Actually the AMD press release has a mistake. The graphics card family is not "RX 5700-series", it's the RX 5000-series" and the RX5700 was used in the demos, as clearly stated in Lisa Su's keynote.
no they didnt. not yet. next year will be more of the same then after that we get the gcn replacement. it really doesnt matter though. people complaining about gcn dont know what they are talking about.
Feels a bit as if people complained about Zen2 being "x86, agaaaaaaain ...?" and that we've all seen with Bulldozer and Netburst how that ISA is THROUGH!
Upgrading GCN is like trying to squeeze water out of a dry rock at this point. Improving GCN performance hasn't scaled well since at least Polaris, and it seems we're stuck with it for another two years.
AMD needs a GROUND-UP overhaul of its GPU arch that can be easily scaled to higher performance and higher perf/W.
From the work in their open source driver stack, it sounds like the architectural changes relative to Vega are somewhat minimal - at least at the ISA-level. The uArch could still have some big changes, I suppose.
Sad to say it's probably a lie, though they probably think of it as "blown-up truth". It's a lie, because they knew people wanted to hear the words "new architecture" for at least two years now, and they couldn't say NO anymore. So they changed enough stuff that could plausibly make it sound like it's really a new arch - even though it totally isn't, unless you use a very twisted interpretation of what a new arch is.
Comparing CPU with GPU ISA is a bad analogy. CPUs have a front-end ISA and a back end with micro-ops, shadow registers, etc. GPUs have no such indirection. This is largely because virtually no one programs GPUs in their native assembly language, so the native ISA is much easier to change.
It's a ground up redesign because they used the same ALUs and just arranged them in a different way. It seems that they simply doubled graphics engine, for 4 to 8, that is now each engine can have up to 5 CU instead of 10, and each engine can now have more ROPs with respect to old GCN versions. If so this does not make it a completely new architecture, as it does not introduces nothing new but a "ground up redesign", that is a different way to use the old blocks with same capabilities, it just improve the number of some of those blocks. Poòaris 10 is not a new architecture with respect to Polaris 11 just because it uses more blocks, and Polaris 10 is not a new arch with respect to Ellsmere just because the number of ACES or geometry engines has increased. Here we have that to increase even more the number of uncore elements (as it seems the max number of ALUS is still 4096 as old GCN) they had to redesign the graphics engine and doubling their number.
I'm quite curious to see ho big these changes are going to impact on die size.
We have seen that nvidia had a quite die size impact on its changes at shader (with the introduction of the INT32 unit), SMX level with their doubling of their front end and with the new scheduler able to work at thread level. This on net of the other die size increase due to tensor and RT core, as we can see for TU116 and TU117. Never the less these changes brought an IPC increase which is about 30% with respect to Pascal and with a reduction is power consumption.
It's not perfect though; the 1660 loses to the 1060 6GB in The Witcher 3 and Middle Earth: Shadow of War, and on average it's more like 15% faster for about the same power consumption, certainly not the 30% increase in performance for a reduction of power consumption as you're stating (source: https://www.techspot.com/review/1813-geforce-gtx-1... ).
Arcturus is just the first chip we started working on after switching back to using engineering code names rather than names like "Navi" that were discussed publicly in roadmaps etc...
I don't think we have ever said anything to suggest that it will be post-GCN or even post-Navi.
"Thanks to the combination of AMD’s architectural improvements and TSMC’s 7nm process, AMD is promoting a 50% increase in performance per watt for Navi"
TSMC deserves no credit if those 50% are just about the architecture....
Why not? They compared the actual Vega chip with this new Navi based one in a game and they measured that the new chip is 1.5x more efficient than the old one. Comparison takes into account everything, from architectural changes to the used PP. Fact is, that 14nm vs 7nm should provide better than that only based on PP capacity. Most probably they just threw a bit of energy efficiency away to use higher clocks and make smaller dies (with less shaders and al the rest).
Of course it has. The biggest example is Intel's KBL-G. A single stack of HBM gave them pretty substantial bandwidth, enough to feed those 24 CUs and offer decent low-power performance rivalling mid-range mobile dGPUs at the time. Eventually I expect Intel will release another MCM setup like that but with their own Xe graphics.
AMD may do something similar at some point, as well. Actually even in an MXM form factor HBM could have size/power benefits for mobile dGPUs. The issue is that it's so expensive that any more than a single stack is going to be reserved for high-end hardware. Navi is going to be a mainstream chip, so GDDR is ideal for keeping price competitive without wrecking margins.
Do like Charlie, always put your business card next to it when taking photos. Not only is it "good branding technique", it's also intelligent. You can later measure things up in Photoshop or some other image editor.
1.5x efficiency is not that high, seen the upgrade from 14nm to 7nm PP. If true, however, it is better than what happened with Polaris, where they stated the same and finished with a beefly OC chip to have performances in the range of the small GP106 chip rendering those claims false.
Moreover that 1.25x more performance per clock may just come from the release of memory bottlenecks with the use of GDDR6. In fact, all test used for the mean were done in 4K resolution, putting more stress on memory bandwidth than other parts of the architecture.
BTW, I have now counted at least 4 cache hierarchy redesign in AMD architectures. It seems they are not able to use caching appropriately and keeps on experimenting on it. Hope this is the right scheme.
So for that up-to 1.5x perf/W, they gave their test as: 12. Testing done by AMD performance labs 5/23/19, using the Division 2 @ 25x14 Ultra settings. Performance may vary based on use of latest drivers. RX-325
And for the 1.25x perf/clock: 11. Testing done by AMD performance labs 5/23/19, showing a geomean of 1.25x per/clock across 30 different games @ 4K Ultra, 4xAA settings. Performance may vary based on use of latest drivers. RX-327
"1.5x efficiency is not that high, seen the upgrade from 14nm to 7nm PP."
The question is if 1.5x is the improvement of just the architecture. If so that would be really really good. If not it would be underwhelming. A new architecture and going from 12nm to 7nm should result in more than 50% better power efficiency.
Yes, it should, but with Polaris they made even bolder statement of 2.5x (combined PP + architecture) and they resulted in a similar end case with power efficiency just barely better than GCN 1.2 and still light year behind competition, despite the new PP. So 1.5x power efficiency may still be the combination of architecture + PP as from the test they did to calculate that you cannot really estrapolate only the merits of the architecture telling them apart from those of the PP.
