning) and we do not ever charge manufacturers for coverage.
Follow us in these locations for more gaming and hardware updates:
t: / gamersnexus
f: / gamersnexus
w: www.gamersnexus.net/
Host: Steve Burke
Camera, Video Editing: Andrew Coleman
Guest: Sam Naffziger 
Игры
Until 11/23/22, code 'THISISFINE' will give you 10% off the GN store! The best way to support our work is through our store: store.gamersnexus.net/
Watch our video about the cool hardware feature on the AMD RX 7900 XTX reference video card: ruclips.net/video/8eSzDVavC-U/видео.html
Lol, nice reference code joke with nvidia cables burning
Did you chew the engineers out about locking down memory clocks?
Love you got this interview. so much better than marketing fluff about tech that companies push. Loved this TY GN
Why does he look so annoyed?
The fact AMD isnt hyping RNDA 3 tells me all I need to know
It's very special when you get 15 minutes from someone like this! Look up his resume to realize he holds over 130 patents, and is the main designer behind many innovations in the computing space. Thanks GN for giving us some time with the man behind it all.
Well I am convinced that Smart Access Memory was named after him.
Well I am convinced he is an alien sent to earth to help uplift our civilisation... aliens are gamers. Fact.
@@TheMastaRob Lol that's my dad, in a nice way this is kinda accurate
I'm loving these engineer discussions. Plenty of people cover the consumer-facing side of hardware, but it's so cool to see what goes into them behind the scenes. Really makes you appreciate all the hard work that goes into our FPS!
Always an excellent opportunity to learn from the engineers when we get a chance to meet up with them!
I really wish GN asked one more bonus question on the future of silicon, or should I say, the lack of it. What technologies is the industry generally looking forward to next? Though in fairness, Sam probably can't talk about that as AMD would have his ass for dinner if he said anything regarding those long term plans.
@@GamersNexus We engineers find it really interesting and I appreciate it, thank you!
Yes, thank you very much.
@@GamersNexus Brilliant! As an "Electronics Junkie" since I was born in 1958, I find this kind of content informative & educational. Thank you...🇺🇸 😎👍☕
The amount of PR training that these guys go though is insane. Engineers such as him know very well what RDNA4 is going to look like and possibly even RDNA5, but to make sure one only speaks in "released product" terms takes a lot of holding back.
It's more impressive they can still speak English. Engineers tend to not enjoy public speaking because it can be tough to translate what you've been doing into terminology everyone can understand.
@@EkiToji My professor always said if you know you can explain it and if you can explain you will never forget.
@@andresguerrero3347 That's a great quote and very true. Sometimes people that are passionate specialists dedicated in things can override their shyness too and be really enjoy/good speaking about it. It's the stuff they love in life and live for so it can make it a lot easier when they're super knowledgeable and/or passionate
When he mentioned the "Decoupling of the clocks" in the shader engine (at the design requirements section of the video 10:50) I kinda got the same feeling of that he could not talk much more about it, felt like there was more thought behind the decoupling.
For me he kinda hinted (maybe unintentionally and maybe I'm reading too much into it) that they have paved the road for RDNA 4 to having a more Chiplet oriented design than a RDNA 3 like MCM design, with this RDNA 3 shader redesign and decoupling of clocks (amongst other new RDNA 3 design choices).
We know that Raster usually doesn't scale well over chips (crossfire/sli) above 2/3 modules so I believe that the great "Zen benefit" like scalability of chiplets will be seen in Raytrace engine designs. RT does usually scale linearly which is amazing for chiplets.
Think we might see in the future a 2-3 chiplet raster design with a X number of raytrace chiplets, maybe for RDNA 4 or 5. That RDNA 4 or 5 will extract the RT part from the current shader engines making the raster engine smaller while having the RT in dedicated RT chiplets (including BVH, denoiser, upsampling).
Yes I might have read too much into it :D I guess this is more of my personal hope of having raytrace being main priority and raster as the "secondary/classic/compatibility" chiplet.
@@solidreactor From my point of view, it makes sense for power efficiency reasons, yeah, maybe one day we will see almost identical amd gpu's which differ only in the performance of ray tracing, the amount of memory and the width of the bus
This is what our investment into this channel gets us. These interviews are great, Steve has developed some great interviewing skills and Sam himself presents his great knowledge in an equally talented way. Between the both of you, this was a fantastic interview. Great watch 8)
Man Sam is just an amazing presenter and engineer. He really does a fantastic job of getting the point across in layman's terms.
