How this tiny GPU invented the Future 

This and learning about the history of AMD64, made me realize that they are an amazingly innovative company. Their stock price was around $2-3 at this time.

I still remember how surprised i was knowing that such a small form factor GPU had a tremendous power to catch up with GTX 980.

This, Zen chiplets, and RDNA 3 MCMs really make it clear that AMD is a company that bets on advanced technology with long lead times compared to its competition. All of these developments have been pioneered by them, often when they had significantly less revenue and cash then others, and didn't see a true benefit until multiple generations of refinement on top of the long R&D phases. It's crazy to think that AMD was close to bankruptcy so many times when they had so many advanced projects in development that are now being adopted by their competitors. If they had gone under, it's hard to imagine we'd have seen anywhere close to the level of advancement in performance and efficiency that we have now. It makes me wonder why Intel doesn't seem to have been developing similar technology until AMD started to see their investments pay off, when they had much more revenue to spend on R&D.

I remember doing repairs on a few of these. I had to do a core transplant on one once. It was easier to tell the customer to buy a new one and swap their painted shroud onto it, with their permission of course. I salvaged the dead core card for it's other components. I eventually tore the core off to see what killed it, and found that the interposer had cracked right below the main GPU die.

I bought the R9 nano in 2015 and still use it today. I love this little piece of hardware.

I love seeing these old interposer designs. I recently finished my PhD on interposer design and silicon packaging, and have worked with Intel on Foveros throughout it. Lakefield was their first product to have my name on some papers, which I'm rather sad about given it's performance, but it was a testbed for so many things. We learned a lot about how to handle multiple architectures in one chip, how to handle caches for those cores, refined earlier Foveros processes, and set the stage for it's spiritual successors, Meteor Lake and Arrow Lake. Alder Lake and Raptor Lake are a different concept, as they have monolithic dies, but also took some lessons from Lakefield as they are hybrids as well, with the RPL-C0 die being a hybrid of hybrids, having Alder Lake's L2 cache and Raptor Lake's L3. I really hope we start seeing more HBM packages in the future. When I saw the first Navi31 pictures from a pal in team Radeon, I was stoked to see what I assumed at the time to be HBM3 stacks, and 6 of them! I was so excited for another HBM card, and I still bought one when I found out those maniacs moves the memory controllers onto those dies instead. It retired a Radeon VII as my primary GPU, which itself replaced a Titan Xp. My hunger for VRAM still grows lol.

Allegedly UMC uses Canon steppers, and Canon makes some "wide-field" steppers which are meant to expose a full-frame CMOS image sensor without stitching and conversely have a 33 x 42 mm reticle. The Fiji interposer would fit into that reticle with relative ease (seems to be around 30x36).

Damn, this takes me back. This thing was an SFF treat, as well as the Fury X for micro-atx cases. Pretty much a 980 for SFF systems. HBM for consumers was a crazy play by AMD, but it really shows how much they learned from this, and improved their packaging technology. I remember all the threads and articles discussing, arguing about the bus width vs mem clock/size. Not to mention the many talks of a 'Titan Killer', how history repeats itself lol. Also, didn't know how long AMD was working on this too. While writing this, remember the Pro Duo/Fury X2? The Pro Duo would be an amazing collection to have with your Nano lol

I have an R9 FURY X in my system, bought it a couple of months after launch, I remember everyone mocking me for “overpaying” for a 4GB card, still rocking after almost 7 years with no signs of slowing down :)

My Fury X was my main card until somewhat recently, it had good performance, but AMD has stopped driver support, and even with modded drivers, it felt like you didn't get proper performance in some modern games. It's part of my collection now of ATI/AMD cards, their use of cutting edge tech is what prompted me to make this collection, HD 4000 series as the first cards to use GDDR5, GDDR4 and the ring bus architecture of the 1950 XTX, and so on, back to the 9700 Pro. Great video, something I'll probably show if someone was curious about what made my Fury X so special :D.

Thanks for the quality content. Keep up the very good work.

