Cost of HBM2 vs. GDDR5 & Why AMD Had to Use It
HTML-код
- Опубликовано: 20 авг 2024
- Talking about how much HBM2 costs vs. GDDR5 and AMD's tricky market position, ultimately requiring HBM2 over GDDR5. Note: Bandwidth should read 409.6GB/s for V56, not 484 (V64).
Ad: Share your keyboard+mouse with Synergy symless.com/sy...
Article: www.gamersnexus...
We have a new GN store: gamersnexus.sq...
Like our content? Please consider becoming our Patron to support us: / gamersnexus
** Please like, comment, and subscribe for more! **
Follow us in these locations for more gaming and hardware updates:
t: / gamersnexus
f: / gamersnexus
w: www.gamersnexus...
Editorial: Steve Burke
Video: Andrew Coleman
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.
Article here: www.gamersnexus.net/guides/3032-vega-56-cost-of-hbm2-and-necessity-to-use-it
Note: Bandwidth should read 409.6GB/s for V56, not 484 (Vega 64 bandwidth). That was based on pre-launch info and should have been updated.
You might like clips from our latest Vega 56 Hybrid overclocking livestream (but we are working on a TLDR): ruclips.net/video/0fc1wti2EG8/видео.html
Gamers Nexus some charts are wrong. The bandwidth of Vega 56 isn't the same as 64
Old chart from before launch, based on pre-launch spec sheets. Should read ~410GB/s for the latest figure. Thanks!
I recall that the motive for using the HBCC controller in the first place is that AMD could make it talk to HBM v1, v2, GDDR5, or GDDR5X. That flexibility is what they talked about in the early press briefings when they first went over the Vega architecture.
Yes, that was one of the early reasons for the HBCC setup. Didn't seem to really pan-out that way, but we did have that conversation with AMD either at CES or before then. HBCC is a protocol, not an architecture, and Vega only has an HBM controller. Oh, and Cataclysm, were you the one suggesting PWM testing on the monitor? Turns out that the Samsung display doesn't use PWM to adjust backlight output and uses DC voltage instead.
Thanks for the video, the research and the hard work behind the scenes, GN. I'm a recent viewer and, as a geek in my early thirties, have gladly eaten humble pie to see that you're the real, hard-core deal and not the "long haired hippie millennial hipster tech wannabe." :-)
Nice to see a video with a simple, non-clickbait title that delivers exactly what it promises. I now know what HBM2 memory costs and why AMD used it. Thanks!
But why ?
Do i need to watch the video ?
AMD - Awesome Memory Dudes
GoodDay :)
AMD - Advanced Mining Devices
@Subi_fan what do you think about AMD now? (ok maybe the drivers succ)
@Subi_fan nice :)
edit: yeah some gpus drivers suck
All of this misses 1 really key point... Vega pulls a ton of power for rather average performance... RAM issues aside, the power draw of Vega feels like R9 290X all over again, and in 2017 that just doesn't cut it anymore...
AMD needed to fix that issue, nothing else is really going to address the problem short of that. Whatever allows Ryzen to have such low power consumption per unit of performance needs to go into Vega. My understanding is that Vega didn't get the same design team benefits that Ryzen got, thus the insane power draw, similar to the FX chips. Perhaps Vega 2 next year can get the Ryzen treatment, if so, good times...
I'm on AMD's side here, I WANT them to do well, but they are in a rock and a hard place... Vega costs too much to make to sell for what it is worth.
They have a Balanced power mode that addresses the problem. Simple as that really.
Only thing that might change now is Performance via driver improvements.
"Vega 2" = Navi ;)
@boothegoo pc do you even know how long development of a new gpu takes? Anything coming out will have had 5 years of development time.
I dont get it.. I have vega 56 with undervolt it's 220W card for 1080 performane..
AMD says it's a long term investment. And they also call it HBC. And also Raja Koduri once hinted that the technology that AMD used for Ryzen CPUs to interconnect dies(Infinity Fabric) is also possible with GPUs. But if they could achieve this with Navi architecture which is also focused on "Scalability" as their road map shows, the HBC might in fact be far more essential than it is today with Vega. And I don't think that Raja Koduri picked HBC as the most impact technology in the long run if not for a reason.