Sam! But I'm sure he appreciates the compliment nonetheless! haha
Amazing
@@GamersNexus Does he not remind you a bit of current Tom Cruise in the face? lol That's the first thing I saw, especially in profile...
@@GamersNexus Doh! I went back and rewatched it to try and get that right and still goofed.. changed!
When you know what you are talking about … you can
I love when we get direct talks from the engineers! Thanks GN for the coverage as well as for all of your other informative and deep, "engineering"-focused reviews! Hope you can make more videos and insightful reviews of PC components and products for years to come!
Agreed, it is great to see a youtube tech channel actually getting into the tech!
so you're telling me the only reason why we don't have a 4 core 13900k right now is because of this amazing man right here... Sam is truly an incredible down to earth engineer. I'm super in debt to you
Literally a huge reason why
The reason is called Jim Keller - The OG Athlon K7-12(including the x64 instruction set) architect who came back to design the first Ryzen CPU's.
@@NPC-fv3nc Sam and Keller designed the first one together
Yeah it's clear that Intel planned on 4c/8t with 8MB cache for the long term.
With Nvidia's insane pricing this is gonna be my first AMD card
yeah! the only thing that was holding me back was their hardware encoder. but with these babies we got AV1 encoders which is amazing in encoding! plus with cheap prices!
Good luck with drivers and new games
@@gudvfqeugb4703 been running AMD cards for 10 years now. Drivers are solid. Plus, my power cables don't melt, so there's that.
Me too , 7900XTX
$900 and $1000 still insane pricing just for gpu alone
every time I learn more about CPUs I'm amazed that computers work at all. Just the sheer amount of tech in this chip that fits in the palm of a hand and we make like thousands every day.
When AMD was presenting RDNA 3 you could hear in his voice, though calm and soft, how excited he was with what they were coming up, like he was really trying to convey how cool all this new approach actually is. Too bad the audience seemed to have missed that, or at least that's how it looked like from the RUclips stream.
went as expected since it was all media.. always the downside to these closed presentations instead of being at the big open events.
And here I thought he sounded incredibly bored during that presentation. I guess it's just how he talks, but he definitely seems somewhat more lively in this one-on-one format.
I can tell he was excited, however, the technical terms were way over my head as a layperson. The only things I think I understand out of it is:
While logic capacity increases significantly from die shrinking, input output and memory do not benefit nearly as much.
Because input output doesn't benefit significantly from additional miniaturization, they can separate that section in larger but cheaper die.
Because the core units are smaller and I/O is divided out, they can fit much more units per wafer than putting them both together, cutting even more costs.
Infinity fabric seems to be an interconnected and deep weave of conduits made from a more advanced substrate that's not silicon.
Gpu I/O functions similarly to cpu so they can perform a similar strategy in designing gpu.
@@eddapultstab2078 A little off on the last point, GPU's can't have a similar strategy to CPU's because of the transfer rates required between the memory and shader engines are so high. However there are still ways that the GPU's can be compartmentalised too allow more robust die design.
@@4kirb the last part was a bit confusing, what I thought was gpu i/o unit and the resulting shaders do not count as the logic scale, which has the highest benefit from shrinkage and are closer to i/o and memory so they can use a larger die on that part.
Sam making the chiplet approach seem like an absolute no brainer in only 15 minutes. So awesome
Yeah, but he also mentions how he had to fight for and champion the approach inside AMD before they did it.
Companies don't like risks, and changing the entire architecture approach is a massive risk, what if all those benefits on paper don't turn out to translate to the real world? It could have put AMD generations behind if it didn't pan out.
Nvidia has shown clearly where the "same thing but better" path lead, AMD has shown us an alternative.
I know which of the two approaches I prefer personally.
@@NATIK001 fact, but also bulldozer was a fuckin disaster (i had the 8350)
@Griftercash that was before Zen of course. Dirk Meyer was a good guy but too many things were going wrong then with endless cash flow going into foundry upgrades and other complacency woes. Zen was a complete rebuild from scratch and take risks approach. It paid off.
He's a good salesman, which is a skill that helps even for the most technical roles you don't normally associate with sales.
I hope he realizes how much we appreciate him taking the time to do this. This was AWESOME!
