[BitsAndChips]390X ready for launch - AMD ironing out drivers - Computex launch

Page 31 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
How come 4GB is fine for high end cards people are buying now but in 1-2 months suddenly a card with 4GB is no good?

4GB? Depending on what card you are talking about it can be 3.5GB and that's absolutely no problem. So, if the 980 ti is 6GB and the 390 is 4GB that one difference alone will end up the only reason to purchase a card. You will absolutely need 6GB and nothing less at that level of performance. Yet when 7970 was 3GB and 680/770 were 2GB, and the 290 was 4GB and the 780's were 3GB, that made no difference whatsoever. Oh sorry, efficiency will matter as well. Unless AMD pulls a rabbit out of a hat and meets/beats nVidia there, then it won't be important again. (Except to those who like AMD. ;))
 

rgallant

Golden Member
Apr 14, 2007
1,361
11
81
4GB? Depending on what card you are talking about it can be 3.5GB and that's absolutely no problem. So, if the 980 ti is 6GB and the 390 is 4GB that one difference alone will end up the only reason to purchase a card. You will absolutely need 6GB and nothing less at that level of performance. Yet when 7970 was 3GB and 680/770 were 2GB, and the 290 was 4GB and the 780's were 3GB, that made no difference whatsoever. Oh sorry, efficiency will matter as well. Unless AMD pulls a rabbit out of a hat and meets/beats nVidia there, then it won't be important again. (Except to those who like AMD. ;))
how about 4gb HBM for game data and a 2gb cut down binned chip [defective] HBM for other stuff , should be ok per 970 owners.= 6gb of vram [on paper]. yea

just kidding
 
Feb 19, 2009
10,457
10
76
Has there ever been any confirmation whether Fiji is going to have high Double Precision compute performance? Or will it be focused only on gaming loads and Single Precision computing, as Tonga and all the Maxwell chips are?

The reason I ask this is because it wouldn't make much sense to focus on DP performance if the card is limited to 4GB by the restrictions of HBM. Most serious professional users who depend on DP are going to want a lot more. AMD's current HPC flagship, the FirePro W9100 (Hawaii-based), has a whopping 16GB of RAM on board. The original Titan's 6GB would seem to be a bare minimum for these kind of scientific and mathematical tasks.

At least some regression in RAM capacity is almost certainly going to happen on FirePro Fiji. No one has ever suggested that 16GB HBM is even technically possible at this time. The only way around this - and I want to emphasize that this is pure speculation on my part - would be if AMD had a memory subsystem that could treat HBM as, essentially, the next level of cache, and then access GDDR5 on top of that if the HBM runs out. Imagine a graphics card with 4GB of ultra-fast HBM and 16GB of GDDR5 backing it up. Does AMD have the technical ability to do this? I looked for some information on the Xbox One's ESRAM cache, and it's not clear exactly how much of this is automated like a normal cache and how much requires developer involvement. Apparently it's needed for the developers to flush it once in a while, though that could be due to its very small size, and the 'necessity' could be improved performance rather than "it'll break if you don't do this". We know that Intel has this kind of technology, with Iris Pro's EDRAM that basically acts as L4 cache.

Good post and good points. Indeed, FirePros have been doing very well gaining marketshare, the good DP compute + high vram definitely is a major selling point.

AMD may be doing an NV and skipping this gen of HPC refreshes. NV did a GK110 x2 for high SP SKU, since GM200 lacks decent DP compute, they are waiting for Pascal for that next update on HPC lineup.

On HBM2, vram capacity will be ample.
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,476
136
Has there ever been any confirmation whether Fiji is going to have high Double Precision compute performance? Or will it be focused only on gaming loads and Single Precision computing, as Tonga and all the Maxwell chips are?

The reason I ask this is because it wouldn't make much sense to focus on DP performance if the card is limited to 4GB by the restrictions of HBM. Most serious professional users who depend on DP are going to want a lot more. AMD's current HPC flagship, the FirePro W9100 (Hawaii-based), has a whopping 16GB of RAM on board. The original Titan's 6GB would seem to be a bare minimum for these kind of scientific and mathematical tasks.

At least some regression in RAM capacity is almost certainly going to happen on FirePro Fiji. No one has ever suggested that 16GB HBM is even technically possible at this time. The only way around this - and I want to emphasize that this is pure speculation on my part - would be if AMD had a memory subsystem that could treat HBM as, essentially, the next level of cache, and then access GDDR5 on top of that if the HBM runs out. Imagine a graphics card with 4GB of ultra-fast HBM and 16GB of GDDR5 backing it up. Does AMD have the technical ability to do this? I looked for some information on the Xbox One's ESRAM cache, and it's not clear exactly how much of this is automated like a normal cache and how much requires developer involvement. Apparently it's needed for the developers to flush it once in a while, though that could be due to its very small size, and the 'necessity' could be improved performance rather than "it'll break if you don't do this". We know that Intel has this kind of technology, with Iris Pro's EDRAM that basically acts as L4 cache.