This aged so well
@@ILoveTinfoilHats Gonna age even better when RDNA 4 comes out with multiple GCDs
6:06 The card itself starts shaking from seeing how expensive it is.
I don't think hbm2 was 'forced' for RX vega. The way I see is that RX vega's are left over chips that didn't go to the more profitable professional market sections or an OEM contract. Since hbm2 limits the total amount of chips you can make; if you were a struggling company would you rather sell it as a $600 card or something like a $7000 radeon pro ssg? I think the hbm2 is just the nature of the beast regarding these chips, the high end gaming market isn't going to bring amd the money they need so they developed a high end compute chip and sell those that don't make the cut as RX chips to recoup the loss.
The elephant in the room with Vega is compute performance. In Indigobench it competes with the 1080ti/Titan X. For those of us Using scaling GPU render engines that makes Vega the best bang for the buck. Period.
Vega forced Nvidia to fully unlock Titan X =)
Ya, only if you consider how many of those who buys gpus ACTUALLY render something and making content, and not just game. Almost none, and their PRO series for that specific task, RX vega series is a consumer grate gpu, that means for playing games on it.
Playing games while still excelling in compute is what RX Vega does. The gaming driver part keeps on improving bring RX Vega 64 especially the LC dangerously close to GTX 1080Ti and sometime beat it. This Fall will be so exciting to watch.
If AMD wants to sell me a "gaming" card for $500 that I then use for compute applications because it outperforms the Titan X which costs $1200 why should I give a shit about the dissatisfaction you feel because you are a gamer and don't benefit? Why should I give a shit about their Pro series of cards when the only Vega variants are functionally Vega 64 with "optimized" drivers I don't need? I don't use Autodesk or Solidworks applications. I'm good. Worry about yourself instead of being a fanboy worried about what peers and reviewers think.
close to 1080-maybe, not to ti. 1080 ti is out of rx vega reach
Thanks alot for the power consumption & price data.
I had no idea GDDR5 could consume so much power. And furthermore, that HBM consumes so much less.
why do you think the 290x was such a power hog?
GN is my go to for any and all tech news. They do a great job breaking the info down and being unbiased in reporting. Great content and video as always!
Watching back old content is really interesting to see how the channel changed and grew (now with better visual presentation, a more refined style, and a faster cadence among other things), and how some stuff are still the same: Highly educational, and how you cross-reference your infos and ditch unreliable infos (the GPU die cost).
I'm still rocking my trusty ole' Sapphire Vega 64, and I'm still surprised at how it can handle things, even at 1440p. I'm even able to run Stable Diffusion quite well and I keep wondering if the HBM2 plays a role in that.
Also, remember when the world was freaked out at Vega 64 energy consumption ?
Bought a vega 56 for an older pc I had for $100 and it's amazing. I barely hear about these cards anymore.
Good on AMD. HMB isn't a fail. AMD is taking the hit, and losing out on profits to further improve their architecture as a whole, to help themselves AND the consumer. HBM is where we're heading whether we like it or not. Sure it's expensive now, but prices will drop, and AMD knows that. GDDR5 is on its way out, and AMD's just future proofing.
Fail prediction
Literally no other TechTuber can even come close to the quality & depth of this content.
I love the depth you go into on your reviews and articles, good stuff helps me stay on top of a hobby I no longer have much time for.
Videos like this is why I love Gamer's Nexus!! AMD have miscalculated the price depreciation of HBM2 memory. On the plus side the miners are helping AMD keep its head above water until SK Hynix starts mass production of HBM2 memory, eventually driving the cost down.
And now, a couple of years later, the situation is somewhat reversed. nVidia uses the more expensive memory while AMD doesn't have to.
Am I the only one who sees the humor in that?
It's more sad to see that HBM has failed, it was a great technology
@@davidyang9902 it hasnt
The voice on those olds videos and the writing is much more composed.
Few factors not mentioned - HBM memory controller is apparently smaller and simpler then GDDR5 memory controller - meaning that AMD can therefore make a smaller die. And also since the memory is on package it lowers the cost of the PCB (routing all that memory chips) somewhat. So there are some savings to offset the hugely higher price.