Answered to the wrong comment below... sorry Unfortunately for AMD no GCN GPUs has ever been clocked at its sweet spot as the architecture is so computing inefficient in 3D pipeline that at the sweet spot they could just barely make competition to a two tier lower nvidia's GPUs with a not sustainable costs.
With more Hz (and consequently more W) the architecture could compete with just a tier lower nvidia GPU, so they could be sold at a loss but for much longer time.
Vega VII is the last of this series of OC GPUs with absolutely abysmal TDP with respect to their performance: Vega VII can barely beat 1080Ti consuming 100W more despite HBM and 7nm, so taking it as one of the best representative of GCN efficiency just tell how bad are the other ones.
It's funny that people try to treat Vega VII as a serious consumer card. It's no more a consumer GPU than Titan V, and yet its price/perf ratio is far better.
Perf/price ratio is only set depending on marketing choices, not technical one. In this case AMD needs to show that it has a product that can compete with a concurrent solution that is placed under the top. The fact that it has to use a GPU thought for the computing market is by itself a sign that it is doing a marketing trick. The facts are simple than the reasons that are brought out to justify them: 1. Polaris is now 2 generations behind the concurrent solutions 2. Vega isn't able to close the gap, despite the W and the latest expensive technology you throw at it, be them 7nm, HBM, tons of cache and bandwidth. 3. AMD has nothing to close the gap until it releases Navi, despite all the criticism about Turing and its prices (which are that level simply because AMD has nothing). 4. AMD eternal weapon is the same since the release of GCN: lower the price of its solution to match the lower tier one of the competition. At this moment it has to lower them 2 tiers to match Turing.
Despite Vega VII efficiency (which is really abysmal with respect to the PP used) , Polaris and Vega are not placed better with respect to the competition.
> I have now counted at least 4 cache hierarchy redesign in AMD architectures. It seems they are not able to use caching appropriately and keeps on experimenting on it. Hope this is the right scheme.
It's funny that you pose this as a matter of correctness. At least some of those changes were probably driven by priorities among their target markets and their needs. Others were probably enabled by technological changes, such as Infinity Fabric, HSA, and HBM.
These are only guesses that try to find other justifications about AMD (un)skill in creating GPU architectures. Some may be or not, whatever: what I see is that their cache schemes have varied in time on the same architecture (GCN) with no other technical changes but the fact that the previous choices were wrong. - IF is not used for the GPU, so it is useless to induce it as a possibile reason - HSA is a waporware thing that only work with iGPU as there are no example on where it has been applied with some advantage. Be it a reason for it, one may wonder why L2 cache and ROPs access changed (ROPs have nothing to do with HSA). - HBM is used on a particular designed architecture so if ever that architecture should have been used with such modifications, but we see modification done on GDDRx architectures. All I can see is AMD changing the cache hierarchy layout now and then saying that the changes is going to make improvements which however we do not see at all.
nvidia is keeping their cache hierarchy the same way it was from Kepler, so probably they created a better thought architecture to evolve since 2012. And as a result of this after 7 years we have Turing from nvidia which goes beyond the classic rasterization fixed units and a rebranded shrunk Vega without RT/tensor/advance geometry from AMD which is still thinking if it better to link ROPs directly to L2 or go directly to VRAM.
Given that the 590 uses 2x the power of the GTX1660 and is slower, I'm not sure if 1.5x efficiency will bring them up to par with NVIDIA. Was hoping for a few more details from them today.
Very bad comparison. GTX 1660 Ti is working close to the sweet spot of the architecture, while RX 590 is not even close to the sweet spot. I haven't read which GCN card was compared against Navi. Vega 56 and especially Radeon VII are way more efficient than RX 590. AMD would need ~40% more efficiency to close the gap to RX 2060/2070/2080.
Unfortunately for AMD no GCN GPUs has ever been clocked at its sweet spot as the architecture is so computing inefficient in 3D pipeline that at the sweet spot they could just barely make competition to a two tier lower nvidia's GPUs with a not sustainable costs.
With more Hz (and consequently more W) the architecture could compete with just a tier lower nvidia GPU, so they could be sold at a loss but for much longer time.
Vega VII is the last of this series of OC GPUs with absolutely abysmal TDP with respect to their performance: Vega VII can barely beat 1080Ti consuming 100W more despite HBM and 7nm, so taking it as one of the best representative of GCN efficiency just tell how bad are the other ones.
Anecdotally, the Division 2 uses no more than 160W on my 590 which is a title that it outperforms the 1660 in. -8% power limit, 1100mV max., that's it. A small undervolt can make a decent difference across the board.
I presume the 1.5x efficiency is versus Vega 14nm, and would suggest a 50% performance improvement for the same power usage, so probably about 150W for Vega 64/GTX 1080 performance as a best case.
$500 price is unlikely. Judging from the benchmark result (Vega was generaly ~10% stronger in Strange Brigade than in other games, plus this is cherrypicked)- I'd expect 5700 to be somewhere in the Vega64- 2070 range. So- pricing could be ~$380-400 for full chip card, $280-$300 for a cut down model, ~$180-200 for RX590, $130-150 for a cut down Polaris30 (or it's replacement, if that is being prepared already).
I do not understand where these rumors about cheap GPUs come. There are some conditions at work that 7nm are not cheap and nvidia has raised the minimum bar. And GTX 1080 performance are now mainstream. This means AMD new GPUs must be a bit fatter to get at nvidia's level and this means some expensive mm^2 at 7nm must be used. More over, GDDR6 is not cheap either and that is a bit of the BOM cost on 8GB boards.
Unless AMD wants to get another big hit in its quarter results it has to sell its GPU in black. At least now that it has the chance to exploit 7nm against nvidia 12nm. If it does not do this now, when nvidia will shrink Turing successor nvidia will stop selling these GPUs if they can't really provide the vantage to fill the gap not with Turing, but with Ampere (or how it will be called). And with a mere 1.25x in performance per clock (which may just the result of new memory handling) and a 1.5x in power efficiency (just enough to get on par with Pascal) I doubt these new arch can stand Turing successor.
Nvidia did claim that They do not use 7nm yet because it would make gpu too expensive! So Yes, 7nm is not cheap to And we all know that Nvidia Gpu Are cheap and according to Nvidia They would be more expensive if made at 7nm... how much 2080ti would cost at 7nm 2000$??? 500$ for 2070 speed is very real estimation and actually it is a leak from GRaphic card manufacturer, so most likely very true compared to fan made ques that we have seen.