He's my dad bro! I made sure he saw :D
@@sethnaffziger1402 Luckyyyyyyyy......your Dad is HIM
Top tier content as usual from GN! Always interesting to learn about the design challenges that the engineers face and how they overcome them with innovative solutions.
I will be brutally honest with you, guys. This is one of the best videos to ever land on Gamers Nexus YT channel. Kudos to you for getting an idea to put an AMD engineer in front of the camera and ask him about the "kitchen". And kudos to Sam, who has a rare talent of being able to explain complex issues in a way that is totally digestible to an average PC hardware fan. Amazing stuff, guys!
Chiplets?... Kitchen?... I don't know what everyone's talking about, but sounds tasty.
@@michaelmonstar4276 always good to know how the sausage is made
Hear, hear!
This is exactly what I was hoping to see more of. Thank you to Sam from AMD for the presentation, as well as the GN team for this continued emphasis on the work that goes into the products.
What I love this channel for is the actual little to no editing with that kind of stuff. Your videos guys are just authentic. It's not easy to pull off, in my opinion, so big respect from me.
For sure. Unlike Linus tech tips which are straight shilling videos. The content on that channel is so fake and there is a weird air to the videos where Linus is a complete ass and his employees are scared to say anything.
@@playernone2110 you couldn't have said it any better, i feel the same way after watching some of the ltt videos
@@playernone2110 oh, how little you know. stop falling for the intentional on camera persona, if they were truly scared they'd leave.. all them have skills that would get them a job with damn near any production company especially in canada..
@@JayPrakash-th4rd he *never* represented the common people
I know Sam a little. He’s a good dude. Did some mtb racing a decade ago with him. He’s an animal!.. It’s been a while but I remember him speaking about chip tech in general. He’s a damn smart guy. He’s responsible for lots and lots of patents. Good to see his work being implemented. I wouldn’t doubt AMDs success being partly to Sam’s work with chiplet design and integration. Great to listen to the design details! Keep up the great work! 🍻🍻
That’s so cool! It seems like MTB is popular with tech types - GN Steve’s big into DH, I do some club racing and enduro XC myself.
Sam comes across like a true engineer - very intelligent but humble about his and AMD’s huge achievements over the past few years. Hope we hear more deep dives from him and other AMD engineers in the future.
Had no idea he's into MTB! I'll ask him about bikes next time, haha
@@GamersNexus it’s been a while but I’d bet he’s still riding in his free time. Also, I remember giving him shit for AMDs power efficiency at the time (2010ish I think). He had recently moved from intel. I bet intel wants him back now haha. I don’t remember the details of our 4-5 hour conversation (driving to and from a race at winter park) but that it was always fun to speak with Sam. And I figured I’d hear about him in the mainstream at some point in the future. And that day has come! Maybe I’ll give him a call sometime soon to catch up. thanks for taking your time to go over things personally. It’s really cool to see the people that are responsible for creating the leading edge of technology we use and love so much!
@@damonm3 He's still an avid rider, does several races every season and he's on some crazy conditioning now (think he's doing 9-10 hours a week) and yeah the infinity cache and a bunch of other things are things he's fought to implement in the GPU space. it's exciting to see all that work paying off
@@sethnaffziger1402 i assume you’re his son? I figured he’s still riding a lot. I was near peak shape (late 20s) and he was faster and more technical without a doubt. The race we did in winter park was staggered start by age class. He started 10 min behind. About a 2 hour race. My only goal was to finish prior. We had a few local nuts pass us in the 60+ bracket (10 min behind him)Then I finished and Sam was maybe 30 seconds behind me so he almost made up the 10 min.. also we were drifting his x5m in the mtns and one gave us a little rush.. haha..good times for sure. Tell him Damon says hi, small world here on the RUclips🍻
The next gen Radeon gpus are turning out to be extremely interesting. Hope it's a good product, regardless of the horrible price market.
i suspect that similar to the launch of ryzen it will take until 2nd or 3rd gen chiplet designs for these gpus to become just as powerful, if not more so than the competition. im hoping that like ryzen they are able to release generations a lot faster than previously as well to really give nvidia a kick to the teeth
Motorsports question: "How fast can I go"? Answer: "How much money have you got"?
@@lilblackduc7312 Truly a Petronas AMG moment.