I've had R9 Nano, RX Vega56. And I'm watching this video using Radeon VII. R9 Nano was a crazy beast! This tiny card performed like a top level NVidia GTX 980. Vega56 was a very average card. There were rumors about 4 stacks version but both 56 and 64 got only 2. Radeon VII is still a good card because of enormous amount of VRAM and bandwidth. It's funny when modern GPUs struggle with low memory bandwidth but old Radeon VII has about 1TB/s of it.

This was an extremely well done exposé! What a cool story that I had no idea about. It’s just incredible that AMD was nearly bankrupt at this time, but were literally innovating well ahead of the industry in many ways. They deserve all the success they’re having now, just fantastic R&D *and* productization.

AMD certainly were ahead of the curve on this one. Now look at every major HPC/AI GPU, they all look like Vega only on more advanced nodes.

The late Vega 56 releases were the spiritual successor to this. They had the same small board and most coolers, like the power color red dragon, used the saved space for flow through fins.

I remember wishing these would me made into mobile GPUs. Without the need for vram chips you could fit a huge power delivery onto an MXM board which is about the same size as the fury nano pcb!

I had a Vega64 for a VERY long time that I got for £200 second hand and only fairly recently replaced it with a 6700XT. Great bit of kit that was!

Amazing video. Worth the wait. Going to have to watch this a couple more times for sure.

I was using Fury X for many years, since its release, to just only over year ago I updated, as it had simply too little VRAM. If it had 8GB I would be still using it. Now I am using 6900XT with 16GB, which should be good. Still I think Fury X was amazing design, and loved it. I do think HBM will come back to consumer cards eventually. HBM is used a lot in other cards, like professional GPUs, AI accelerator, ultra high frequency oscilloscopes, and ultra fast network switches and routers, where its performance is advantageous despite higher price. GDDR will reach its limits eventually, and only way up would be to copackage HBM close by.

Another well edited video with great research and content. And a stripped down Fury Nano!

This channel is a gem of knowledge, thank you very much, I happily subscribed! Great presentation too!

I have the Fury Nano, and only replaced it with a newer card (along with the rest of the PC) this year. Aside from memory capacity issues, the card still performed well. The computer it is in is now on loan to a friend who cannot afford to buy a new computer. I say "on loan" because I had the stipulation that when it was to be replaced, or it died, I wanted to reclaim the Fury Nano, take it apart, and pot the PCB into clear resin. It is the most interesting GPU I have ever owned, and always felt like it was a little slice of significant computer history

I used to have a R9 Fury, but I upgraded to and RTX 2080 Super, as some of the games I were playing started struggling on the Fury. I bought it because of the HBM. The technology seemed fascinating to me, and I wanted to see how it performed in real life.

This is a brilliant presentation. Thanks very much!

I have a Sapphire Pulse Vega 56 and an Asus Strix Vega 64 in my pc and my sons one. Definately surprised by the board size which is really short all the volume of the cards is their heatsinks. They are still decent in performance in games, especially the 64, and can be used with Resizable Bar, or SAM as AMD calls it, which helps for smoother framerates and much better lows with a simple Registry hack. HBM works more like cache rather than classic RAM ddr chips. It is so close to the GPU and has such a wide bus that I believe it deserves to be called cache memory and paved the way for 3DVCache technology we find on the modern Ryzens. Very nice video keep up!

I know this is gonna be a good content when i’m hear the jazz while disassembling

Great historical video on Fiji and its line of Fury cards. A few months ago, I retired my Sapphire Vega 56 Pulse from daily use as a 1080p card. In late 2019, I bought the used Vega 56 for $225 after tax. Power hungry and dumped a lot of heat at full blast, I love the Vega 56's compute power when I did video rendering. From my experience, the Vega 56 is my second most stable card after my first card, RX 580. Undervolted Vega 56 was a great card for my use case. With Vega 56, I got a taste of HBM. I can't lie; I want more, but we are not going to get HBM gaming cards for the foreseeable future. As a result, I want to collect the other HBM cards like Fury, Vega 64/64 Frontier Edition, Radeon VII, and Titan V.

Very nice and helpful review. Thank you.

I have/had all 3. I love the concept and that it is something different, so I went with it, despite the shortcomings. Still have my Fury non-x and my Vega 56, but sold my VII. But if I find one, I will buy it "back" to have them side by side.