AMD's R9 390X according to leaked slides has 8 GB HBM using a dual ink interposing method.

http://videocardz.com/55146/amd-radeon-r9-390x-possible-specifications-and-performance-leaked

btw that is using 4Hi HBM stacks of 1GB each. Hynix has been producing the 1 GB 4Hi HBM chip from late 2014.

https://www.skhynix.com/products/support/databook.jsp

Hynix also supports 8Hi HBM and will start manufacturing some time later in H2 2015. btw this is a recent change from earlier presentations which said only 4 Hi will be supported with HBM1.

http://www.hardwareluxx.de/images/stories/newsbilder/aschilling/2015/gtc2015/gtc2015-skhynix-2.jpg

There is a small mistake in the slides comparing 4 Hi HBM1 and 8 Hi HBM1 with 4Hi HBM2 and 8 Hi HBM2

http://www.hotchips.org/wp-content/...Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf

According to Hynix graphics databook 4 Hi HBM1 supports only 1GB using 2 Gigabit DRAM chips. This 2 Gigabit DRAM chip will be connected to 2 128 bit channels. The latest Hynix slide confuses that to 2 Gb per 128 bit channel.

So in summary we have 4Hi HBM1 with 1 GB capacity shipping now and 8 Hi with 2 GB capacity shipping later. I believe the 8 Hi chips will be used for the Firepro versions of R9 390X. So with 8 chips using dual link interposer we have ( 2 * 8 = 16 GB) capacity.

AMD co-designed HBM with Hynix and presented to JEDEC who adopted it as JESD235 standard in Oct 2013. AMD has been working on HBM from 2010. AMD very well knew the limitations of HBM1 and thats why they designed the dual link interposer. We are likely to see a Firepro based on R9 390X sometime in Q4 2015 or early Q1 2016.
 

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
If the 390x does indeed only ship with a 4GB variant, I see this is a sign of desperation from AMD.

Using less VRAM than the market expects is not a sign of desperation but rather a sign of failing to meet the market's needs by under-delivering/missing out on those expectations. Desperation is the exact opposite - it's when you throw useless amounts of VRAM on a card to justify high price premiums (for example R9 290X 8GB or putting too much VRAM on slow cards like FX5200 or GT630). Desperation would be throwing excessive VRAM on a card now fast enough to take advantage of it to justify the high price. Titan X comes to mind here as well, or 6GB of VRAM on the original Titan - pure waste for gaming.

8GB of HBM i'm going to assume is not at all cheap to produce, and given their position in the market comaperd to NVidia, AMD can't charge the same premium and still sell their products as readily as NVidia can, hence the compromise of only a 4GB card.

1) It's already been noted by more than 1 source that dual-link interposer allows for HBM1 to support 8GB on 1 GPU. Therefore, if we are to believe those engineering leaks of how dual-link interposer works, why would "only 4GB card" be the case? I really don't get where you got this from and why so many other posters keep ignoring all the slides and sources that did discuss dual-link interposer?

2) Even the supposed AMD leaked slides show up to 8GB of HBM1 for 390X WCE, which suggests there will be 4GB and 8GB card options.

The long term question is going to be how well it stacks up to say a 6GB 980Ti in a years time.

The better question is: Is it even worth it to pay a premium for > 4GB of VRAM for this generation for non-4K gaming when this gen is shaping up to be a stop-gap generation?

Real world scenario: one can just buy a $500 R9 390 4GB, save $ on not buying the R9 390X 8GB/GM200 6GB, and just reinvest the savings + resale value on 40-50% faster 14nm/16nm GPUs around May-June 2017. This strategy has been so bullet-proof for decades, it's hard to imagine why it would fail again. With so much historical data behind us, the fastest cards have never been able to outlive the 2nd tier cards from AMD/NV. Pick any generation you want and this is true. 6800UE didn't outlive 6800GT, X850XT PE didn't outlive X800XT, 8800GTX Ultra didn't outlive 8800GTS 512MB, GTX680 didn't outlive GTX670, HD7970 didn't outlive HD7950, etc. I remember how Newegg had fire-sale on a solid Asus DirectCUII GTX480 for $299 while the brand new GTX580 was $499. Just to have the latest and greatest, many people ignored the 480 and threw $200 at the 580. How did that work out? Pure waste of $ right there. 580 never outlasted the 480 either.

