AmazingAveryAge of Conan AdvocateMemberUncommonPosts: 7,188
edited August 2017
I think there are some major issues with Vega 10.
Vega is supposed to bring much better resource utilization and perf/sp than Fiji and definitely higher than Polaris.
There are obvious architectural bottlenecks - perf/clock has gone up a meager 6% over Fiji. Thats even lower than Polaris which gained 7% over Tonga.
Massive regression in perf/watt and perf/sq mm over even the lackluster Fiji. For a 486 sq mm Vega 10 with HBM2 to struggle to beat a Fury X by avg 35% while drawing more than 300w compared to Fury X at 275w was truly embarassing. The 14nm node brings roughly 2x transistor density and 60% lower power at same perf or 40% higher perf at same power.
70-80% perf gain is what Nvidia got from GM102 Maxwell to Pascal GP102. AMD have got half of that. Thats truly disastrous. The GF excuse just doesn't wash, GCN flopped on 28nm and Nvidia got similar result on Samsung's 14LPP with GP107. i guess that thecommon denominator on failure is AMD/GCN, not GF really.
Process limitations and missing target clocks - AMD Polaris and Vega both represent clear examples of AMD missing their frequency and power targets. GF has failed badly with 14LPP implementation and AMD suffered the consequences when they tried to ramp clocks beyond optimal ranges. On paper, with the feature set Vega has, it should have been much better than what we are seeing (and heavily lead to believe), and with over 18 months in development, it is looking more and more like that there are hardware issues that must be bypassed by software (drivers) to fix.
Launch drivers not supporting some of the main hardware features like primitive shaders and HBCC. Features like Tile based rasterization were rushed as press received launch review drivers only 2 days before launch. The software aspect of Vega launch is very disturbing as AMD seemed to have made strides with drivers after RTG was created. But we again see AMD fail hard with drivers on such a major GPU launch.
It is the exact same thing people were saying after bulldozer launched, AMD is dead, they can't compete, they are going bankrupt, and so on.
Reality showed that AMD can actually compete, once they have good leadership & direction, so, for all we know, Navi could be the comeback kid, like Ryzen was.
(And NOBODY will believe anything from AMD's crappy PR machine about Navi, until benches are done.)
Vega is more than 50% larger in die size, has 60% more transistors, and consumes 120+ watts more when going toe-to-toe with the GTX 1080. The 390x was much closer in die size and transistors to it's immediate competitor (the GTX 980) while consuming similar or less power than RX Vega 64.
Vega is not a step in the right direction. Fiji was a huge step in the right direction. Polaris was a small stumble, Vega is lame duck. AMD can't keep throwing transistors and TDP at the problem; they're at the wall. If they don't heavily innovate they will be even worse off next round because I guarantee you Volta will not be some 5-10% update like Intel's CPU's were for years which allowed AMD to catch Intel in the CPU space.
Vega is more than 50% larger in die size, has 60% more transistors, and consumes 120+ watts more when going toe-to-toe with the GTX 1080. The 390x was much closer in die size and transistors to it's immediate competitor (the GTX 980) while consuming similar or less power than RX Vega 64.
Vega is not a step in the right direction. Fiji was a huge step in the right direction. Polaris was a small stumble, Vega is lame duck. AMD can't keep throwing transistors and TDP at the problem; they're at the wall. If they don't heavily innovate they will be even worse off next round because I guarantee you Volta will not be some 5-10% update like Intel's CPU's were for years which allowed AMD to catch Intel in the CPU space.
When you put it this way:
They crammed 60% transistors into a 50% space, which is impressive in and of itself (6 lbs of crap in a 5lb sock?). And given that 390X and Fury X's TDP were 275W, to only come in at 295W means the power draw only went up by about 10%. But we got about 50% more computing/graphics power out of the generation, and nearly double the memory bandwidth & texture fill rate.