I suspect you are misinterpretting Jen Hsun's statement that they didn't yet need to yet pay the cost to shift to 7nm, to mean that it would be "too expensive." It's not about an absolute statement of fact, that it is or isn't. It's that Nvidia decided it could preserve both high lead and high margins, while AMD didn't have that luxury, they have to pay higher costs of the 7nm process just to maintain competitveness to Nvidia SKUs still on 12nm++
Also my calculator chip is very power efficient. All chips at lower clocks get more efficient. But their performances also decrease. Efficiency is not an absolute measure, is the ratio of two. We are speaking about performance/W here. And Polaris is not efficient in any way you look at it. It becomes efficient if you want to build an even larger die at a low clock to do slow work. It would just became more expensive of what is already now in OC mode.
7nm is not so new anymore, and memory prices also continue to drop. Maybe we need an update regarding them. Anyway- if nVidia can sell a huge 445mm^2 die + GDDR6 + huge proffits, then AMD should be about as competitive at selling ~275mm^2 7nm die + GDDR6 + lower profits as well.
It all depends on how good are the yields at 7nm. The HP version of the process is new (do not get confused with the LP version of it) and because nvidia does not need to use an expensive immature PP to create something that leads the market, they simply aren't. They also stopped hurry up with the HPC version of their architecture, as Intel withdraw from the competition with their Xeon Phi solutions.
nvidia will most probably adopt a later PP (7nm or Samsung 6nm with EUV) when they'll need to make something better than Turing (not necessarily cheaper, just better in performance without needing to reach 600mm^2 of die area or 300W of power consumption). And they need something better when: 1. AMD will create a competitive architecture 2. Their Turing architecture will be old enough to need a replacement to restart the selling hiatus.
Seen AMD is a couple generation back with GCN with lower performances and no RT/AI/DL advanced geometry engine support, they are not in a hurry to give % points of gross margin just to adopt an early PP which they do not really need at this moment.
This is the same situation that was in 2008 when nvidia remained at 65nm while AMD was in a hurry for 40nm having used 55nm (and GDDR4, can someone remember that?) with not practical advantage.
In a single statement: Who leads does not need to adopt more expensive technology to obtain the same or better performance than the competition.
" and no RT/AI/DL advanced geometry engine support " no biggie.. i think AMD even said that as well.. might see it in their cards next generation... but they are looking into it now though.... personally.. with the performance hit that RT does.. id rather not have it in the cards, and have to pay for something i wont use, or will give too much of a hit to use ...
Everyone speak about "hits" in performance when activating RT forgetting that you are not forced to play at 4K@60Hz (as if it were the normality) and full HD with raytracing illuminations is way better than 4K with stretched textures. As the VRS is going to be adopted (as now nvidia and Intel support it in HW, we have to see if the Navi is too) graphics quality can be enhanced with even less performance hit. It is better to have new advanced features than claiming to be faster of the competition offering nothing of the modern technologies the market is going to use.
well..most of the tests seem to push 4k and RT, not RT and 1080p.... still.. why buy something that only a small handful of games support, when there is a card available for less, that is faster or on par? and the fact, that the 20x0 series.. is just way too expensive to begin with....
"Last but certainly not least of course is overall power efficacy." :Cringe: ! Rewrite (AFAIR, you're normally AT's most careful writer, Ryan): "Last, but not least is the overall power efficiency."
"It should be noted however that this statement didn't come with a qualifier - if it's 50% more efficient at the same clockspeeds as Vega or a given total card TDP - which can impact the meaningfulness somewhat." :Cringe:, another one: "However, this statement didn't specify if it's 50% more efficient at the same clockspeeds as Vega or a given total card TDP." # The next statement drives the point home so the last part of this sentence can be chopped off as being way too verbose.
I don't reach much WCCFTech, but I've only seen them claiming the generation after Navi is post-GCN. I read that, again, in the latest Navi rumors article they posted less than a month ago!
And it's still going to consume too much power and get very hot. Especially since it wants to catch up to Nvidia's latest offering. I think I'll sit this one out, also.
And you know this because you work in top management at AMD. Or because you are extrapolating from the past couple of years? Completely ignoring any technical and economical factors that lead to AMD's products being less efficient in gaming than NV's. And you're assuming nothing will ever change and that AMD with a history of producing decent, even excellent GPUs over the years will now just repeat Vega over and over?
1.25x IPC is interesting, if it ignores the 8 shader engines with 5CUs each that were leaked, since that theoretically doubles geometry and raster performance (to 8 triangles/clock without programmable geometry shaders) of Navi versus Vega. Though, Hawaii had trouble utilizing the extra geometry engines (via tessellation) when they last doubled shader engines from Tahiti (2 to 4). So, I wouldn't expect an outright 2x performance increase from that due to other parts of the architecture and how AMD is handling the load balancing of the shader engines now (if the leak proves true). But, if they're getting Vega56+ performance with only 2560 streaming processors and 8 shader engines, that's a significant improvement versus the 3584-4096 SPs they had to use in Vega56/64.
TU106, Navi 10's main competition, comes in at 175W TDP, so Navi will need to get near that power consumption to be perf/W competitive with Nvidia.
Same power with an advantage of an entire PP? They will be dead GPUs soon, as soon as nvidia speak about Turing successor at 7 or better PP, even without showing anything.
I believe the reason why AMD took the veil of Navi with a low/mid-range GPU (RX 5700) is that this is the architecture for the upcoming PS 5, and probably also the next Xbox. I would be amazed if Sony doesn't end up moving 5-10 million PS5 consoles in its launch year, and I doubt that AMD will even ship half that number of Navi-based dGPUs in the same time. With that being said,I wonder if Ryan and colleagues have heard any news on AMDs console business?
If you consider RXT 2080 Ti to be high-end, which it is, then RXT 2080 is mid-range and RTX 2070 would be low/mid-range because we have yet another lower number 2060 which is low-end. The rest of them GPUs under RTX 2060 are obsolete and irrelevant (I wanna buy a video card in 2019 to get 55 fps@1080p, give me a break!) If you're thinking about the cost, well you know... What,you people don't have iPhoneX, don't drive Mercedes? ;)
I'd consider the 2080 Ti to be enthusiast, 2070/2080 to be high-end, 1660 Ti/2060 to be upper mid-range, 1060/1660 to be lower mid-range, 1650 to be budget and anything below to be entry-level. I know that's rather broad but given that most people still game in 1080p and that mid-range cards generally offer that level at 60fps with high or ultra details and sell the most, that's the mainstream or mid-range, with 1440p and 4K at the same settings being the mark of high-end and enthusiast respectively. That's just my opinion, though.