What makes you think R9 390X/GM200 6GB are even worth the money over their slightly cut-down alternatives?

Historical reflection: was it worth it to buy GTX780 6GB to "future-proof", or was it worth it to pay $700 for a GTX780Ti to "future-proof" over 780Ghz? If you are objective and look at the benchmarks and evolution of price/perf. in the last 1.5 years, then the answer is: definitively NO. With GPU landscape evolving so fast, it's way better to upgrade more often than spend $200-300 extra per card for that minimal 15% performance gain. And when we discuss CF and SLI, things go bonkers in no time. 1.5 years ago GTX780Ti SLI cost $1400 and R9 290X CF cost $1100. Today, you can buy that performance for just $600 in GTX970 SLI / R9 290X CF. In reality, by mid-Sept 2014 we had $1400 GTX780SLI performance in a $660 GTX970 SLI; so it took less than 1 year! The whole idea of paying hundreds of dollars extra today on the ultra high-end to future-proof for 'years time' is bullocks and it has never worked. That's why while most people on this forum are salivating over 980Ti and R9 390X for the rumoured $799 price, the sleeper cards imo will be a cut-down GM200 and R9 390 that when overclocked should deliver 85% of the performance of the best cards for a fraction of their prices.

I would say for anyone gaming on 4K or higher or multiple 1440P screens, 6-8GB is worth it but otherwise, paying hundreds extra to have 8GB for future-proofness is going to be a waste of money for this gen. With HD7950/7970 vs. 670/680 it was different because AMD never charged hundreds of dollars extra for more VRAM -- it was just a free bonus. However, when a gamer is faced with paying $100+ more to go from 4GB to 8GB, it's far less clear cut. That means we won't be able to tell how much of a 'failure' a 4GB 390 card is until we see the price/performance of a GM200 6GB equivalent competitor.
 
Last edited:

2is

Diamond Member
Apr 8, 2012
4,281
131
106
When I say "compromise" I'm not talking about a technical compromise, but one of cost savings. I thought this was pretty clear. Anyway, that was all "IMO" and it still is. You're obviously free to disagree with it. It's not a secret that our view on AMD isn't the same.
 

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
When I say "compromise" I'm not talking about a technical compromise, but one of cost savings. I thought this was pretty clear. Anyway, that was all "IMO" and it still is. You're obviously free to disagree with it. It's not a secret that our view on AMD isn't the same.

I think you used the word desperate in the wrong context. Compromise would have a completely different meaning. Having said that you and I are probably in a different market segment for GPU purchases. I am the type of guy who would go for 2nd best cards in pairs rather than the absolute top cards. The only reason I picked HD7970s last round is because all those extra shaders contributed to mining. Otherwise I would have picked up 670s or 7950s. All my NV/AMD cards before that were never flagships. I find the best cards a total waste of $ and I'd rather upgrade more frequently to 2nd tier cards. Since 7970s cost me nothing, I didn't really care about their price/performance vs. 7950s. However, if today I am faced with a choice between an $500 R9 390 4GB with 85% of Titan X's performance and an AMD/NV card with 105% of Titan X's performance for $799, I wouldn't even look at the latter. Also, since I am not gaming at 4K, overpaying for 8GB cards would be a total waste for me. Based on how many gamers bought 4GB 290X CF, 970/980 4GB SLI, I think millions of gamers agree with my view that 8GB is a total waste for anything but 4K or multiple 1440P monitors.

As long as AMD provides both 4GB and 8GB options, I don't have a problem. While you might view 4GB as a compromise, I view 8GB as a money grab for resolutions below 4K and I'd much prefer the 4GB option so I don't have to throw $100+ per card on something I'd never use. I think AMD realizes there are different customers and should provide these 2 options. Certainly my personal view looking at 20 years of GPU history is the fastest cards with 15-20% more performance and more VRAM never outlast 2nd tier cards (i.e., 580 OC never outlasted a 570 OC). I generally recommend going for more VRAM or spending a bit more if the price/performance curve scales well, but it doesn't look like it will be the case with the top cards this gen. That's why I am way more interested in AMD's/NV's $500-550 options.

BTW, in hindsight, we can probably agree that 680 2GB SLI never outlasted 670 2GB SLI for gaming. So basically no matter the gen you pick, the top cards are only worth it if you can afford to get them every single generation. For example, someone who bought 670 SLI and sold them, would have used $200 saved from not getting 680s that you did and upgraded to 970 SLI (or has $200 extra for this round). The point is that last 10-20% is usually a total waste of $ when it comes to future-proofness. With 14nm GPUs less than 2 years away, this makes the 4GB vs. 8GB argument moot for anyone not using 4K or multiple 1440P monitors. I bet in 18 months a $400 card will match $700 GM200 6GB.
 