Honestly, I think they are right to go that route. Unfortunately being the underdog means that no matter how good your architecture is, if it isn't aligned with the leader architecture, performance will always be lacking due to the lack optimization. While it is a problem that is unlikely to be fixed in the game industry, the productivity area sees more and more alternative software that can make use of GPU computing power. Banking on it is I think a great long term investment as it is bound to happen, especially if they refine it power wise for the laptop market.
Currently there are no pure productivity alternative in that area where you either get a low core/thread with a good gaming gpu or an average core/thread still crappy cpu with integrated graphics. If ryzen get ported to the laptop market with 6/8 cores variants with different models ranging from integrated graphics to vega powered gpu , AMD will get a clear win on that Market against its competitors.
Intel getting into the GPU market is making more sense now.
One thing not mentioned is that it's necessary to push for the development and production of HBM if one wants the cost of the tech come down in price anytime soon. Such a reduction in price would be important because HBM could be AMD's ace-in-the-hole for APUs, especially if it could double as system memory on 7nm Zen or the like (super low-latency memory anyone?).
Oh, and also it saves die area... 64b GDDR5 bus estimate 6-8mm^2. 1024b HBM/HBM2 bus
Its great to see how in your previous vids your like" hmmm hbm2 why maybe this maybe that" and next post you go out and get the answers. Its like hopping on the memory mystery bus with you bud. Keep up the good work agent burk
Your passion fuellls the nexus!!!
0:18 Lol, the good old days of cheap prices.
Well, AMD's GPU engineers should learn from CPU's engineers.
I subscribed to you because of this video. Not many PC tech RUclipsrs give this kind view of the video card market. I was getting tired of seeing the oversaturation of standard benchmark videos of shitty music montages.
It would be fun to be able to overclock the HBM to hell and back (I'm talking somehow increasing the voltage that goes through it), and see where diminishing returns start to show up
1100mhz, vega 64 has unlocked voltage
V64 has 1.35V on the HBM2 it's not unlocked voltage control
Dimishing returns you mean each module will start to explode from heat lol
I love your content man, more than other techtubers you do tons of research and i like how you broke down the compenents and calculated the power. Keep up the good work. I support vega we need competition in the market
I kinda wrote off HBM2 before this. It seemed like gddr6 was the future, but now I really want a 1180 or 1180 ti with HBM2. I'd upgrade from my 1080 for that. The lower power draw could improve clocks.
Excellent report Steve. To paraphrase one revelation "You think Vega burns through a great deal of power now watch what you get if you re-engineer the board with GDDR5".
With this report in mind now for the Follow -On questions. What does the future hold for AMD GPUs ? I'm specifically thinking of Navi. The language used by AMD in their statements leaves little doubt that GDDR5x being used on any future iteration of Vega is out of the question. Does that logic extend to using GDDR6 on Navi ? It will most likely be an entirely new architecture.
The only way out of the stew AMD has cooked up for itself with HBM2 Memory is for them to produce a GPU that blows out of the water anything and everything in the way of GPUS NVIDIA has on the market. And then price it according to cost which would put the price figure into 4 digits. But AMD is no where close to the 1080Ti in performance right now. No matter how you look at Vega the card shaping up as a GPU with the potential gain the reputation the Edsel had as an automotive product for Ford Motor Company.
Cost per die candidate coming from GloFo is probably between 30-40 dollars. Yields, binning, shipping, et cetera will drive up a lot obviously :)
Hence huge cost difference...
Great video, didn't know some of this and it's nice to hear it broken down into details
@gamers Nexus This was a fantastic and interesting video. I'd be interested to see future "cost breakdowns" of other products, and logical conclusions based on those to give insight to design decisions. As always, keep fighting the good fight. You guys are amazing.
So basically the vega architecture is at fault and not HBM 2, right? (
Yes.
Yes.
Epic Listening correct. Given how Vega seems to be bandwidth starved, a 3072b version would be far better, I think. Plus a bit of cost savings being able to use 6GB
Pretty much.
You can think of it as a big F-350 truck, but instead of a 120-litre fuel tank with the fuel being gravity fed through carburetors, it has an 80-litre fuel injection system. Smaller fuel tank means less weight, more efficient travel. Except it's still an F-350.