It's better than Fury's launch sitatution, which apart from HBM1 was a pretty lackluster improvement over the Hawaii. So much so that AMD not only kept Hawaii around, but refreshed it just one month prior to Fiji's release as the 390X (with only very minor clock bumps and a VRAM upgrade), and included it in their marketing for people who where so worried about the 4GB VRAM limitation on Fury.
So taken in light of the competition, yeah, doesn't look so good. But here your making sound like AMD hasn't improved - they have, and Vega represents a lot more improvement than Fiji did. And honestly, in today's chips, TDP is whatever you want it to be with PowerTune/Turbo settings within the operating constraints of the silicon - it's just a matter of how high your willing to go: nVidia didn't want to fool with water cooling, because they didn't have to in order to claim a performance crown, but I would be willing to bet dollars to donuts that if the field were reversed (like, oh say, Fermi), they would have no problem cranking up the clocks and TDP be damned in order to be competitive.
Not to take away from the fact that nVidia is doing a better job, but when you do stop to think about it, what AMD is impressive in it's own right. It's just too little too late. Like coming out and scoring in an NBA game, but only after your opponent has run up 37 points, or hitting that buzzer beater, even though you lost the game.
I do think your right about Vega/Navi vs Volta, that even estimating aggressively, AMD is several years away from retaking any performance/efficiency crowns away from nVidia. But that doesn't mean it's impossible, the tables have turned a number of times over the generation, and all it takes is one good architecture and a lucky break. Every manufacturing process is a bit different, and one particular design may work great on it, then utterly suck on the next one around.
Besides, we don't ~need~ AMD to beat the 1080Ti, as much as I would have liked to see computing power overall advance and push the envelope a bit more. We just need AMD to remain solvent and competitive. AMD is still very much competitive with Polaris in the low-mid tiers, and in a lot of ways superior to nVidia, and that's a much higher volume market.
Vega will probably sell about like Fiji did, as there are just too many parallels here: tepid but not insignificant initially, and one good price cut and it all of a sudden becomes a very good looking buy for consumers (albeit with, probably, nearly no margin for AMD). Vega may even do a bit better than Fiji without the price cut now that FreeSync is out in the wild and there's a bit of installed monitor base out there. And there's always the volatility of the mining segment, and we've all see what impact it can have on things. Right now Vega doesn't look that attractive based on perf/watt for miners, but that could easily change.
Imagine a world where nVidia's only graphics competition is Intel.
AmazingAveryAge of Conan AdvocateMemberUncommonPosts: 7,188
It's not double the performance of Polaris, but it is double the efficiency.
Granted, MSRP for Polaris (580) is $220US, and for Vega64 is $500, so it's more than double the up front price, and availability and price markup is still an issue for all of them. I'm not sure how the increase in efficiency will play out with ROI for miners, but it's probably safe to assume that it has gotten their attention.
Outside of combo deals, RX580's are going for $725US right now on Newegg, whereas reference Vega64's are going for $699.
It's not double the performance of Polaris, but it is double the efficiency.
Granted, MSRP for Polaris (580) is $220US, and for Vega64 is $500, so it's more than double the up front price, and availability and price markup is still an issue for all of them. I'm not sure how the increase in efficiency will play out with ROI for miners, but it's probably safe to assume that it has gotten their attention.
Outside of combo deals, RX580's are going for $725US right now on Newegg, whereas reference Vega64's are going for $699.
The price on the RX580 seems odd, they had a bunch at Microcenter the other day for $220, there only stipulation is you can only buy one.
It's all because of availability. Newegg just has third party resellers right now that have any stock and that's why the price is through the roof. I don't think AMD has changed the MSRP on any of the cards.
Comments
Vega is more than 50% larger in die size, has 60% more transistors, and consumes 120+ watts more when going toe-to-toe with the GTX 1080. The 390x was much closer in die size and transistors to it's immediate competitor (the GTX 980) while consuming similar or less power than RX Vega 64.