Last edited:

2is

Diamond Member
Apr 8, 2012
4,281
131
106
There was no 670 when I bought my 680, nor was there a 4GB option at the time. It was either that or the 7970, which at the time cost more and didn't perform as well. I actually don't have a problem with 2nd best cards if the only difference is clock speed.
 

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
There was no 670 when I bought my 680, nor was there a 4GB option at the time. It was either that or the 7970, which at the time cost more and didn't perform as well. I actually don't have a problem with 2nd best cards if the only difference is clock speed.

That's not a great explanation. How many generations before that one did NV's 2nd tier card provide nearly the same performance compared to the top card? Many. For starters, GTX470 OC vs. 480 OC and 570 OC vs. 580 OC. Before that GeForce 4400 vs. 4600, 5900XT vs. 5900U, 6800GT vs. 6800U, I could go on. We knew from day 1 when 680 launched that 670 would be a better buy anyway. This continued with 970 vs. 980. All of this is not surprising given NV's track record. That's why the entire focus on 390X vs. 980Ti at $799 level is pretty amusing for those of us who have paid attention wrt to 2nd tier vs. 1st tier GPUs.

It's even more important to point out that on the AMD side, since AMD retains the back-end (ROPs and memory bus) more or less completely intact, once their 2nd tier and top tier cards are overclocked, the performance difference shrinks from 15% to about 5-7%.

Now if AMD has a 900mhz 3840 SP 4GB R9 390 for $500 and a 1050mhz 4096 SP 8GB R9 390X for $800, it might seem there is a major gap in performance here that justifies the premium - 24%. However, 5 minutes of overclocking and both cards running at 1150mhz, we are down to 7% performance delta. If you paid attention to 5850 vs. 5870, 6950 vs. 6970, 7950 vs. 7970, R9 290 vs 290X, this happened every single AMD generation, except for 4850 vs. 4870. That's why your point on 4GB vs. 8GB misses the most criticial piece - where does R9 390 come in vs. R9 390X in terms of core specs?

"Our look at the reference Radeon HD 7950 and HD 7970 cards shows there to be a 15 per cent gap in performance in favour of the range-topping GPU. But raise the HD 7950's clocks to the higher speeds of the 7970 and the gap melts away to around five percent, often a little less, depending upon gaming title."
http://hexus.net/tech/reviews/graphics/34761-amd-hd-7950-vs-hd-7970-clocks/?page=10

That's why if R9 390 packs the same 4096-bit bus, the same number of ROPs and doesn't suffer a large SP/TMU reduction in specs compared to R9 390X, but has 4GB HBM1 for hundreds of dollars less, R9 390X is an easy skip for those of us not on 4K. The emphasis in this thread on 4GB being a failure/disappointment completely misses this crucial dynamic that has existed between 2nd tier and the flagship AMD card since HD5800 gen! As I have repeated, once you go dual cards, it gets even more dramatic ($1000 R9 390s vs. $1600 R9 390Xs for 5-8% more performance once both are overclocked if history is anything to go by?! Ya the 390Xs would be a giant waste of $ for 1080P-1440P gaming against the standard 390s). All of a sudden, $600 saved towards 14nm GPUs and the entire argument of 4GB vs. 6-8GB is moot. And that's how a $500 R9 390 could be the real star of the entire show.
 
Last edited:

raghu78

Diamond Member
Aug 23, 2012
4,093
1,476
136
That's not a great explanation. How many generations before that one did NV's 2nd tier card provide nearly the same performance compared to the top card? Many. For starters, GTX470 OC vs. 480 OC and 570 OC vs. 580 OC. Before that GeForce 4400 vs. 4600, 5900XT vs. 5900U, 6800GT vs. 6800U, I could go on. We knew from day 1 when 680 launched that 670 would be a better buy anyway. This continued with 970 vs. 980. All of this is not surprising given NV's track record. That's why the entire focus on 390X vs. 980Ti at $799 level is pretty amusing for those of us who have paid attention wrt to 2nd tier vs. 1st tier GPUs.

It's even more important to point out that on the AMD side, since AMD retains the back-end (ROPs and memory bus) more or less completely intact, once their 2nd tier and top tier cards are overclocked, the performance difference shrinks from 15% to about 5-7%.