Exept that there less than 5% of total selled cards will be used for actual content creation. PPL buys gpu's mostly for games, and it should be optimesed for one, with good drivers and low power consuption, deal with it. For pros there pro series of cards, and they have nothing to do with RX vega. More raw computing power in consumer grade cards, that can't be fully utilised in games means a good deal for miners. Heck, there more miners nowadays than content creators.
Hmm now that you've mentioned that HBM2 is more power efficient, I wonder if they can do a Polaris refresh with HBM2.
GTX 1070 never dropped in price for the MSRP unfortunately, it SHOULD have but I don't recall them ever doing it. It's the reason there's a $200 price difference between 1060 & 1070
i just moved from 980ti sli to vega 6 sli. I know i know its stupid and the gains are minimal. But i also have a Rx fury on another computer and i was very happy with the support that card received in long term driver updates. It slowly crept up on my 980ti s . Here is hoping vega gets the same treatment.
You can easily estimate the die cost for the Vega chip using the cost per sq. mm of the cheapest pentium from Intel (4400?) minus some sort of below-average markup but scaled up to the area of the Vega chip. The Intel chip will be a little cheaper because it will have a higher yield, but I imagine that GlobalFoundries is killing it with their 14mm process as the Ryzen chips are rumored to all have very high yield and good overclockability. Anyway, this gives you an estimated floor cost for Vega GPU (manufacturing - not including design costs).
Vega needs a faster memory bus because it is bandwidth limited. As miners already know, you can run the core at 1100mhz, and it works BETTER. Up memory frequency and Vega kicks ass! I would say with a 4096 bit memory bus Vega would be insane.
The problem is why is the VEGA architecture draws so much power? It kinda looks like a derivation from their past GPU architectures. And imagine if the 1070 had HBM2, clearly it shows just how much more efficient the Pascal architecture is.
anyone just randomly curious where the performance in terms of gaming would have landed? I mean think about it for just a second if the card was 100 cheaper because of nerfing it with GDDR5 vega 64 could have been priced like a 1070, would GDDR5 be enough of a performance killer for 64 to still lose to the 1070 in that particular scenario, if we just treated Vega like Polaris and ignored attempting for a high end card how might it compare, theoretically speaking?
I don't doubt that AMD needs HBM2 to get the bandwidth to feed their GPU and power savings.
But just image the GTX1080 on HBM2. If the GDDR5 and 5x cards suck up 50-75W of power of the card's budget the GPU die of the 10 series only eats up ~100W. Crazy.
I know it's a different controller, power setup, etc. But it would be cool to see a 1070, 1080, 1080 Ti, v2 that features HBM2.
That would be a cool refresh for 2018.
I wonder if there's a possibility for a lower clocked vega chip, using gddr5, to offer something between the 1060 and 1070?
I recall GN reporting that vega can be quite efficient at lower clocks, twinned with gddr5 at a 256 bit. Could be interesting?
In the video he says both the 480 and 580 have the same tdp but when i checked bith reference cards it said the 480 was 150 while the 580 was 180 anyone got an explenation
I just hope for cpus with HBM L4 Cache. Something like 4 or 8 gigs (according to the processor) as a faster pre-ram
Just hope that doesn't lead to a slot 1 situation
Man idk the way you explain stuff and what you explain is pretty cool
Great explanation as usual. Heh, I'm getting so good at this from watching you, RGT, adored, and Bullziod I was finishing sentences while watching, and understood everything. Bullziod still leaves me befuddled sometimes though.
GooD fuN!
They could also solve the problem by making Vega architecture more efficient overall, so it wouldn't need so much memory bandwidth and power consumption on the dice in order to compete with Nvidia. But I guess that`s out of their legue...
how did they plan the power consumption years prior to release? or did they think 'let's overclock our mid arch as much as possible by dropping memory power'? what about gddr5x?
Vega is a stepping stone. Setting HBM2 as a standard in their gpus will be required for their long term plan for Navi with NexGen Memory. Going crossfire then with infiinity fabric should be able to scale the (say 2 of them) gpus about 95%+. Yes, that's actually mostly accurate on a theoretical aspect. Practical though we do have an example: Ryzen. Look at Threadripper, it's ridiculously powerful AND cheap o_O
AMD: Just Wait (tm)
Good food for thought. I really hope it will pay off though (like in the CPU market, we do need the competition in the GPU market).