Vega is not a step in the right direction. Fiji was a huge step in the right direction. Polaris was a small stumble, Vega is lame duck. AMD can't keep throwing transistors and TDP at the problem; they're at the wall. If they don't heavily innovate they will be even worse off next round because I guarantee you Volta will not be some 5-10% update like Intel's CPU's were for years which allowed AMD to catch Intel in the CPU space.
They crammed 60% transistors into a 50% space, which is impressive in and of itself (6 lbs of crap in a 5lb sock?). And given that 390X and Fury X's TDP were 275W, to only come in at 295W means the power draw only went up by about 10%. But we got about 50% more computing/graphics power out of the generation, and nearly double the memory bandwidth & texture fill rate.
It's better than Fury's launch sitatution, which apart from HBM1 was a pretty lackluster improvement over the Hawaii. So much so that AMD not only kept Hawaii around, but refreshed it just one month prior to Fiji's release as the 390X (with only very minor clock bumps and a VRAM upgrade), and included it in their marketing for people who where so worried about the 4GB VRAM limitation on Fury.
So taken in light of the competition, yeah, doesn't look so good. But here your making sound like AMD hasn't improved - they have, and Vega represents a lot more improvement than Fiji did. And honestly, in today's chips, TDP is whatever you want it to be with PowerTune/Turbo settings within the operating constraints of the silicon - it's just a matter of how high your willing to go: nVidia didn't want to fool with water cooling, because they didn't have to in order to claim a performance crown, but I would be willing to bet dollars to donuts that if the field were reversed (like, oh say, Fermi), they would have no problem cranking up the clocks and TDP be damned in order to be competitive.
Not to take away from the fact that nVidia is doing a better job, but when you do stop to think about it, what AMD is impressive in it's own right. It's just too little too late. Like coming out and scoring in an NBA game, but only after your opponent has run up 37 points, or hitting that buzzer beater, even though you lost the game.
I do think your right about Vega/Navi vs Volta, that even estimating aggressively, AMD is several years away from retaking any performance/efficiency crowns away from nVidia. But that doesn't mean it's impossible, the tables have turned a number of times over the generation, and all it takes is one good architecture and a lucky break. Every manufacturing process is a bit different, and one particular design may work great on it, then utterly suck on the next one around.
Besides, we don't ~need~ AMD to beat the 1080Ti, as much as I would have liked to see computing power overall advance and push the envelope a bit more. We just need AMD to remain solvent and competitive. AMD is still very much competitive with Polaris in the low-mid tiers, and in a lot of ways superior to nVidia, and that's a much higher volume market.
Vega will probably sell about like Fiji did, as there are just too many parallels here: tepid but not insignificant initially, and one good price cut and it all of a sudden becomes a very good looking buy for consumers (albeit with, probably, nearly no margin for AMD). Vega may even do a bit better than Fiji without the price cut now that FreeSync is out in the wild and there's a bit of installed monitor base out there. And there's always the volatility of the mining segment, and we've all see what impact it can have on things. Right now Vega doesn't look that attractive based on perf/watt for miners, but that could easily change.
Imagine a world where nVidia's only graphics competition is Intel.
Yeah - once the inventory reduction sales have cleared the prices for new last-gen hardware go through the roof.
https://www.reddit.com/r/Amd/comments/6xqpot/vega_64_435_mhs_130w/
It's not double the performance of Polaris, but it is double the efficiency.
Granted, MSRP for Polaris (580) is $220US, and for Vega64 is $500, so it's more than double the up front price, and availability and price markup is still an issue for all of them. I'm not sure how the increase in efficiency will play out with ROI for miners, but it's probably safe to assume that it has gotten their attention.
Outside of combo deals, RX580's are going for $725US right now on Newegg, whereas reference Vega64's are going for $699.
The price on the RX580 seems odd, they had a bunch at Microcenter the other day for $220, there only stipulation is you can only buy one.