@@lilblackduc7312 haha so true! That applies to everything
Since this seems to be the first gpu gen he is working on I wouldn't expect it to be good already. Maybe in the next gen.
Nothing more beautiful than engineer to tell you how things works and what they did, not to send some marketing waxed boy to tell you basic things and that's about it.. I love things like this, community is really interested in a knowledge which this gentleman and other engineers have, this is the best way to see how much people respect their work, passion and intelligence which they have.
what if I want to see an engineer who doubles as a wax boy / stripper?
@@ubuu7 What if I want to hire a stripper as a computer engineer?
Man, that vastly simplified, imagine just how complex these devices really are! I mean I've been in IT for 30 years, and I am always just amazed at how they design any kind of silicon (or plastic now) lol Great content, thanks. =)
I read "How amd is FLIRTING WITH NVIDIA with RDNA3" and fear the worst.
So sweet
🤭
we don't want that fanfic
@@GamersNexus but I do.
"Would you be the SLI to my CrossFire?"
Sam is absolutely right about engineers wanting to work on the cutting edge. Back when I was a product engineer for a Telco equipment company, I was excited to work on the IP over wireless tech. You guys know it as WiFi. But the other part of my work was dealing with PSTN, which is fixed telephone lines carrying 56k modem stuff.
Sounds like to me a company would have to pay an engineer more to work on normal shit and cutting edge guys would get paid less LMAO
@@theq4602 youd be surprised, very few engineers work on "cutting edge" stuff. unless you have a graduate degree, working on R&D and anything related to cpu/gpu chip designs requires an extensive graduate background. "normal" engineers like me are working on all the "lower level" stuck like pcb routing, etc etc
I think this reflects what has bring AMD to a new "golden" era, it's not about working hard, you need to work smart. Something insanely true if we see the ARCs GPUs.
You can see that Sam's clearly passionate about this and knows a lot a lot about chipset design. He's truly a remarkable person and I loved seeing an in-depth yet simple interview with him. Hope to see more content like this!
I'm sure it would be asking a lot to get more time than you already did but I could see an hour long segment just devoted to discussing the various aspects of this design. More than anything, I'm just keen to see how it performs in-practice.
So excited for benchmarks, been putting off building a new rig to finally upgrade from my 5 year old 2200g build without a dGPU for months now until we know what AMD truly has to offer with RDNA3...
To be fair, any current gen hardware is better than 5 years ago, because what you are doing is mostly gaming.
Games can get much better once UE5 titles comes out with nanites and lumen utilized, but until then, keep your expectations low for other game engines like Unity, for no matter how much of a super computer you have, Unity engine will run like dirt if its going to be that unoptimal.
@@blueplanet... upgraded myself from 8700k to 7950x. Other then just setting some stuff up I do feel that I will want to upgrade to a more recent midtier card over my 1070.
Please please keep making these. I love these videos more than anything else.
You can tell he fought for this simply by how easily he can explain it in the simplest of terms. He clearly has experience selling this idea to bean counters who already knew nothing about the technology or the manufacturing process beyond timescales & costs.
A couple of obvious things from this presentation that AMD is likely going to do in the future and it centers around re-using more chiplets across more product lines.
1) Spin off the PCIe, display interfaces and encoding logic on their GPUs to their own chiplet. This single chiplet will be spread across their entire GPU lineup and even serve as part of the desktop IO. On desktop the memory controller die which also contains the basic links to the CCD.
2) Common dies for stacked cached that are used on the CCD on server processors and MCD on graphics. This isn't about performance but economies of scale as it permits AMD flexibility in how they allocate these cache dies.
3) Common Infinity Fabric link between CCD, FPGA, GPU and specialized IO dies. This dives into AMD's semicustom side but if a large customer *cough* Microsoft *cough* wanted say a four CCD and two large FPGA dies in a SP5 package for a server, AMD can leverage their existing die portfolio and quickly validate a custom package for release.
I think much of this first wave of AM5 and SP5 products are purely a stepping stone in their strategy. I strongly suspect that AMD is leveraging the same Infinity fabric interface between Zen 3 CCD, Zen 4 CCD, the socket AM4 IO die and socket AM5 IO die. The reason behind this strategy is that AMD could have launched AM5 products leveraging the Zen 3 CCD if it made sense to launch a DDR5 based platform earlier. Similarly if DDR5 was not ready in the market, AMD could have launched Zen4 on socket AM4. Essentially AMD had an insurance plan on how the market move and always get a new product out to continue competing.