Yes it is. Just a reworked GCN with a different number of graphics engine, but substantially it is the same architecture underline. A different architecture would have provided new features, like new advanced geometry engines (and not only more than the old simple ones), tensor core and possibly new cores for RT acceleration and INT execution in parallel with FP to improve IPC for real.
As it appears now it is only the same GCN shaders differently arranged in new graphics engine with a new (again) cache hierarchy.
I would add that the changes in the graphics engine are not going to help GCN work better in the HPC/DC world, but they are only new transistors aimed at making the 3d pipeline work better/faster and try to shorten the gap with respect to nvidia architecture that works like a charm with respect to the resources and power it uses to male the same work. That's why Navi is not going to be brought to HPC/DC market. It will just be bigger with no practical advantage for that kind of works.
The RX5700 series are mid-range chips, intended as competition to RTX2070/2060. A better question to ask would be: why is the RTX2070 only roughly as fast as last generation's high-end product?
All you are so damn concerned about percentages and numbers posted that most of you fail to accept the fact, that the only thing that really matters is the actual performance/price.
If Nvidia has anything at that price range and performance range, you should only be concerned about comparing it with that.
It really, really does not matter if Nvidia pulls a rabbit out of their ass, if it can't compete for same performance at same or lower price.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
99 Comments
Back to Article
KOneJ - Sunday, May 26, 2019 - link
https://www.globenewswire.com/news-release/2019/05...LemmingOverlord - Tuesday, May 28, 2019 - link
Actually the AMD press release has a mistake. The graphics card family is not "RX 5700-series", it's the RX 5000-series" and the RX5700 was used in the demos, as clearly stated in Lisa Su's keynote.Chaitanya - Monday, May 27, 2019 - link
Good to see AMD finally moving away from GCN.del42sa - Monday, May 27, 2019 - link
really ? do they ? With NAVI ?Opencg - Monday, May 27, 2019 - link
no they didnt. not yet. next year will be more of the same then after that we get the gcn replacement. it really doesnt matter though. people complaining about gcn dont know what they are talking about.dr.denton - Monday, May 27, 2019 - link
Feels a bit as if people complained about Zen2 being "x86, agaaaaaaain ...?" and that we've all seen with Bulldozer and Netburst how that ISA is THROUGH!Krysto - Tuesday, June 11, 2019 - link
It's not the same at all.Upgrading GCN is like trying to squeeze water out of a dry rock at this point. Improving GCN performance hasn't scaled well since at least Polaris, and it seems we're stuck with it for another two years.
AMD needs a GROUND-UP overhaul of its GPU arch that can be easily scaled to higher performance and higher perf/W.
Jinkguns - Monday, May 27, 2019 - link
I'm very confused. They literally say RDNA is a new clean sheet architecture. Lisa Su said it is not GCN.mode_13h - Tuesday, May 28, 2019 - link
From the work in their open source driver stack, it sounds like the architectural changes relative to Vega are somewhat minimal - at least at the ISA-level. The uArch could still have some big changes, I suppose.Krysto - Tuesday, June 11, 2019 - link
Sad to say it's probably a lie, though they probably think of it as "blown-up truth". It's a lie, because they knew people wanted to hear the words "new architecture" for at least two years now, and they couldn't say NO anymore. So they changed enough stuff that could plausibly make it sound like it's really a new arch - even though it totally isn't, unless you use a very twisted interpretation of what a new arch is.TEAMSWITCHER - Monday, May 27, 2019 - link
In name only.tipoo - Monday, May 27, 2019 - link
The SIMD setup change is a pretty fundamental redesign. How much can change while still calling it GCN?WinterCharm - Tuesday, May 28, 2019 - link
Lisa Su, CEO of AMD -- "Its a ground up redesign"Armchair internet commenter -- "It uses the same ISA so it's not!"
Guess what? The i9 9900k, AMD Bulldozer, and AMD Zen all use x86/64. they are not the same design.
mode_13h - Tuesday, May 28, 2019 - link
Comparing CPU with GPU ISA is a bad analogy. CPUs have a front-end ISA and a back end with micro-ops, shadow registers, etc. GPUs have no such indirection. This is largely because virtually no one programs GPUs in their native assembly language, so the native ISA is much easier to change.CiccioB - Wednesday, May 29, 2019 - link
It's a ground up redesign because they used the same ALUs and just arranged them in a different way.It seems that they simply doubled graphics engine, for 4 to 8, that is now each engine can have up to 5 CU instead of 10, and each engine can now have more ROPs with respect to old GCN versions.
If so this does not make it a completely new architecture, as it does not introduces nothing new but a "ground up redesign", that is a different way to use the old blocks with same capabilities, it just improve the number of some of those blocks.
Poòaris 10 is not a new architecture with respect to Polaris 11 just because it uses more blocks, and Polaris 10 is not a new arch with respect to Ellsmere just because the number of ACES or geometry engines has increased.
Here we have that to increase even more the number of uncore elements (as it seems the max number of ALUS is still 4096 as old GCN) they had to redesign the graphics engine and doubling their number.
I'm quite curious to see ho big these changes are going to impact on die size.
We have seen that nvidia had a quite die size impact on its changes at shader (with the introduction of the INT32 unit), SMX level with their doubling of their front end and with the new scheduler able to work at thread level. This on net of the other die size increase due to tensor and RT core, as we can see for TU116 and TU117.
