[UPDATE] AMD Announces Polaris Architecture - 4th Generation GCN Arriving in Mid 2016

[Notional]

Intel's lived up to it twice, once with Nehalem and again with Sandy Bridge.

 

Intel's NAND is 22nm FinFET. It should come as a surprise to no one that Samsung's 20nm or 14nmFF is better at being low voltage. Samsung's node designs are for low-power SOCs, not an all-rounder like Intel's. It's not that impressive.

 

And they have been fined for manipulation and lies also. None of the 3 big C/GPU manufacturers are innocent in this. But AMD seems to be the least worst of the bunch throughout history. 

 

Afaik we are still to see any products using 14 nm FF LPP (Low Power Plus) node products. Only their LPE (Low Power Early) are on market so far. If AMD can manage to launch CPU's on this node, I don't see any problem (the question of course if max hertz attainable). Either way, none of us can conclude anything on the matter until products have launched/been tested by non biased sources.

 

Well it's impressive when Samsung's 3D NAND is the best performing on the market. The 950 pro makes a joke out of Intel's 750. Sure the Intel is a tiny bit faster at very high queue depth, but slower at lower, but it also costs a LOT more and is a multitude times larger in size.

 

Samsung's/GloFo's 14nm FF LPP has still yet to show results, but we already see on the IPhone 6s, that Samsung's 14nm FF LPE (NOT LPP that AMD uses), is outperforming TSMC's 16nm FF. This will be very interesting to see the consequences for NVidia here.

[DarkBlade2117]

So I can assume this new architecture won't be good because they've been alternating between good and bad every generation-

tahiti- good

hawaii- bad

fiji- good

polaris- bad?

good

bad

good

FAN-FUCKING-TASTIC

 

I said it

[Notional]

good

bad

good

FAN-FUCKING-TASTIC

 

I said it

 

I still don't get why people say Hawaii is bad?!? It completely out performs NVidia's Kepler today, is still relevant and even outperforms mid-end Maxwell cards, especially in DX12.

That shitty stock cooler really screwed over AMD.

[shawno]

I still don't get why people say Hawaii is bad?!? It completely out performs NVidia's Kepler today, is still relevant and even outperforms mid-end Maxwell cards, especially in DX12.

That shitty stock cooler really screwed over AMD.

Agreed. My crossfired r290xs still crush 4k with ease. Not bad for a "bad" card...

[NuclearPeace]

Hawaii murdered GK200.

 

The 290 was faster than the 780, which costed $250 more at the time. It was doing it even by using the horrible reference cooler which actually have the 290 a disadvantage. Despite that, it was still faster than the 780.

 

The 290x was faster than the Titan which costed almost double at the time. Again, it was being weakened by the lack of cooling.

 

Hawaii was also much smaller than GK200 and had better compute abilities. Hawaii only used maybe 20w more than a GK200 chip in a worse case scenario. Plus, Hawaii had more memory controllers and thus its GPUs had more VRAM and bandwidth.

[patrickjp93]

And they have been fined for manipulation and lies also. None of the 3 big C/GPU manufacturers are innocent in this. But AMD seems to be the least worst of the bunch throughout history. 

 

Afaik we are still to see any products using 14 nm FF LPP (Low Power Plus) node products. Only their LPE (Low Power Early) are on market so far. If AMD can manage to launch CPU's on this node, I don't see any problem (the question of course if max hertz attainable). Either way, none of us can conclude anything on the matter until products have launched/been tested by non biased sources.

 

Well it's impressive when Samsung's 3D NAND is the best performing on the market. The 950 pro makes a joke out of Intel's 750. Sure the Intel is a tiny bit faster at very high queue depth, but slower at lower, but it also costs a LOT more and is a multitude times larger in size.

 

Samsung's/GloFo's 14nm FF LPP has still yet to show results, but we already see on the IPhone 6s, that Samsung's 14nm FF LPE (NOT LPP that AMD uses), is outperforming TSMC's 16nm FF. This will be very interesting to see the consequences for NVidia here.

 

I reserve judgment until the 770 appears. The 750 was mostly testing the waters. I don't think anyone expected a maxed out version of NVMe coming out so quickly.

[patrickjp93]

Hawaii murdered GK200.

 

The 290 was faster than the 780, which costed $250 more at the time. It was doing it even by using the horrible reference cooler which actually have the 290 a disadvantage. Despite that, it was still faster than the 780.

 

The 290x was faster than the Titan which costed almost double at the time. Again, it was being weakened by the lack of cooling.

 

Hawaii was also much smaller than GK200 and had better compute abilities. Hawaii only used maybe 20w more than a GK200 chip in a worse case scenario. Plus, Hawaii had more memory controllers and thus its GPUs had more VRAM and bandwidth.

Actually the bandwidth of the 290/X is lower than the 780TI. It's 320GB/s vs. 336 GB/s. (384/8) * 7.010 = 336.48 vs. (512/8) * 5.000 = 320.