Not covered in this presentation is how much energy it takes to move data across various wires. While chiplets provide lots of flexibility in terms of design, they require more power than a monolithic design. I would predict that in the Zen 5 generation, AMD is going to replace the traditional wire based packaging with a full interposer or EMIB-like technology to reduce wireless lengths and drop power consumption in this area. AMD has given a strong hint that this is the direction they're going by some of their HPC parts based around CNDA will include a Zen 5 CCD die(s) in the same package.
They will spin out every part they can spin out in order to make room for the gpu logic. Expect NVidia to follow they are a few years behind in this process, but the current gen clearly shows they will hit a wall soon. The high prices definitely partially at least stems from costs and binning problems. The Ryzen CPU designs are great and while Intel marketing made a mockery about glued processors in the early Ryzen years their engineering clearly probably saw that AMD struck gold with that design for at least 10-15 years.
It seems that large number of wires required to interconnect is the limiting factor here. Once they figure it out, it will be like lego, I hope.
@@sudeshryan8707 Doing that would probably require some kind of die-on-die direct stacking, but I don't even know if that's a thing right now. I suspect they'll try really hard to go full modular.
Maybe an Crossfire-like approach, if directly cutting up the current design isn't feasible? Maybe the downsides of a "multi-GPU" setup could be mitigated with some proper love and drastic latency reduction.
@@werpu12 I think NVIDIA already hit a wall with the 4090's. The high prices definitely because of wafer defects and binning problems. They are also hitting the power ceiling with their 600 watt limit as well. Their designs hint that they are simply just brute forcing tangible performance gains by throwing more power at the problem. AMD is now ahead of NVIDIA by a few years in power/performance efficiency ratio so it's almost a guarantee that AMD's next line-up of gpu's after the 7900xt will no doubt demolish NVIDIA's own lineup.
@@luminatrixfanfiction Tbh, 600W power limit for NVIDIA is bit overcranked on their size... it can be around 480 almost without performance loss. But they wanna sell these 3-5% performance for 120W+
Adored tv was right again, and he was stunned by the realization of how far ahead amd is at this point in time. I was stunned as well. This will be a huge game changer.
I really love it when you let engineers talk about the design. You know he really knows about his design by the moment the question is fires his eyes shine and immediately knows what slide to show
This is genuinely one of the coolest interviews I've ever seen. Thank you Steve (and team!) for always pursuing and sharing these opportunities
I like how he pointed out that the engineers don't like to port the IO stuff and much rather work on the "cool" stuff 😁
That was fantastic, love the insight. If they can have a "Zen" moment with RDNA, it'll truly be great for gaming. It'll even reverb through the next console cycle likely as well.
Pretty excited for these GPU's on mobile going forward, love these types of videos
i know this is gonna be litty spaghetti
You guys have really been crushing it with your industry interviews and hook ups recently. Great job.
I really admire the actual team working for AMD. Has been a fantastic uphill recovery since the dark ages of Hector Ruiz where I would every day expect to hear of AMD bankruptcy. Well done!
Steve and Wendle, taking PC hardware breakdowns to a whole new level.
Yeah it makes me feel like I am finally in the group that has that bleeding edge. In the past, I never would have thought that my current info would come from youtube. Yeah, steve and wendie have their ear on the train track.
These are my favourite videos you guys do, where you get to talk to the people behind the tech and get a sense of their vision.
THIS is what I love to see! Let the engineers speak! I would watch an hour of this with rapt enthusiasm - really wish we could get an uncut version of the full presentation. Love hearing about all the ways the engineers are pushing their designs and the technologies they have access to in new and different ways to achieve their goals!
Man. You gotta consider that a hell of an interview catch. And from AMDs perspective, they've got an engineer that's personable enough along with a mastery of the engineering to make this interview a big win for them too. If nothing else, getting information out to a few hundred thousand people capable of understanding it (at one level or another). Good job.
I could listen to him talk about this topic in detail for hours. Happy for the 15 minutes you managed to get. :)
Wow...that sure was interesting. I could definitely spend hours just listening to the various design decisions being made to bring this design to fruition. Well done again GN!
I'm guessing Sam is always the smartest guy in the room.