Never the less these changes brought an IPC increase which is about 30% with respect to Pascal and with a reduction is power consumption.
silverblue - Wednesday, May 29, 2019 - link
It's not perfect though; the 1660 loses to the 1060 6GB in The Witcher 3 and Middle Earth: Shadow of War, and on average it's more like 15% faster for about the same power consumption, certainly not the 30% increase in performance for a reduction of power consumption as you're stating (source: https://www.techspot.com/review/1813-geforce-gtx-1... ).Santoval - Monday, May 27, 2019 - link
No they didn't. Their first post-GCN card is going to be Arcturus.bridgmanAMD - Wednesday, May 29, 2019 - link
Arcturus is just the first chip we started working on after switching back to using engineering code names rather than names like "Navi" that were discussed publicly in roadmaps etc...I don't think we have ever said anything to suggest that it will be post-GCN or even post-Navi.
mode_13h - Wednesday, May 29, 2019 - link
Thanks for that.Now, I'm unclear on the distinction between the two classes of names - is it just that Arcturus hasn't been publicly announced?
brakdoo - Monday, May 27, 2019 - link
"Thanks to the combination of AMD’s architectural improvements and TSMC’s 7nm process, AMD is promoting a 50% increase in performance per watt for Navi"TSMC deserves no credit if those 50% are just about the architecture....
nevcairiel - Monday, May 27, 2019 - link
It literally says a combination of architecture and process.CiccioB - Monday, May 27, 2019 - link
Why not?They compared the actual Vega chip with this new Navi based one in a game and they measured that the new chip is 1.5x more efficient than the old one. Comparison takes into account everything, from architectural changes to the used PP.
Fact is, that 14nm vs 7nm should provide better than that only based on PP capacity.
Most probably they just threw a bit of energy efficiency away to use higher clocks and make smaller dies (with less shaders and al the rest).
LemmingOverlord - Tuesday, May 28, 2019 - link
It *LITERALLY* says it's a combination of process & architecture. Are you blind?Gachigasm - Monday, May 27, 2019 - link
someone said 128 sp per cu, so I bet on 64?also estimated die size?
Ryan Smith - Monday, May 27, 2019 - link
It's 64 SPs per CU. AMD's press release confirmed it.Threska - Monday, May 27, 2019 - link
"The products will also be AMD's first video cards using faster GDDR6 memory. "So does this mean they're abandoning HBM?
Ryan Smith - Monday, May 27, 2019 - link
No. Future HBM products have not yet been announced, but HBM is clearly not going anywhere for HPC cards.Xajel - Monday, May 27, 2019 - link
Nope, AFAWK; Navi will start with mainstream to low-end markets. No high-end for now.That's why it uses GDDR6 as HBMx is still considered as high-end duo to it's current implementation cost compared to GDDRx
CiccioB - Monday, May 27, 2019 - link
Have HBM had any meaning outside HPC market?Alexvrb - Monday, May 27, 2019 - link
Of course it has. The biggest example is Intel's KBL-G. A single stack of HBM gave them pretty substantial bandwidth, enough to feed those 24 CUs and offer decent low-power performance rivalling mid-range mobile dGPUs at the time. Eventually I expect Intel will release another MCM setup like that but with their own Xe graphics.AMD may do something similar at some point, as well. Actually even in an MXM form factor HBM could have size/power benefits for mobile dGPUs. The issue is that it's so expensive that any more than a single stack is going to be reserved for high-end hardware. Navi is going to be a mainstream chip, so GDDR is ideal for keeping price competitive without wrecking margins.
brakdoo - Monday, May 27, 2019 - link
How wide is that thumb? Die area?Ryan Smith - Monday, May 27, 2019 - link
We didn't have a ruler with us for the GPU shot, unfortunately.As for the thumb, all I can tell you is that it's 1 thumb in width, plus or minus a few layers of epidermis.
ats - Monday, May 27, 2019 - link
Also, hope you don't use your thumbprint for security...LemmingOverlord - Tuesday, May 28, 2019 - link
Do like Charlie, always put your business card next to it when taking photos. Not only is it "good branding technique", it's also intelligent. You can later measure things up in Photoshop or some other image editor.ET - Monday, May 27, 2019 - link
Precisely my thought. To see this picture without a die size estimate was disappointing.Ryan Smith - Monday, May 27, 2019 - link
We now have an estimate from some additional photos: 275mm2.=)zepi - Monday, May 27, 2019 - link
Modern phones should be able to run all kinds of AR-measurement apps, you should have tried some of those.mode_13h - Wednesday, May 29, 2019 - link
They're not terribly accurate, though.CiccioB - Monday, May 27, 2019 - link
1.5x efficiency is not that high, seen the upgrade from 14nm to 7nm PP.If true, however, it is better than what happened with Polaris, where they stated the same and finished with a beefly OC chip to have performances in the range of the small GP106 chip rendering those claims false.
Moreover that 1.25x more performance per clock may just come from the release of memory bottlenecks with the use of GDDR6. In fact, all test used for the mean were done in 4K resolution, putting more stress on memory bandwidth than other parts of the architecture.
BTW, I have now counted at least 4 cache hierarchy redesign in AMD architectures. It seems they are not able to use caching appropriately and keeps on experimenting on it. Hope this is the right scheme.
jabbadap - Monday, May 27, 2019 - link
Well yeah you have obviously read the press release footnotes, so let's give the link to it for all of us:https://www.amd.com/en/press-releases/2019-05-26-a...
So for that up-to 1.5x perf/W, they gave their test as:
12. Testing done by AMD performance labs 5/23/19, using the Division 2 @ 25x14 Ultra settings. Performance may vary based on use of latest drivers. RX-325
And for the 1.25x perf/clock:
11. Testing done by AMD performance labs 5/23/19, showing a geomean of 1.25x per/clock across 30 different games @ 4K Ultra, 4xAA settings. Performance may vary based on use of latest drivers. RX-327
gruffi - Monday, May 27, 2019 - link
"1.5x efficiency is not that high, seen the upgrade from 14nm to 7nm PP."The question is if 1.5x is the improvement of just the architecture. If so that would be really really good. If not it would be underwhelming. A new architecture and going from 12nm to 7nm should result in more than 50% better power efficiency.
CiccioB - Monday, May 27, 2019 - link
Yes, it should, but with Polaris they made even bolder statement of 2.5x (combined PP + architecture) and they resulted in a similar end case with power efficiency just barely better than GCN 1.2 and still light year behind competition, despite the new PP.So 1.5x power efficiency may still be the combination of architecture + PP as from the test they did to calculate that you cannot really estrapolate only the merits of the architecture telling them apart from those of the PP.
dr.denton - Tuesday, May 28, 2019 - link
Depends on whether you compare architectures or actual production GPUs. Polaris is very power efficient - just not at clock speeds well above 1Ghz.CiccioB - Tuesday, May 28, 2019 - link
Answered to the wrong comment below... sorryUnfortunately for AMD no GCN GPUs has ever been clocked at its sweet spot as the architecture is so computing inefficient in 3D pipeline that at the sweet spot they could just barely make competition to a two tier lower nvidia's GPUs with a not sustainable costs.