I do think that Vega is a good design, just poor tuning on it's reference model. Hope that AIB partners do some fine tuning on these Vega chips and get its performance and efficiency back.
With the low power draw/space per performance for HBM2. Do you think this opens up some opportunities for efficient gaming laptops? Also with the addition of "radeon chill"+ cheap freesync screens they might be onto something?
Research and Developtment costs of HBM,the Exclusive production with Hynix
Power Consumption
Form Factor
Memory Density
Bandwidth
APUs
I love the gamersnexus in depth analysis on architectures it helps everyone understand so much more
Considering Navi is expected sometime Q4 2018 or H1 2019 I'm curious if Vega as an architecture or even broader GCN/NCU are a mainstay. Or if Vega as a whole will be a replay of TeraScale 3, and the swan song of GCN. (more of a goose song though..)
TeraScale3 was an architecture that only saw 2 high end discrete SKUs (flagship and cut down version, much like 64 and 56) and limited use as mobile GPUs (which will manifest as APU iGPUs in Vega's case). And was the end of 5 year major overreaching architecture. Well it's been well over 5 years for GCN.
The 6950 and 6970 were far more competitive too, yet still had poor long term prospects, if Vega (& NCUs) see the same limited scope it's long term prospects will be rather poor.
especially if Navi is a drastic architectural shift, which it rather seems like from the limited info we have.
Unless NCUs are a forward thinking architecture more suitable for the supposed modular "scalable" interconnects of Navi.
if not "fine wine" will be a blazing ball of nope for Vega.
Also it's funny to consider ATI were also ahead of the curve with GDDR5 being first to market by a significant timeframe if I recall.
So, Hynix failed its time-target to provide HBM2. Some people say for long now that Vega delay is most due to this, than anything else. Drivers aren't ready either, but that's classic for AMD new arch. Vega 11 which will succeed 580-570 will have GDDR memory imo to lower costs.
I know the vega cards are overclocked out of the gate, but whats the clock speed they would need to run at to hit the efficiency curve of the 14nm lpp. I would love to see vega locked at that specific clock to see the performance delta and power difference.
How effective are 56 and 64 with productivity compared to rx vega frontier? You think you can do some benchmarks to see how these 3 cards stack up?
Great analysis, thanks Steve! I'm sure it will play out well for AMD in the long run... for now, gamers will have to settle with slightly less-than-desired raw performance and low quantities. I just don't understand how the Vega architecture is so power hungry that it needs HBM2 and can't stomp the GTX 1080? Even in the most AMD-optimized games, Vega 64 doesn't pull that far ahead of the 1080. I guess we have to keep saying the same thing as always... hoping for nice gains from driver optimizations and game optimizations as well.
It will be interesting to see how HBM2 fairs against GDDR6 in real world application, e.g. how well nVidia does with it regarding their first architecture to use it.
@2:45 Not if you take potential losses into account if it drives away those consumer who just want the watercooled graphics card.
HBM cracked me up where it said " Low Power" (3:31). =)))
Could that David Kanter estimate be targeted at 8hi stacks? I thought it was an older quote and Vega did launch with 8hi stacks. Vega 56 and Vega64 however use 4hi stacks, don't they?
Saying that Vega architecture draws more power is heavily inaccurate statement.
If 1070 or 1080 had unlocked voltage and 1.2V put to the core they would most likely draw the same or more then V56 at same game performance.
Vega is pushed to far with voltage, for no known reason to me, and maybe couple tens of mhz too much to make it look better, which failed unfortunately.
Hello! AMD named it's APU Kaveri because Kaveri is the widest river on the planet, hint-hint- it's NOT that AMD HAD to use HBM2, it's that they were designing their products FOR HBM as far back as 3-4 years ago,. AMD has had a long term plan as far back as 3 years with all implementations beginning to converge in this generation of products-( AMD/Hynix designed the interposer 2-3 years ago that's now used in Vega and soon their APU's/HBM2). Do the research.!
Great video. Love the straightforwardness of this channel.
Excellent video. This is something I've wondered for a while now.
Isn't HBM central to the memory cache controller's strategy?
I seem to recall that the strategy requires low latency, something GDDR memory cannot provide.