Sam- "Smart access memory"
Yea, he's talking about hella complex stuff and there's no umms, no blips in his vocabulary, I'm only smart enough to know I'm really not even the same species as dudes like Sam. I can't talk about my dogs with the confidence, ease, and agility he talks about obscure computational minutiae
Love listening to engineers, massive respect to him for giving us a look behind the scenes about this stuff even if it's AMD allowing him to speak to help show things. Super cool thing to do occasionally and everyone wins. Hope he doesn't mind having to do this speaking stuff especially with all the PR training they put people under for good reason, he's a great presenter and we really appreciate it
Sam strikes me as a smart and capable guy who might not be aware of just how many hundreds of thousands of people are going to view this video. He's about to become quite a bit more famous among AMD fans
This is great, more interviews like this please! :)
I would have loved a two hours video like that. It's so interesting when you can interview the biggest brains of a company :D
always a treat to hear from the mind behind the more interesting and impactful designs and processes, many thanks to steve and GN team for grabbing this fun little q&a!
It's presentations and interviews like this that provide the real roadmaps for platforms and its always a treat to see them come up, thanks for getting some time with the man himself GN!
Incredibly interesting to hear some background perspective and it was very digestible, even as someone that doesn't have as big of a clue about the details behind this kind of hardware. Sam seems incredibly competent and makes me hopeful that there will be many more great technological advancements from AMD with people like him working on it!
Looking forward to more
He could make a 4 hour presentation and I would watch excitedly ... can we get more of this ? ... Sam ? Steve ? Please?!
Thanks to Steve and the team for providing this kind of in depth coverage. It is good to get some actual in depth info beyond the usual benchmarks and marketing speak. Great job!
I wonder which was the last question you didn't have time to ask Sam. Amazing interview, this kind of coverage is something that is hard to find among the tech press, focusing on technology and not only on raw performance on synthetic benchmarks. Thank you guys!
This is so cool! I loved the discussion you had recently with Intel and Nvidia but this was something different. Its insane to think how much this one person shaped the current CPU space and will also maybe or hopefully shape the GPU space.
Thank you all for recording this interview. I am super excited for more. I hope Intel, AMD, Nvidia and others see the benefit of these kinds of discussions and will give you more opportunities to bring us more "behind the dies" content!
I literally always wait for your video!!!
GN is killing it with these interviews! Well done Steve and crew, I love the trajectory you're going in.
Great interview, always great to learn about some of the concepts that these engineers come up with, and it was great that the language wasn't so technical I couldn't follow along.
I'm glad they decided to share some of this background. I know there's a tendency to keep everything a secret with technology companies.
Great! very informative thank you
Great video man! And it's really great to have a man from inside with all the experience to tell all these things.
Never stop doing these, pls!! We need more engineer interviews.
Seems AdoredTV's guess of about 1000 on a wafer was correct. Also looking at the slide around 8:55, when fabbing a 16 core chip, taking 16 chiplets as a 100% value it seems chiplets are also scale to larger designs better than the monolithic approach. The monolithic version going from 8 cores -> 16 cores is a 2.1X increase in cost, but in the chiplet version it only costs an additional +0.25X to go from an 8 -> 16 core, that's not just a little saving that is significantly better.
I am extremely interested in this generations performance and in how they iterate on this.
I'd love to see this or similar packaging technology used for APUs. We know how much of a problem MS and Sony are having getting the costs down for the silicon they use in their consoles, which is probably down to the use of a monolithic die, and the fact that a die shrink doesn't automatically save costs in the way it used to. If the CPU and GPU were on separate dies and communicated via infinity fabric / UCIe, or whatever, presumably yields would be a lot better. Could potentially be exciting for other products that need a single high performance APU part, like gaming laptops or handhelds.
Amazing content. So happy y’all are getting opportunities to work with key industry players. As usual, high quality output. Bravo.
This was awesome, you were right at the beginning! Thanks for sharing, cant believe this is free^^
F'ing TRAIL BLAZERS! Sam and GN both, in their own arenas. Bravo to all involved!
I understand that he may not be able to talk about it, but it would have been really interesting to hear him talk more about slicing up the graphics logic into several separate dies. He said that the bandwidth requirements would be too large, but is it something that they are working on for the future?
the cost is prohibitive. You just need to look at what apple did with the M1 Ultra were the entier side of one of the dies is used to bridge the 2 dies and this uses a silicone interposes (as he said very costly). This is something that just doe not make economic sense currently for consumer products (even gpus costing over $1k).