With more Hz (and consequently more W) the architecture could compete with just a tier lower nvidia GPU, so they could be sold at a loss but for much longer time.
Vega VII is the last of this series of OC GPUs with absolutely abysmal TDP with respect to their performance: Vega VII can barely beat 1080Ti consuming 100W more despite HBM and 7nm, so taking it as one of the best representative of GCN efficiency just tell how bad are the other ones.
mode_13h - Wednesday, May 29, 2019 - link
It's funny that people try to treat Vega VII as a serious consumer card. It's no more a consumer GPU than Titan V, and yet its price/perf ratio is far better.CiccioB - Wednesday, May 29, 2019 - link
Perf/price ratio is only set depending on marketing choices, not technical one.In this case AMD needs to show that it has a product that can compete with a concurrent solution that is placed under the top. The fact that it has to use a GPU thought for the computing market is by itself a sign that it is doing a marketing trick.
The facts are simple than the reasons that are brought out to justify them:
1. Polaris is now 2 generations behind the concurrent solutions
2. Vega isn't able to close the gap, despite the W and the latest expensive technology you throw at it, be them 7nm, HBM, tons of cache and bandwidth.
3. AMD has nothing to close the gap until it releases Navi, despite all the criticism about Turing and its prices (which are that level simply because AMD has nothing).
4. AMD eternal weapon is the same since the release of GCN: lower the price of its solution to match the lower tier one of the competition. At this moment it has to lower them 2 tiers to match Turing.
Despite Vega VII efficiency (which is really abysmal with respect to the PP used) , Polaris and Vega are not placed better with respect to the competition.
Krysto - Tuesday, June 11, 2019 - link
Considering it barely beats Nvidia's competing products on 14nm, I'm going to assume the 1.5x efficiency is from the 7nm node alone.mode_13h - Wednesday, May 29, 2019 - link
> I have now counted at least 4 cache hierarchy redesign in AMD architectures. It seems they are not able to use caching appropriately and keeps on experimenting on it. Hope this is the right scheme.It's funny that you pose this as a matter of correctness. At least some of those changes were probably driven by priorities among their target markets and their needs. Others were probably enabled by technological changes, such as Infinity Fabric, HSA, and HBM.
CiccioB - Wednesday, May 29, 2019 - link
These are only guesses that try to find other justifications about AMD (un)skill in creating GPU architectures.Some may be or not, whatever: what I see is that their cache schemes have varied in time on the same architecture (GCN) with no other technical changes but the fact that the previous choices were wrong.
- IF is not used for the GPU, so it is useless to induce it as a possibile reason
- HSA is a waporware thing that only work with iGPU as there are no example on where it has been applied with some advantage. Be it a reason for it, one may wonder why L2 cache and ROPs access changed (ROPs have nothing to do with HSA).
- HBM is used on a particular designed architecture so if ever that architecture should have been used with such modifications, but we see modification done on GDDRx architectures.
All I can see is AMD changing the cache hierarchy layout now and then saying that the changes is going to make improvements which however we do not see at all.
nvidia is keeping their cache hierarchy the same way it was from Kepler, so probably they created a better thought architecture to evolve since 2012. And as a result of this after 7 years we have Turing from nvidia which goes beyond the classic rasterization fixed units and a rebranded shrunk Vega without RT/tensor/advance geometry from AMD which is still thinking if it better to link ROPs directly to L2 or go directly to VRAM.
AndrewJacksonZA - Monday, May 27, 2019 - link
"(ed: you still owe me a steak)"*chuckle * I really enjoy the personality that shows through in this site. I like the culture here. :-)
sorten - Monday, May 27, 2019 - link
Given that the 590 uses 2x the power of the GTX1660 and is slower, I'm not sure if 1.5x efficiency will bring them up to par with NVIDIA. Was hoping for a few more details from them today.sorten - Monday, May 27, 2019 - link
edit: GTX 1660Tigruffi - Monday, May 27, 2019 - link
Very bad comparison. GTX 1660 Ti is working close to the sweet spot of the architecture, while RX 590 is not even close to the sweet spot. I haven't read which GCN card was compared against Navi. Vega 56 and especially Radeon VII are way more efficient than RX 590. AMD would need ~40% more efficiency to close the gap to RX 2060/2070/2080.CiccioB - Monday, May 27, 2019 - link
Unfortunately for AMD no GCN GPUs has ever been clocked at its sweet spot as the architecture is so computing inefficient in 3D pipeline that at the sweet spot they could just barely make competition to a two tier lower nvidia's GPUs with a not sustainable costs.With more Hz (and consequently more W) the architecture could compete with just a tier lower nvidia GPU, so they could be sold at a loss but for much longer time.
Vega VII is the last of this series of OC GPUs with absolutely abysmal TDP with respect to their performance: Vega VII can barely beat 1080Ti consuming 100W more despite HBM and 7nm, so taking it as one of the best representative of GCN efficiency just tell how bad are the other ones.
bridgmanAMD - Wednesday, May 29, 2019 - link
How about the Fury Nano ?mode_13h - Wednesday, May 29, 2019 - link
See my above comment about Vega VII. You're simply off-base, trying to use it as some kind of trend.silverblue - Wednesday, May 29, 2019 - link
Anecdotally, the Division 2 uses no more than 160W on my 590 which is a title that it outperforms the 1660 in. -8% power limit, 1100mV max., that's it. A small undervolt can make a decent difference across the board.I presume the 1.5x efficiency is versus Vega 14nm, and would suggest a 50% performance improvement for the same power usage, so probably about 150W for Vega 64/GTX 1080 performance as a best case.
zodiacfml - Monday, May 27, 2019 - link
$500? I guess, we won't be seeing $200 or less 7nm cards this yearneblogai - Monday, May 27, 2019 - link
$500 price is unlikely. Judging from the benchmark result (Vega was generaly ~10% stronger in Strange Brigade than in other games, plus this is cherrypicked)- I'd expect 5700 to be somewhere in the Vega64- 2070 range. So- pricing could be ~$380-400 for full chip card, $280-$300 for a cut down model, ~$180-200 for RX590, $130-150 for a cut down Polaris30 (or it's replacement, if that is being prepared already).CiccioB - Monday, May 27, 2019 - link
I do not understand where these rumors about cheap GPUs come.There are some conditions at work that
7nm are not cheap and nvidia has raised the minimum bar. And GTX 1080 performance are now mainstream.