So aside from all the power arguments, AMD's memory cache controller would be ineffective with GDDR5 making cards like the SSG little more than a graphics card with an SSD on board and the memory cache just a fancy marketing term.
Very interesting news..on point & to the heart of the matter .
Whats crazy is that other tech tubers will continue to bench against VEGA although they are now $675 cards - technically irrelevant at that price .
Kenny
Joe macri said the hbm development started inside of AMD ...!!! That was a article of 2015 and they started it 7years before that in 2008 ;-) making this sort of innovation happen was a broadly collaborative effort. AMD did much of the initial heavy lifting,designing the interconnects,interposer, and the new DRAM type. HYNIX partnered with AMD to produce the DRAM, and so on.... so if they started it they cant just drop it ;-) and now it's 2017 so it's allready 9 years in process.
HBM2 enables some features on their workstation cards like the SSG.
If i was Amd i would just leave the high-end gaming GPU market.
Polaris was a good sign from Amd as it was actually efficient for the performance it offered
Here's my question for you guys, what should we expect from the 'Raven Ridge' APU family? Will the HBM payoff there? How many TFLOP do you expect on the IGPU? Will we see 1080p high with 60+FPS on popular games with it? Very interested in the Zotac MA551 myself but not sure what to expect.
Fugasnaya we won’t see HBM2 or any memory on an APU... APU’s use ram as shared memory
You're right of course. But the thing is there were many rumors that suggested there would be a version with HBM2 on die. Looking into it now I can see it is not confirmed, but I saw this rumor so many times I stopped thinking about whether it made sense lol.
Fugasnaya Just from the size of the HBM2 chips compared to the size of a RYZEN die, I can say they would have to take out at least half the cores, with the integrated GPU, other half, so we would end up with a 2 core CPU with a crappy integrated GPU and a base price of 175$ just from the HBM2 cost. I doubt anyone would pay almost $300 for a 2 core APU...
It will be the *Threadripper APU* !
There there is plenty of free space on that Die for *4* of those HBM2 cubes.
LowLightVideos Thought it would be a normal RYZEN APU, because if you have enough money to buy a threadripper motherboard you should also have enough to buy a decent GPU. I'm not very well informed of the latest tech news because I spent the last few weeks on vacation without internet.
I know that Radeon Technologies Group really fucked up many parts of Vega, but I kind of admire them for actually trying to push the technology envelope regarding new technology (HBM2, high bandwidth, mantle (Vulkan)). But I really can't accept that they are investing so much in all of these technologies while they are screwing up the basics. Even so, If Vega 56 is stabilizing at the 399$ and AMD can crank out another 5% + in performance from driver optimization I think every consumer is doing themselves a disservice if they chose the 1070 instead. In the long run, this is what would be best for all of us consumers. The worst thing that could happen to us as consumers is if a company lands a to domination market position like intel for the last 10 years and what Nvidia is about to acquire. Then we very probably will see incremental minor gains from a slightly improved base architecture but for ridiculous prices. On top of this, the frequency for major technology jumps would most likely decrease.
This is a very interesting video. That's too bad that the architecture decisions during initial development forced AMD into a corner too far down the line.
Would 352-bit or 384-bit gddr5x provide comparable bandwidth without requiring too much more power? I now understand why gddr5 wasn't good enough, but not why gddr5x couldn't have been an acceptable solution.
Why would Ford use rubber tires instead of wood. Wood cost less and works fine. The real reason is the car weighs so much now it needs Rubber as wood would not hold up, nor would it be able to get as good gas mileage on wood. The world should be using wood - FOREVERRRRRR! And the other companies are not using rubber so it must not be a good idea. If the other companies used Rubber we would have to pay more even though it might be better for the Industry. The manufacturers would have to raise their prices to use Rubber, why? Well, because I said so.
Why not 4 stacks of HBM @ over 800gb/s? How much performance could it get?
i thought hbm provided higher bandwidth and that meant it didn't need as much memory. If so will wouldn't 4gb hbm2 560 make sense?
The bundles make the cards more attractive to miners as it actually makes them cheaper. They get to sell on those extra products as brand new sealed items.
what circumstance would make VEGA get higher frames per second or lower frame times due to HMB2? i'm currently working with high resolution textures in unreal engine 4 which make the engine use more VRAM than I have. would this make a difference? I would test this myself but I don't have a bunch of GPUs to test with.