Doated my sapphire furyX to a friend during the covid & crypto era .. its still going strong. great little card.

I’ve still got mine as shelf art.

I still own Vega 56. Not in my PC, it got retired for 6800XT due to lacking in performance and memory capacity. But it was really nice card, gonna keep it as collectible, maybe put PCB in some frame just to exposed the package with HBMs. Maybe it's good idea for also to hunt for dead GPUs with HBMs just to make some nerd keychains for myself, tho probably size would be bit too much.

IIRC the main downside to HBM is the strangely high latency, which never made sense to me, how could something so close to the die, have a higher latency than GDDR placed so far away from the die.

Actually Virtual Visions Finland did vram embedded graphics chip in 1994ish, it had 3d accelerator. It was built irl and was much faster than that era's other gpus. Some vga manufacturer bought em. But they invented the future..

High Yield : you forgot to mention that the HBM technology is a proprietary technology owned by AMD with pattened status. Nvidia uses it on Tesla cards but firstly experimentally used it on the Nvidia GTX Titan V 12GB HBM2 back in 2017. You also forgot to mention that anytime Nvidia manufactures a GPU with HBM, they have to pay AMD for the rights to use it and they have to buy HBM from AMD :).

Very interesting! Thank you.

I love how he got straight into it!!

Very interesting, i remember looking at those cards and i didn't get the point, thanks for explaining, very interesting.

This was my dream GPU at the time, the size and power it had! Just amazing would be nice if had 8GB!

When I see chiplet designs I always think of the chip they show off in Terminator 2 when they visit the engineer's house.

Great video, very cool. Got a sub here! Keep it up!

Interposers are used in other products such as Intel's EMIB in product like sapphire rapids and Apples direct die interconnect "UltraFusion" for M2 Pro Ultra.

Really interesting video sir

Hbm's death was a real bummer since eol hbm2 (titan V) to this day DOGS on modern gddr6x

It was doing a custom system that used piezoelectric actuator to slide the silicon under a sliding negative during prototyping this was changed to a a set of expanding and colomating optics lenses the reduced the minimum feature size. We have newer methods today

Buddy of mine in the US has the rig AMD used to tour and show off this GPU, the BS-Mods AMD Nano I-Beam, it's super cool

Well this explains where AMD's chiplet packaging of its CPUs came from. When it was first introduced people were surprised by such an advance and wondered why AMD had it production-ready before Intel the much larger company.

The HBM on Vega (or the card's IMC) was temperature limited. I got a ref. V56 from the first production run that hit European shelves, put it under water and flashed Vega 64 bios. That HBM has been doing 1045Mhz no problem since then. With the stock cooler, this just wasn't possible to get stable. Almost all people who also did water cooling, put a Morpheus cooler on their card or got an AIB model with sufficient cooling will tell you that they could crank their HBM clocks up by quite a bit. Later in the series, even with Hynix memory as live VRAM timings editing became a thing.

For AI, what you do is you have the memory distributed amongst the processing silicon that does the hardware multiplies. You need to store millions of weight values so you can use memristors that store them as analogue values. Tsinghua University has a working prototype. It's far more energy efficient.

I was looking to upgrade my graphics card back in 2019 I think, and I was looking mostly on the used market for Vega 56 as well as GTX 1070 tis. I found a good price for a 1070 ti first, so never got to mess around with a Vega card, which would have been interesting. The Fury cards, with their HBM memory, kinda blew my mind at the time. It got me excited, but then AMD pivoted back to regular memory again. I have since upgraded again to a 6800 XT that I got on the used market, but I am still curious to know how the Vega cards are aging these days against the Pascal and Turing series.

I remember the launch and the reviews from jayz2cents, linus, GN, bitwit, and Paul's hardware. The cost of that HBM was what made the R9 forgotten, around $200 higher than the 980 Ti. It just didn't make sense paying more for a card with the same performance and 2gb less vram. I'm still rocking a vega 56 I got second hand in 2019. It's slowly showing it's age and it's next life is going to be running my couch gaming setup once I get my hands on a newer card.