This means AMD new GPUs must be a bit fatter to get at nvidia's level and this means some expensive mm^2 at 7nm must be used.
More over, GDDR6 is not cheap either and that is a bit of the BOM cost on 8GB boards.
Unless AMD wants to get another big hit in its quarter results it has to sell its GPU in black. At least now that it has the chance to exploit 7nm against nvidia 12nm. If it does not do this now, when nvidia will shrink Turing successor nvidia will stop selling these GPUs if they can't really provide the vantage to fill the gap not with Turing, but with Ampere (or how it will be called).
And with a mere 1.25x in performance per clock (which may just the result of new memory handling) and a 1.5x in power efficiency (just enough to get on par with Pascal) I doubt these new arch can stand Turing successor.
Opencg - Monday, May 27, 2019 - link
rumors are these will all be significantly cheaper than their nvidia counterparts. this is from leakshaukionkannel - Monday, May 27, 2019 - link
Nvidia did claim that They do not use 7nm yet because it would make gpu too expensive!So Yes, 7nm is not cheap to And we all know that Nvidia Gpu Are cheap and according to Nvidia They would be more expensive if made at 7nm... how much 2080ti would cost at 7nm 2000$???
500$ for 2070 speed is very real estimation and actually it is a leak from GRaphic card manufacturer, so most likely very true compared to fan made ques that we have seen.
Qasar - Monday, May 27, 2019 - link
they did ?? where did nvidia say that ?haukionkannel - Tuesday, May 28, 2019 - link
When Radeon 7 was released...Qasar - Tuesday, May 28, 2019 - link
hmmm dont remember reading that....MadManMark - Tuesday, May 28, 2019 - link
I suspect you are misinterpretting Jen Hsun's statement that they didn't yet need to yet pay the cost to shift to 7nm, to mean that it would be "too expensive." It's not about an absolute statement of fact, that it is or isn't. It's that Nvidia decided it could preserve both high lead and high margins, while AMD didn't have that luxury, they have to pay higher costs of the 7nm process just to maintain competitveness to Nvidia SKUs still on 12nm++CiccioB - Tuesday, May 28, 2019 - link
Also my calculator chip is very power efficient.All chips at lower clocks get more efficient. But their performances also decrease.
Efficiency is not an absolute measure, is the ratio of two.
We are speaking about performance/W here.
And Polaris is not efficient in any way you look at it.
It becomes efficient if you want to build an even larger die at a low clock to do slow work. It would just became more expensive of what is already now in OC mode.
neblogai - Monday, May 27, 2019 - link
7nm is not so new anymore, and memory prices also continue to drop. Maybe we need an update regarding them. Anyway- if nVidia can sell a huge 445mm^2 die + GDDR6 + huge proffits, then AMD should be about as competitive at selling ~275mm^2 7nm die + GDDR6 + lower profits as well.CiccioB - Monday, May 27, 2019 - link
It all depends on how good are the yields at 7nm.The HP version of the process is new (do not get confused with the LP version of it) and because nvidia does not need to use an expensive immature PP to create something that leads the market, they simply aren't.
They also stopped hurry up with the HPC version of their architecture, as Intel withdraw from the competition with their Xeon Phi solutions.
nvidia will most probably adopt a later PP (7nm or Samsung 6nm with EUV) when they'll need to make something better than Turing (not necessarily cheaper, just better in performance without needing to reach 600mm^2 of die area or 300W of power consumption).
And they need something better when:
1. AMD will create a competitive architecture
2. Their Turing architecture will be old enough to need a replacement to restart the selling hiatus.
Seen AMD is a couple generation back with GCN with lower performances and no RT/AI/DL advanced geometry engine support, they are not in a hurry to give % points of gross margin just to adopt an early PP which they do not really need at this moment.
This is the same situation that was in 2008 when nvidia remained at 65nm while AMD was in a hurry for 40nm having used 55nm (and GDDR4, can someone remember that?) with not practical advantage.
In a single statement:
Who leads does not need to adopt more expensive technology to obtain the same or better performance than the competition.
Korguz - Monday, May 27, 2019 - link
" and no RT/AI/DL advanced geometry engine support " no biggie.. i think AMD even said that as well.. might see it in their cards next generation... but they are looking into it now though.... personally.. with the performance hit that RT does.. id rather not have it in the cards, and have to pay for something i wont use, or will give too much of a hit to use ...CiccioB - Tuesday, May 28, 2019 - link
Everyone speak about "hits" in performance when activating RT forgetting that you are not forced to play at 4K@60Hz (as if it were the normality) and full HD with raytracing illuminations is way better than 4K with stretched textures.As the VRS is going to be adopted (as now nvidia and Intel support it in HW, we have to see if the Navi is too) graphics quality can be enhanced with even less performance hit.
It is better to have new advanced features than claiming to be faster of the competition offering nothing of the modern technologies the market is going to use.
Qasar - Tuesday, May 28, 2019 - link
well..most of the tests seem to push 4k and RT, not RT and 1080p.... still.. why buy something that only a small handful of games support, when there is a card available for less, that is faster or on par? and the fact, that the 20x0 series.. is just way too expensive to begin with....mr_tawan - Monday, May 27, 2019 - link
Looking forward for in-depth technical analysis from you guys :).mode_13h - Wednesday, May 29, 2019 - link
Who? The site's authors or the commentariat peanut gallery?mikato - Thursday, May 30, 2019 - link
You're selling yourself short mode_13hboredsysadmin - Monday, May 27, 2019 - link
Small Typo in article : not RNDA, but RDNAballsystemlord - Monday, May 27, 2019 - link
Spelling and grammar corrections:"Last but certainly not least of course is overall power efficacy."
:Cringe: ! Rewrite (AFAIR, you're normally AT's most careful writer, Ryan):
"Last, but not least is the overall power efficiency."
"It should be noted however that this statement didn't come with a qualifier - if it's 50% more efficient at the same clockspeeds as Vega or a given total card TDP - which can impact the meaningfulness somewhat."
:Cringe:, another one:
"However, this statement didn't specify if it's 50% more efficient at the same clockspeeds as Vega or a given total card TDP."