Also in the end you end up with a lot of extra die area needed for the interconnect (on the M1 Ultra that might well be as much as 5% of the die area ends up being used for interconnect).
Thank you for making this video and to Sam for taking the time for it. This was the most interesting video about RDNA3 I've watched so far and got me to appreciate the technology that goes into it even more.
Loving those engineer discussions! My lingering doubt about the chiplet approach was always the interconnect bandwith bottlenecks, so interesting to hear how they work around it!
I really can't wait to see end user benchmarks, I really hope it's not a disappointment. I'd really like RDNA3 to be my next upgrade.
Man this is great insight on where amd is going for the next 3 years (or more). It lets me know how to budget for servers and desktop performance and what I need it for. This is one of the few videos you guys have that is bookmarked. Good stuff. This guy had to hold back talking, but it put out some good nuggets|vcabulary of info (that I can search in the underground).
I love how Steve is always way up on the conversation - understanding everyhtin and asking pertinent questions. lets the video go to great places.
This was so good I actually got to the end & was sad it was over, hope to see more like Steve mentioned!👍
Very interesting video, the chiplet approach looks hard to implement in GPUs but appears to have massive benefits. So far it has apparently allowed AMD to stay below $1000. I wonder if the RX 7800XT will use chiplets too.
they'd just have to remove 2 of the MCD's but are likely focused on doing the same thing as nvidia and trying to get rid of as much 6k series as they can since their AIB's also have a surplus of 6800, 6900, 6950's.
Navi 31 and Navi 32 are both chiplet designs, according to the leaks which turned out to be accurate about Navi 31. The 7800 XT will either be a more heavily cut down Navi 31 or a Navi 32, so it will definitely be chiplets.
As for the price, it's actually quite high. AMD will be making killer margins at those prices. If nVidia had priced their cards sanely, AMD would have plenty of room to move down.
@@sirmonkey1985 but on the bright side is that AMD is heavily discounting their gpus to get rid of them instead of milking them for more than they're worth like ngreedia.
@@TrueThanny NVidia probably cannot even price their cards sanely, given their monolithic design. I assume their lower end of the wiggle room is where AMD is with MSRP atm.
I wonder if the MCDs offer such a benefit that they obsoletify HBM. With that many memory channels and that much cache acting as an accelerator for the GDDR, the total effective memory bandwidth to the GCD may be greater than with HBM.
Amazing video. I’d love to see more. Is there a longer video in the works or perhaps the original presentation video available?
This was absolutely incredible. I’m not one ‘in the industry’ but hearing Sam speak at a level that even I can understand was really interesting.
Hope everyone at Gamers Nexus has a great Thanksgiving.
7900XTX for Me!!! Cmon December 13th!!!
good gracious, can't wait!!!
That guy's time is valuable. Thanks for stepping up and answering questions Mr. Naffziger. And thanks to Gamers Nexus for providing the content. Great job Steve.
These technical discussions are fantastic! Love to see this.
When I was a kid in the 1990s in Austin I had a friend whose dad was a senior engineer at AMD (they had a Fab here forever) and on his office wall he had like a 5' by 5' schematic of a chip he helped design earlier in his career. It was so amazingly dense and detailed I couldn't imagine how human brains could come up with it. Looking back now I wonder if it might have been a 29k chip. The house my friend lived in was very very big let me tell you.
I'm his daughter and remember seeing his whole wall of patents in his new office at AMD, and realizing 'oh, my dad invented a lot of things!' It's fun to see the hype over tech that I still don't fully understand. The STEM brilliance gene skipped me lol.
I will say though that we grew up pretty grounded! He grew up below poverty line on a ranch and literally got the letter of acceptance/scholarship to CalTech while living in a trailer park. He did *not* want us to be trust fund babies. My hometown had some other engineer's kids who had Audis and my parents were like nope you'll learn to drive stick :)
It's great to hear from actual engineers. It's much more interesting and gets me more excited about the products than the normal corporate nonsense they put out does.
Wish you could have had more time, so much to learn from this. Excellent questions from your side!
Loved this discussion, great interview, Steve, thank you!