I had one. It required an undervolt to get working reliably due to thermal constraints. The memory could be overclocked. I shipped it off to a friend in Canda when I upgraded and unfortunately the drivers are too old for him to use it.

The best and worst thing about the Nano is that it was greatly power-starved, and nothing made that more obvious than it being the only card of its time that could easily run Furmark without any risk of thermal throttling, because it just couldn't ramp up enough to overheat in that application.

I am reminded of the Hybrid Memory Cube, imagine if that technology won the memory market.

I used to have an R9 Fury (which died during OC) and a Vega FE onto which i had to mod a Raijintek Morpheus cooler. Cool products, held back by the GCN architecture and, in case of the Vega, the cooler. Seriously, with the stock blocker, it wouldn't run it's full boost, as it would quickly hit 95 on core, 105 on the HBM2 and like 115 junction. Tuning and undervolting made them much more usable.

I still have my old Vega64. I still pull it out every now and then because for high memory intensity workloads it *still* stomps around like Godzilla.

Great video.

I am using fpga with hbm technology for few years now. It was available for intel's(altera )FPGAs and XILINXs , aquired not so long time ago by Amd.

i bought a arez vega only for that tech marvel , using it in a x79 platform as a pcie pass-through for win10 vm; also the possibility to cache system memory to increase video memory is another benefit for the 4ch ram workstation boards

The R9 Fury was the first and only AMD GPU that I ever wanted. Tiny form factor and integrated water cooling. But of course I didn't have the money. Is it a good idea to find a used one today just for collection purposes? I assume the water cooler must be not working after such a long time

I have one of those and 2 liquid cooled Fury's. Worth every penny when they came out

It is an interesting use of 1920 vintage VRM.

Vega56 user user here. That thing is still running my main PC, underpowered to ~130W, in passive mode 99% of the time. When it gets into retirement I'm going to open it up so it can shine some silicone on my bookshelf :)

Great stuff

if gaming cards do switch back to HBM it will probs be a lower generation of it as the newer stuff will be reserved for Ai cards

Nice. That was so cool. Didn’t know AMD was just a pioneer

Great video

You're right about GCN. It was a good architecture at the start, but it just didn't scale with compute capacity. Just compare the Radeon VII with the 5700 XT. Both are very close to the same performance in games, but the Radeon VII has 50% more compute capacity, and even more memory throughput (it has four stacks of HBM2 delivering 1TB/sec of throughput). RDNA is massively better for gaming than GCN.

very well done video

I recently upgraded computers for few of my familly members with used Vega 56s. Even I still own vega 56 in my pc (actually downgraded from 5700xt because I don't really need that much performance) and for fun got one of those weird Chinese versions. Pcb looks like it was made with some weird parts but hey, it works. It took me a while to tune it just right to not draw too much power but still do it's job.

Clockspeed is a factor as well, latency is a big factor. Hbm's biggest Achilles heel was the increased latency compared to gddr, but bandwidth for sure was there even though it was still not enough with only two hbm2 stacks for vega cards which was proven with r7. Fiji was not fast enough for the large bandwidth of 4 hbm stack but the latency was what dragged down the perf of the fiji cards. Even vega cards would see a healthy perf increase with oced hbm2 even though the theoretical bandwidth was faaaar above what it needed.

Problem with hbm is you can't repair it when its gone while reballing or replacing ddr modules is relatively easy. But i do believe that hbm will be the way forward as integrated graphics become more powerful and more power efficient than traditional pc-s.

0:40 a vapor chamber that goes to 2 u shaped flattened heatpipes? Interesting

10 years ago i was seen a box of an gpu on 1024 bits long bus. what video card was that ? or how much cost it then ?

Still running my Vega 64 with full cover water block, with a 5950x, runs everyting I need it too. Got on release, XFX reference card. Maybe someday ill upgrade, but right now still happy with it..

I got one of those. It competes handily against the RX 580 performance-wise. I probably need to replace the thermal paste, but I'm terrified of damaging those chips.

Is there something similar available for partable mini PC in this size?