This was really informative Steve(your name is Steve right?) , any chance you could compare the dueling architectures? I'd be interested to see if Nvidia actually has something architecture related that makes it render video games faster (as in more FPS) or if Nvidia is designing it's cards to take advantage of how games are currently designed
Honestly when the hell are gpu prices going to be normal. I sold my Rx 470 in June and still am selling RX cards to this day. Its great for me but I wish I could just buy a gtx 980 ti for a low price.
at one point you refer to the GPU as the ASIC - that's highly inaccurate
ASIC: application specific integrated circuit
which is perfectly applied to GPU because graphics are pretty much specific application. GPGPU changes a lot ofc, but still.
Anything about GDDR6 debut ? My guess will be Volta.
So, if the Vega core is starved for memory bandwith, that means that the core was really overbuilt and AMD will need just small improvements at the time and much faster memory each generation to have better GPUs? Unfortunately is impossible to overclock HBM2 that much but I would've like to see at what total bandwith there are diminishing returns or no gain in performance.
well vega 56 memory is bios locked to low voltage (artifically lowering the performance of the 56), but on 64, where its unlocked, you can hit almost hit 1150mhz on the hbm2 clock which does give the core enough bandwidth
Thanks for the reply, I wasn't aware of that. Can you link me to an article with performance using 1150MHz on the HBM2?
Also clearly r and d for the next gen consoles as its low power suits them perfectly.Paired with a ryzen and this will undoubtedly be the next gen console.
No mention of them not having the budget to develop another gpu arch, they use it because the they can only afford to make 2 chips the polairs and vega, they cant afford to make a chipset for just high end gaming like nvidia. Your points are valid but miss that major piece
These are some of the best videos for this category. Thanks for the insight!
Wonder if the hbm use will increase the longevity of the card. I wish you guys did a comparison video of long life cards.maybe you already did and i missed it ?
So the LARGE DIE of the GPU due to architectural choices (that need newer games to benefit) seems to be the CRUX of the problem?
A smaller, more power efficient GPU might have made more sense then they could use GDDR5/5X and even if performance dropped a bit they should be more profitable. Then maybe they could have worked on refining the FAB PROCESS to release a higher-end VEGA later.
Of course hindsight is 20/20. No doubt AMD would change some things if they could but they basically painted themselves into a corner with the HBM2 memory requirement (IMO at least).
HBM2 offers extra bandwidth and efficiency but no notable performance gains :/
I'm just hoping that RX 460, 470, and 480s start hitting the used market soon.
I just don't see how Vega 64 is any better than if they just made a bigger Polaris architecture. How about a Polaris Crossfire comparison to a single Vega 64 in terms of power, performance and "suggested" MSRP? Anyone?
They shouldn't have released vega for gaming. They should have released it only for compute market and maybe make another architecture for gaming, albeit it will be much more expensive.
Too many comments to check, but surely the gpu die costs more than the hbm+interposer, or at least as much- so 175$ mem/controller+175$ min vega die cost+15-20$pcb(with parts)+air cooler(fan 20$+vapor block heat sink~20$)= 410$ add incidentals like metal shroud and retail packaging- plus questionable although still costly marketing- is about 425-450$ cost estimate. Shipping plus retail margin is even more cost, so close to 500$ is my guess- so break even at best, some loss is more likely, maybe if hbm2 is gddr5 cost in awhile, some profit.
Such a great and informational video! Thank you so much!
Reminds me of RAMBUS, Intel troughted it with massive memory bandwidth and all the paper specs looks great but didn't perform much different than SDRAM and then DDR killed it. it's like Intel's RAMBUS fail all over again and AMD didn't learn from it because they went DDR so they should have learned.
It's surprising the number of people in the comments saying so much nonsense in a GN video, GN makes so information rich content one would assume people who watch their videos would be a lot more educated.
At 11:15 mark, then why the hell they keep running hot????
I think this is what sinks Vega. They have to price lower than Nvidia because their cards aren't up to the 1080 Ti, but then again Vega cards are more difficult and expensive to produce because of HBM2. So AMD doesn't make the money off Vega that Nvidia makes off Pascal. Not a good combination.