# The next statement drives the point home so the last part of this sentence can be chopped off as being way too verbose.
ballsystemlord - Monday, May 27, 2019 - link
EDIT: Telepathically added a comma, need to actually type it next time:"Last, but not least, is the overall power efficiency."
ballsystemlord - Monday, May 27, 2019 - link
@Ryan, when you guys bench these please run them through the full battery of compute tests instead of your recent lighter version.Thanks!
R3MF - Monday, May 27, 2019 - link
Laughed my boobies off reading WCCFtech waxing nostradamus saying Navi was a new uarch, and he called it all along.mode_13h - Wednesday, May 29, 2019 - link
I don't reach much WCCFTech, but I've only seen them claiming the generation after Navi is post-GCN. I read that, again, in the latest Navi rumors article they posted less than a month ago!Valis - Monday, May 27, 2019 - link
And it's still going to consume too much power and get very hot. Especially since it wants to catch up to Nvidia's latest offering. I think I'll sit this one out, also.dr.denton - Tuesday, May 28, 2019 - link
And you know this because you work in top management at AMD. Or because you are extrapolating from the past couple of years? Completely ignoring any technical and economical factors that lead to AMD's products being less efficient in gaming than NV's. And you're assuming nothing will ever change and that AMD with a history of producing decent, even excellent GPUs over the years will now just repeat Vega over and over?Krysto - Tuesday, June 11, 2019 - link
Next year seems to be Navi+, too, so we'll have to sit that one out, too. Gosh darn it, AMD!a_throwaway? - Tuesday, May 28, 2019 - link
Really? No one is gonna comment on the solder bridge in the press pic of the underside of the chip? :)JasonMZW20 - Tuesday, May 28, 2019 - link
1.25x IPC is interesting, if it ignores the 8 shader engines with 5CUs each that were leaked, since that theoretically doubles geometry and raster performance (to 8 triangles/clock without programmable geometry shaders) of Navi versus Vega. Though, Hawaii had trouble utilizing the extra geometry engines (via tessellation) when they last doubled shader engines from Tahiti (2 to 4). So, I wouldn't expect an outright 2x performance increase from that due to other parts of the architecture and how AMD is handling the load balancing of the shader engines now (if the leak proves true). But, if they're getting Vega56+ performance with only 2560 streaming processors and 8 shader engines, that's a significant improvement versus the 3584-4096 SPs they had to use in Vega56/64.TU106, Navi 10's main competition, comes in at 175W TDP, so Navi will need to get near that power consumption to be perf/W competitive with Nvidia.
CiccioB - Tuesday, May 28, 2019 - link
Same power with an advantage of an entire PP?They will be dead GPUs soon, as soon as nvidia speak about Turing successor at 7 or better PP, even without showing anything.
eastcoast_pete - Tuesday, May 28, 2019 - link
I believe the reason why AMD took the veil of Navi with a low/mid-range GPU (RX 5700) is that this is the architecture for the upcoming PS 5, and probably also the next Xbox. I would be amazed if Sony doesn't end up moving 5-10 million PS5 consoles in its launch year, and I doubt that AMD will even ship half that number of Navi-based dGPUs in the same time. With that being said,I wonder if Ryan and colleagues have heard any news on AMDs console business?silverblue - Wednesday, May 29, 2019 - link
The RTX 2070 is low/mid-range? Damn it, have I been asleep for the last few months?Gastec - Thursday, May 30, 2019 - link
If you consider RXT 2080 Ti to be high-end, which it is, then RXT 2080 is mid-range and RTX 2070 would be low/mid-range because we have yet another lower number 2060 which is low-end. The rest of them GPUs under RTX 2060 are obsolete and irrelevant (I wanna buy a video card in 2019 to get 55 fps@1080p, give me a break!)If you're thinking about the cost, well you know... What,you people don't have iPhoneX, don't drive Mercedes? ;)
mikato - Thursday, May 30, 2019 - link
No no no :) The RX 580 is still the mid-range card to get.silverblue - Friday, May 31, 2019 - link
I'd consider the 2080 Ti to be enthusiast, 2070/2080 to be high-end, 1660 Ti/2060 to be upper mid-range, 1060/1660 to be lower mid-range, 1650 to be budget and anything below to be entry-level. I know that's rather broad but given that most people still game in 1080p and that mid-range cards generally offer that level at 60fps with high or ultra details and sell the most, that's the mainstream or mid-range, with 1440p and 4K at the same settings being the mark of high-end and enthusiast respectively. That's just my opinion, though.Moody66 - Wednesday, May 29, 2019 - link
She said they would continue gcn with the vega line for professionals. NAVI is not GCN.CiccioB - Wednesday, May 29, 2019 - link
Yes it is.Just a reworked GCN with a different number of graphics engine, but substantially it is the same architecture underline.
A different architecture would have provided new features, like new advanced geometry engines (and not only more than the old simple ones), tensor core and possibly new cores for RT acceleration and INT execution in parallel with FP to improve IPC for real.
As it appears now it is only the same GCN shaders differently arranged in new graphics engine with a new (again) cache hierarchy.
CiccioB - Wednesday, May 29, 2019 - link
I would add that the changes in the graphics engine are not going to help GCN work better in the HPC/DC world, but they are only new transistors aimed at making the 3d pipeline work better/faster and try to shorten the gap with respect to nvidia architecture that works like a charm with respect to the resources and power it uses to male the same work.That's why Navi is not going to be brought to HPC/DC market. It will just be bigger with no practical advantage for that kind of works.
imaskar - Sunday, June 2, 2019 - link
Can someone explain, why doing 10% better, than competitor's pre-flagship is good? It means it is actually looses to RTX2080, right?dr.denton - Monday, June 3, 2019 - link
The RX5700 series are mid-range chips, intended as competition to RTX2070/2060. A better question to ask would be: why is the RTX2070 only roughly as fast as last generation's high-end product?marksilvester07 - Monday, June 3, 2019 - link
All the contents are mentioned in this article. I will keep it in mind, thanks for sharing the information keep updating, looking forward to more posts. https://www.quickenassist.com/blog/how-to-recover-...RobJoy - Monday, June 24, 2019 - link
All you are so damn concerned about percentages and numbers posted that most of you fail to accept the fact, that the only thing that really matters is the actual performance/price.If Nvidia has anything at that price range and performance range, you should only be concerned about comparing it with that.
It really, really does not matter if Nvidia pulls a rabbit out of their ass, if it can't compete for same performance at same or lower price.