Now if AMD has a 900mhz 3840 SP 4GB R9 390 for $500 and a 1050mhz 4096 SP 8GB R9 390X for $800, it might seem there is a major gap in performance here that justifies the premium - 24%. However, 5 minutes of overclocking and both cards running at 1150mhz, we are down to 7% performance delta. If you paid attention to 5850 vs. 5870, 6950 vs. 6970, 7950 vs. 7970, R9 290 vs 290X, this happened every single AMD generation, except for 4850 vs. 4870. That's why your point on 4GB vs. 8GB misses the most criticial piece - where does R9 390 come in vs. R9 390X in terms of core specs?

"Our look at the reference Radeon HD 7950 and HD 7970 cards shows there to be a 15 per cent gap in performance in favour of the range-topping GPU. But raise the HD 7950's clocks to the higher speeds of the 7970 and the gap melts away to around five percent, often a little less, depending upon gaming title."
http://hexus.net/tech/reviews/graphics/34761-amd-hd-7950-vs-hd-7970-clocks/?page=10

That's why if R9 390 packs the same 4096-bit bus, the same number of ROPs and doesn't suffer a large SP/TMU reduction in specs compared to R9 390X, but has 4GB HBM1 for hundreds of dollars less, R9 390X is an easy skip for those of us not on 4K. The emphasis in this thread on 4GB being a failure/disappointment completely misses this crucial dynamic that has existed between 2nd tier and the flagship AMD card since HD5800 gen! As I have repeated, once you go dual cards, it gets even more dramatic ($1000 R9 390s vs. $1600 R9 390Xs for 5-8% more performance once both are overclocked if history is anything to go by?! Ya the 390Xs would be a giant waste of $ for 1080P-1440P gaming against the standard 390s). All of a sudden, $600 saved towards 14nm GPUs and the entire argument of 4GB vs. 6-8GB is moot. And that's how a $500 R9 390 could be the real star of the entire show.

When was the last time we did not see a VRAM increase from one generation to another at the high end for an AMD GPU ? Maybe HD 4890 1GB to HD 5870 1GB. Even there we had few HD 5870 2GB skus. For 3 generations we have seen AMD increase VRAM - HD 6970 2GB, HD 7970 3GB, R9 290X 4GB (with few SKUs with double VRAM 4/6/8GB). The very reason for AMD to design the dual link interposer was for higher capacity.

Sorry but if you are buying a USD 550+ GPU in 2015 then you better pack more than 4 GB VRAM. R9 390 needs to improve on VRAM size over the R9 290 and R9 290X from just a generational comparison leave alone the competition GM200 with 6 GB.

My guess is the R9 390 is going to be fighting Titan X with similar perf and there is no way 4 GB is enough for such a powerful GPU. Even at 1440p a game like Dying Light or Watchdogs (MSAA 4x) or GTA V (with MSAA 4x) easily exceeds 4 GB VRAM at the highest settings.

http://www.hardocp.com/article/2015...gtx_titan_x_video_card_review/14#.VTs6TyGqqko

http://www.hardocp.com/article/2015...video_card_performance_preview/3#.VTs6ESGqqko

What is the point of such a powerful GPU if you cannot run at MSAA 4x or SSAA 4x depending on the game at 1440p. btw imo AMD's R9 380 and R9 380X will ship with 4 GB HBM and be perfect in the USD 300 - USD 500 segment. So people who want a 1440p card with 4 GB should pick these. For the rest who want to go above USD 500, AMD will have 8 GB R9 390 and R9 390X at USD 600 - USD 800 price range.
 
Last edited:

AtenRa

Lifer
Feb 2, 2009
14,003
3,362
136
90%+ of gamers will be fine with 4GB of ram today and I expect for the next few years for 1080/1440p. 4K will not be mainstream anytime soon.
Now for all of you 4K users, im expecting an 8GB card will be available. ;)
 

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
Sorry but if you are buying a USD 550+ GPU in 2015 then you better pack more than 4 GB VRAM.

Sure, there will be options like we had 7970Ghz 3GB and 6GB or 290X 4GB and 8GB. To imply that a $550 card absolutely needs 8GB of VRAM is absurd. Not everyone is a 1440P or 4K gamer.

My guess is the R9 390 is going to be fighting Titan X with similar perf and there is no way 4 GB is enough for such a powerful GPU.

Why is 4GB not enough for 1080P-1440P? According to you then 99.9% of all GPUs today are worthless now?

Even at 1440p a game like Dying Light

Did you see that review? R9 295X2 suffers from a driver issue in that old review. 980 4GB has no such problems. From the review:

"As you can see, with all the video cards operating at the highest settings the game allows, GeForce GTX TITAN X blows the doors off of everything in this game. AMD Radeon R9 295X2 cannot keep up even with the single-GPU GeForce GTX 980. Keep in mind both have 4GB of VRAM available, so if this was about VRAM bottlenecking, both would have the same trouble, yet GTX 980 is clearly significantly faster than AMD R9 295X2."

or Watchdogs (MSAA 4x)

I think the reviewer made a mistake here. Again, 980 has smooth performance but CF suffers badly. If 4GB was the culprit, 980 would have bombed too, but it doesn't.