I still have 3 of these and a fury x. great cards

I got two Radeon Pro Duo's. Their PCB's are a joy to behold.

I wanted to get my hands on a Radeon VII so badly, and it really struck a nerve when the editor-in-chief of Rock Paper Shotgun used it as a doorstop.

wow. what a solid GPU. and smol

Could you make a video about flash memory? I tried to find out how the memory block is connected to the controller (i.e. with how many connections) but could not find anything.

I'm still using the r9 nano with modded drivers and it's still give me a satisfaction when play game in 1080p !

3D stacked V-cache plays a bigger part in consumer gaming GPUs. When that stops providing a benefit then HBM will be looked at again ?

I think that the von neuman architecture where memory is separated from cpu will one day be replaced with something where memory is part of the calculation path, like do the neurons. My bet, let's see.

Highly informative video! Keep up good work. I did find however that the use of graphics/diagrams during your delivery, often with packed with lots of information unrelated to what you were touching on at the time, distracted me away from your delivery. Perhaps you could consider less info-rich graphics or coming up with your own diagrams to better complement what you are saying?

Been a fan of AMD since i bought s sempron wich was a Barton core that was on the athlon XP an fixed all the laser cut cache back an used a pin mod with wires in the socket to get 2.50v. Was an insanely clockable chip was à 1.6ghz that clocked to 2.9ghz on air. Boosted performance massively, not bad for 40 quid. Sought out the certain Samsung ddr1 memory on brainpower pcbs that could hit 500mhz, was a killer rig that bairly cost me 250 quid with a 6600gt that was a really good clocking chip too was almost as good as a stock 6800.

Heeeeey, I got one of these in my backup PC. These things still sell for up to 250 on ebay and are sometimes not available at all. It's worth more than my primary 5700xt! I love my little Nano, I wish HBM was still a thing in GPUs. The Nano was the neatest GPU I have ever owned

FYI it's "WHAT will the future of video memory look like?" I notice this error in a lot of RU-vid videos and I'm not sure why but I though you might want to know

Which is why all gaming GPUs now use that setup

Ive used AMD (amd+ati) exclusively from 2004 till 2021(when i had to buy a laptop amd models were not widely available in my country). They are a great company very affordable too.

I remember being sad they dropped HBM, HBM2. and HBM in general, despite actually having mostly forgotten about it until it was mentioned here.

My question is why we donçt see more cpu-like high level cache in gpus. Surely it would have similar results, right? Then pcb memory chips for mass call storage.

Ive got a R9 Nano. Cool little gpu. Idk why they dont make compact cards like this anymore

And now graphics cards are chunkier than a 1980's answering machine.

Its worth saying that amd start developing infinity cache about the same time. It was originslly made for APUs but that turned out to be a bit to hard. Specially now when amd made sort of the comrpmise solution of cache at the memory controller. Getting sort of the best of both worlds.

for the current memory problem something like HBM is a good solution, and for AI you might actually start looking more at something like a FPGA, but also with some new form of memory directly build into the compute logic itself, essentially AI often uses many weights and such, now they are stored in ram, instead it could be stored more analog or paralel like actually having logic weights buffering memory, where you can even buffer weights directly to logic, kind of use that as ram, this could allow insane speed increases. with a fpga this could be simulated early on just on a small scale to test it. but essentially when done right it is ideal for AI since then it doesn't really need to compute it anymore, instead it just knows it, and it is kind of like analog but then abstract and perhaps digital still but it is like analog as in more like a quantum computer where you insert values/changes somewhere in the logic itself. perhaps inserting it to a special core like thing which decides/manages where it needs to be input and then doing that. this essentially means that the ram no longer needs a speed and just is direct basically, and instructions are not given but instead written to the logic which also is the ram, this means the main bottleneck is a possible clockspeed/settle time and the speed at which the instructions can be handled, probably mostly the last one which right now tends to not really be concidered a problem.

One of the problems with HBM is that they cannot be repaired, if one memory chip falls off, the entire crystal can be thrown into the trash.

I wonder if they are cooking up something similar to this. Making a product with more forward thinking technologies that is probably way ahead of it's time.

Gotta cherish my shelved R9 Nano even more