I truly am cheering for AMD to be successful in both the CPU and GPU space. We need a strong AMD to bring out the best in the competition. For myself, I can't leave Nvidia because a lot of software is clearly superior on Nvidia because of CUDA. I use almost the entire Adobe suite of software and they are heavily CUDA dependent. I know that AMD is stretched thin and would rather focus on the bigger markets like gaming. But if they could make their GPUs fully competitive in the markets like video editing, it would make me consider AMD GPUs. I wish I could use a PC only for gaming but the reality is that my daily work also needs a fast GPU. Time is money but more importantly, longer times is greater cost. That said, go AMD! I'm still on Ryzen (rocking a 5900X system) and will be upgrading to the 7900X. But right now, I'll stick with my 3080 for now.
But why? 7900X is slower than a 13700k in almost every instance and the AM5 platform itself is more costly than Intels offerings(Z690/790). Most Adobe software still prefer Intel architectures, sometimes even by a big margin.(Premiere) Of course competition is important and necessary but this should not end up into users buying a lesser setup for their needs, just because they personally favor one manufacturer over the other. If AMD has the better offerings for my needs performance wise with also the better pricing, I go with it. Is it the other way around I go with Intel. Easy as that. No brand loyalty and fanboism needed.
@@Squilliam-Fancyson Adobe is actually slightly faster on AMD 7XXX than Intel 13XXX with the exception of the 13900K? We're talking marginaly faster, but they're absolutely not worse than Intels CPUs for these mostly single threaded tasks.
@@Squilliam-Fancyson I think you're the fanboy
@@MLWJ1993 Yeah yeah seems like I watched benchmarks with wrong numbers. They are actually all in the same ballpark performance wise. Still Intel has the cheaper plattform costs and overall slightly better gaming performance. As well as more oc headroom if engough cooling power is applied. AM5 Mainboards, are way to pricey atm, which makes Intel plattform more attractive. No fanboism needed for this conclusion.
I'm curious, what do you do that requires a fast GPU in adobe software? Most adobe software is very cpu bound in video editing like premiere and after effects.
What's wild to me is that for so long, the progression was to move more and more things onto the main die with the processor (even more so with mobile devices), and now we're finally going back in the opposite direction and breaking the big monolithic chip into pieces.
GPUs were always monolithic if I remember correctly. It was the CPUs which had memory controller and north bridge seperate. Those were integrated on 1 chip by Intel with 1st gen Core i7 in 2008 and were seperated again by AMD with 1st gen Ryzen in 2016.
Having the memory controller seperate from the processing parts is a 1st for GPUs with RDNA3.
Thanks, It's hard to get any kind of accurate information on anything these days on the internet. I appreciate your channel and candid reporting.
That was great interview and Sam did great job breaking down their tech.
People talking about Driver issues when it launches, like Bro, its not only AMD that has that problem. Even now, Nvidia is spitting bad errors with MW2 And even minecraft. None of them is exempt on driver issues.
Good video again guys keep it up
Wow! You watched 15 minutes in 1 minute?!
@@GamersNexus I think u guys r just that good
@@GamersNexus youtube just added the 15x feature
@@GamersNexus With GN's consistent quality it's safe to assume that every video will be good at this point. That said, I assume that was a comment to please the algorithm lol.
@@GamersNexus well I was watching it on my PC and posting that it was good at the same time. So yes. Multitasking
Love the talks with engineers with nvidia and AMD
Awesome interview and great job Sam Naffziger!
I think having you and AdoredTV here to discuss the chiplet approach would have been awesome. He did a video too trying to look at why AMD did the chiplet approach the way they chose to, and what it could mean in comparison to nVidia. Very interesting to see where this goes in time, if they do something like move more onto the smaller chiplets or not.
The tech media doesn't seem to like Jim. Suspect it goes to his early days when he called out much of their poor "journalism", which ironically made many of them vastly improve to the point the comments now regularly commend/appreciate that integrity. He should be thanked for changing the tide in-favor of the consumer, regardless of how peeps view everything else he does.
Latencey man, latencey
I'm actually understanding the basics of chip manufacturing! Thank you all for the in depth knowledge!
Excellent video, it was interesting to see the engineering point of view of those products design, i hope to see more of those in the future keep up the good work!
GN, STOP hyping up RDNA3, I want to buy a 7900 XTX at launch lmao. Just tell everyone to go buy a 4080 so there's more cards for me ;)