This looks more like a CF/game driver issue for R9 295X2 than a VRAM bottleneck given how smooth 980's fps delivery is.

1428990246ap5USD9KPf_8_3_l.gif


or GTA V (with MSAA 4x) easily exceeds 4 GB VRAM at the highest settings.

Again, I think the review is making a mistake here talking about VRAM. R9 295X2 has higher minimums than the Titan X. How in the world is he concluding that 4GB of VRAM is the bottleneck? Besides R9 295X2's smoothness factor improves at 4K vs. 1440P in this game, which sounds like a driver/game issue.

1429511282q5iVvFquHG_3_3_l.gif


Also, MSAA in GTA V is one of the worst decisions:

1) The performance hit is nearly 50%.

2560_MSAA.png


By the time we get to 4K, the performance hit from 4xMSAA is so great, everything is unplayable. Thus, I don't agree with your insinuations that R9 390X/Titan X are some uber powerful cards -- they are for 1440P and below -- but for 4K, they won't be fast enough. If someone is buying 2 or more of those, 6-8GB starts to make a lot more sense, but if you are only in the market for 1 of those cards, the GPU is too slow imo to need 6-8GB. Maybe future games will prove me wrong but so far I haven't seen any game where this is the case. Right now I recommend the Titan X over 980 SLI not because of the extra VRAM, but because when SLI doesn't work, Titan X OC is 50% faster than 980 OC. That's a huge advantage that's hard to ignore. However, looking at pure GPU pixel shading power, the Titan X is way too weak for 4K on its own.

3840_MSAA.png


Notice how the Titan X is at 28 fps but R9 295X2 is at 36 fps and 980 SLI is at 37 fps? If VRAM was a bottleneck, those dual cards would never manage higher FPS than the Titan X. I myself made a mistake thinking GTA V actually needs > 4GB of VRAM but I looks like I was wrong. The game just dynamically varies the VRAM but it doesn't require it.

2) With so many shimmering textures in GTA V, MSAA doesn't solve any of those issues. It's a terrible method for this game to be honest as it has questionable IQ and insane performance hit.

What is the point of such a powerful GPU if you cannot run at MSAA 4x or SSAA 4x depending on the game at 1440p.

30-40% faster than a 980 isn't that powerful in the context of how inefficient MSAA works in modern games/engines. Also, it's WAY not enough to even talk about SSAA at 1440P. If we were talking about a card 2-2.5X faster than a 980, I'd agree. Again, what about all those gamers on 1080P and 1200P? Why would they want to spend extra for 8GB option? Sounds like a waste.

btw imo AMD's R9 380 and R9 380X will ship with 4 GB HBM and be perfect in the USD 300 - USD 500 segment. So people who want a 1440p card with 4 GB should pick these. For the rest who want to go above USD 500, AMD will have 8 GB R9 390 and R9 390X at USD 600 - USD 800 price range.

I think AMD needs R9 390/390X 4GB versions since they'll make a lot more sense for gamers at 1440P and below than the more expensive and rather wasteful 8GB versions. That's why I questioned the whole 8GB is required or a fail idea from the start. 6-8GB for 4K when discussing dual 390Xs? Fine. 8GB for 1080P-1440P? Overkill for now. You haven't provided good examples to prove how > 4GB helps at 1440P and below in any modern title. I am anxious to see Batman AK, Project CARS and the Witcher 3. All those 3 games should give us more data points to work with. The next Deus Ex, Star Citizen, Mad Max and Just Cause 3 might push the VRAM requirements further but those games are ways out.
 
Last edited:

raghu78

Diamond Member
Aug 23, 2012
4,093
1,476
136
Sure, there will be options like we had 7970Ghz 3GB and 6GB or 290X 4GB and 8GB. To imply that a $550 card absolutely needs 8GB of VRAM is absurd. Not everyone is a 1440P or 4K gamer.
Its pretty shortsighted to buy a GPU as powerful as R9 390X or R9 390 with 4GB VRAM. These GPUs have the grunt to power the latest games at 1440p with MSAA. But if you cripple them with 4GB then you are going to handicap them badly.

Why is 4GB not enough for 1080P-1440P? According to you then 99.9% of all GPUs today are worthless now?
4GB is fine for 1080p. For 1440p its just about barely enough as we see titles like WatchDogs and GTA V with MSAA which need the extra VRAM

Did you see that review? R9 295X2 suffers from a driver issue in that old review. 980 4GB has no such problems.
I think the reviewer made a mistake here. Again, 980 has smooth performance but CF suffers badly. If 4GB was the culprit, 980 would have bombed too, but it doesn't.
This looks more like a CF/game driver issue for R9 295X2 than a VRAM bottleneck given how smooth 980's fps delivery is.
My reason to show the hardocp reviews was to show the Titan X using more than 4GB at the highest settings in Wathdogs, Dying Light. GTA V with FXAA takes 4GB. So with MSAA its easy to see that go way above 4 GB. the R9 295x2 has erratic framerates because of the VRAM limit being hit. CF has VRAM overheads compared to single GPU and so the same 4GB is a handicap. GTX 980's third generation color compression is also helping to alleviate the problem quite a bit.

http://www.hardocp.com/article/2015..._gtx_titan_x_video_card_review/8#.VTtewfAXaZc

"The AMD Radeon R9 295X2 struggles, and here is why. The video card only has 4GB of VRAM available to it for each GPU. But wait you say, so does the GTX 980 and it performed smooth. Yes, but with two AMD Radeon R9 290X GPUs the performance with the AMD R9 295X2 has shot way up. Simply put, the performance potential is pushing the limits of VRAM capacity in this game because it wants to go faster and utilize more VRAM, but it cannot, so it bottlenecks. "

Again, I think the review is making a mistake here talking about VRAM. R9 295X2 has higher minimums than the Titan X.How in the world is he concluding that 4GB of VRAM is the bottleneck? Besides R9 295X2's smoothness factor improves at 4K vs. 1440P in this game, which sounds like a driver/game issue.
Sorry but GTX Titan X has way better minimums than R9 295x2 in hardocp gameplay runs. :biggrin:

Also, MSAA in GTA V is one of the worst decisions:

1) The performance hit is nearly 50%.
Even with MSAA 4x at 1440p, the GTX Titan X can do 40 - 60 fps in GTA V and is perfectly playable. If you want more fps you can overclock and gain another 15 - 20%. But the gameplay experience is good because Titan X does not run into VRAM bottlenecks at 1440p with MSAA.

By the time we get to 4K, the performance hit from 4xMSAA is so great, everything is unplayable. Thus, I don't agree with your insinuations that R9 390X/Titan X are some uber powerful cards -- they are for 1440P and below -- but for 4K, they won't be fast enough. If someone is buying 2 or more of those, 6-8GB starts to make a lot more sense, but if you are only in the market for 1 of those cards, the GPU is too slow imo to need 6-8GB. Maybe future games will prove me wrong but so far I haven't seen any game where this is the case. Right now I recommend the Titan X over 980 SLI not because of the extra VRAM, but because when SLI doesn't work, Titan X OC is 50% faster than 980 OC. That's a huge advantage that's hard to ignore. However, looking at pure GPU pixel shading power, the Titan X is way too weak for 4K on its own.
I disagree strongly. Titan X and R9 390X as a single GPU is capable of maxing any game at 1440p provided they have more than 4GB ( in the case of R9 390X) . Similarly anybody who wants to enjoy 4k gaming they have 2 options. If they are going to stay away from MSAA then a single R9 390X and Titan-X is enough especially when overclocked. But again they need more than 4 GB VRAM. For 4k with MSAA they need two R9 390X or two Titan-X. Then too they need more than 4 GB VRAM. So imo R9 390X needs greater than 4GB for sure.

Notice how the Titan X is at 28 fps but R9 295X2 is at 36 fps and 980 SLI is at 37 fps? If VRAM was a bottleneck, those dual cards would never manage higher FPS than the Titan X. I myself made a mistake thinking GTA V actually needs > 4GB of VRAM but I looks like I was wrong. The game just dynamically varies the VRAM but it doesn't require it.
you show me techspot benchmark which uses the in game benchmark and has no minimum fps. I show you hardocp benchmark which is a game play run with fraps recording and actual game play experience. which one do you think is more relevant :whiste:

30-40% faster than a 980 isn't that powerful in the context of how inefficient MSAA works in modern games/engines. Also, it's WAY not enough to even talk about SSAA at 1440P. If we were talking about a card 2-2.5X faster than a 980, I'd agree. Again, what about all those gamers on 1080P and 1200P? Why would they want to spend extra for 8GB option? Sounds like a waste.
Sorry I disagree that Titan X isn't that powerful. At the same clocks GM200 is roughly 40% faster than GM204. So you are likely to get the same performance with MSAA at 1440p that you get with FXAA at 1440p assuming MSAA takes a 30 - 40% perf hit.

http://www.computerbase.de/2015-04/...diagramm-grafikkarten-benchmarks-in-2560-1440

I think AMD needs R9 390/390X 4GB versions since they'll make a lot more sense for gamers at 1440P and below than the more expensive and rather wasteful 8GB versions.
thats why there is R9 380 and R9 380X. I am confident the R9 380 aka Fiji XT will be around 15% faster than GTX 980 and be perfect for 1440p gamers. These will be cost and power optimized GPUs. R9 380X with 3072 sp and 4 GB HBM priced at USD 400 - 450 makes a lot of sense. Since this chip will be around 400sq mm yields will be better and the pricing will be quite good.

That's why I questioned the whole 8GB is required or a fail idea from the start. 6-8GB for 4K when discussing dual 390Xs? Fine. 8GB for 1080P-1440P? Overkill for now. You haven't provided good examples to prove how > 4GB helps at 1440P and below in any modern title. I am anxious to see Batman AK, Project CARS and the Witcher 3. All those 3 games should give us more data points to work with. The next Deus Ex, Star Citizen, Mad Max and Just Cause 3 might push the VRAM requirements further but those games are ways out.

Whatever examples I provide you try to argue that MSAA is not necesary. So I think you are moving goalposts. Whether MSAA or SSAA is needed is upto the gamer to choose. The performance is there and the VRAM needs to be there. I am not going to agree that a GPU 40% faster than GTX 980 is well suited with 4GB especially for 1440p with MSAA and 4k with FXAA.
 
Last edited:

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
From the same review
In the graph above we are running at the playable settings of the GeForce GTX 980. The AMD Radeon R9 295X2 has more than enough performance to be faster than the GTX 980 here, but it isn't. You see, the VRAM is bottlenecking quite a bit with "Ultra" textures and keeping performance bottlenecked to low performance.
1428990246ap5USD9KPf_8_3.gif

The 980 is 4GB too. Where he's coming up with a VRAM bottleneck for the 295X2 being the issue compared to a card that has the same amount of VRAM, I don't understand.
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,476
136
From the same review
The 980 is 4GB too. Where he's coming up with a VRAM bottleneck for the 295X2 being the issue compared to a card that has the same amount of VRAM, I don't understand.

R9 295X2 is a CF solution. There is a VRAM overhead for multiple GPUs. So you do not get the full 4GB for storing data as you do with GTX 980.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
146
106
R9 295X2 is a CF solution. There is a VRAM overhead for multiple GPUs. So you do not get the full 4GB for storing data as you do with GTX 980.

What overhead would that be?

The only thig I can see is that 1 GPU may have to store and pass the frame from the second one. But thats not exactly much memory. So whats the amount of overhead and what does it go to? If we talk 10-20MB its essentially 0.
 

railven

Diamond Member
Mar 25, 2010
6,604
561
126
It's almost unfathomable. But you have to leave the knowledgeable realms; we do not represent a significant margin of the userbase. Go to places like NeoGAF. Read their GPU discussions, and PC recommendations. It's frightening how some of the masses think.

No kidding. I don't consider myself any where near as knowledgeable as some of the posters where, but holy cow - @Neogaf I feel like Stephen Hawkins.

And that place got some weird AMD obsession because those are the chips in their beloved consoles.

At this point I'm so itching to buy a new monitor, and with it a shiny new vidya cawd!
 

2is

Diamond Member
Apr 8, 2012
4,281
131
106
90%+ of gamers will be fine with 4GB of ram today and I expect for the next few years for 1080/1440p. 4K will not be mainstream anytime soon.
Now for all of you 4K users, im expecting an 8GB card will be available. ;)

Top of the line cards aren't marketed towards the 90% nor are they priced for the 90%

So this is likely a card that's priced for the 10% with ram adequate for the other 90 if these rumors are accurate.
 

stahlhart

Super Moderator Graphics Cards
Dec 21, 2010
4,273
77
91
This thread is for discussing the upcoming R9 390X, not for stalking other forum members. Stay on topic.
-- stahlhart
 

tential

Diamond Member
May 13, 2008
7,348
642
121
No kidding. I don't consider myself any where near as knowledgeable as some of the posters where, but holy cow - @Neogaf I feel like Stephen Hawkins.

And that place got some weird AMD obsession because those are the chips in their beloved consoles.

At this point I'm so itching to buy a new monitor, and with it a shiny new vidya cawd!

Neogaf or any other more casual forum.
And I don't know what you're talking about. Over on NEOGAF, AMD is bashed like no other there. It's a nonstop hate fest.