 

Get your facts straight. Actually the 780 and 780TI wiped the floor with the 290 and 290X respectively until Maxwell launched, and while it's been proven Nvidia did NOT gimp Kepler, they have been ignoring it.

[SurvivorNVL]

I reserve judgment until the 770 appears. The 750 was mostly testing the waters. I don't think anyone expected a maxed out version of NVMe coming out so quickly.

We should hopefully see a host of NVME products at CES.  OCZ, Plextor, etc.  I'm still leaning more towards the Intel 750 for my personal build.  800gb capacity and PCIE instead of M.2.  That and an actual heatsink + power-loss-protection.  

[Notional]

I reserve judgment until the 770 appears. The 750 was mostly testing the waters. I don't think anyone expected a maxed out version of NVMe coming out so quickly.

 

The Samsung SM951 was just testing the waters. These products are what's on the market, and what is available. The 950 Pro is still a lot more impressive for the reasons I stated.

 

Seeing improvements is of course welcomed in newer products.

 

Actually the bandwidth of the 290/X is lower than the 780TI. It's 320GB/s vs. 336 GB/s. (384/8) * 7.010 = 336.48 vs. (512/8) * 5.000 = 320.

 

Get your facts straight. Actually the 780 and 780TI wiped the floor with the 290 and 290X respectively until Maxwell launched, and while it's been proven Nvidia did NOT gimp Kepler, they have been ignoring it.

 

Higher bandwidth on the 780ti sure, but those cards were still vram limited, when the next gen cards launched. Wiped the floor was only the case on GameWorks titles for obvious reasons. Still doesn't explain why Hawaii is still competitive with Maxwell 970/980.

 

We should hopefully see a host of NVME products at CES.  OCZ, Plextor, etc.  I'm still leaning more towards the Intel 750 for my personal build.  800gb capacity and PCIE instead of M.2.  That and an actual heatsink + power-loss-protection.  

 

That will be expensive af. Hopefully CES will show newer products at lower price and/or better performance. Power loss protection is nice, but heatsink is not necessary on the 950 pro though.

[patrickjp93]

We should hopefully see a host of NVME products at CES.  OCZ, Plextor, etc.  I'm still leaning more towards the Intel 750 for my personal build.  800gb capacity and PCIE instead of M.2.  That and an actual heatsink + power-loss-protection.  

 

I thought OCZ declared bankruptcy and was in the middle of dissolution.

[patrickjp93]

The Samsung SM951 was just testing the waters. These products are what's on the market, and what is available. The 950 Pro is still a lot more impressive for the reasons I stated.

 

Seeing improvements is of course welcomed in newer products.

 

 

Higher bandwidth on the 780ti sure, but those cards were still vram limited, when the next gen cards launched. Wiped the floor was only the case on GameWorks titles for obvious reasons. Still doesn't explain why Hawaii is still competitive with Maxwell 970/980.

 

 

That will be expensive af. Hopefully CES will show newer products at lower price and/or better performance. Power loss protection is nice, but heatsink is not necessary on the 950 pro though.

Hawaii isn't that competitive with Hawaii outside a small handful of titles.

 

The 950 Pro lacks both a heatsink (while being horrifically susceptible to thermal throttling) and power loss protection. Intel knows its user base isn't just performance enthusiasts, but workstation users.

[SurvivorNVL]

I thought OCZ declared bankruptcy and was in the middle of dissolution.

Nope.  They were gobbled up by Toshiba awhile ago, and since then have made a steady come-back in terms of reliability and performance / pricing.

[SurvivorNVL]

The Samsung SM951 was just testing the waters. These products are what's on the market, and what is available. The 950 Pro is still a lot more impressive for the reasons I stated.

 

Seeing improvements is of course welcomed in newer products.

 

 

Higher bandwidth on the 780ti sure, but those cards were still vram limited, when the next gen cards launched. Wiped the floor was only the case on GameWorks titles for obvious reasons. Still doesn't explain why Hawaii is still competitive with Maxwell 970/980.

 

 

That will be expensive af. Hopefully CES will show newer products at lower price and/or better performance. Power loss protection is nice, but heatsink is not necessary on the 950 pro though.

http://www.newegg.com/Product/Product.aspx?Item=N82E16820147467

http://www.newegg.com/Product/Product.aspx?Item=N82E16820167359

$20 difference, but the Intel is Intel and there is always NAND over-provisioning on that end, where as Samsung is always performance based.  More space to the Samsung, but more reliability to the Intel.  The 800GB is $700 now, but a week ago it was $569.

[Notional]

Hawaii isn't that competitive with Hawaii outside a small handful of titles.

 

The 950 Pro lacks both a heatsink (while being horrifically susceptible to thermal throttling) and power loss protection. Intel knows its user base isn't just performance enthusiasts, but workstation users.

 

It's competitive with both Kepler and Maxwell mid end cards. Not sure why you say otherwise.

 

Lol it doesn't. Let's look at PCPer's conclusion: http://www.pcper.com/reviews/Storage/Samsung-950-PRO-256GB-and-512GB-M2-NVMe-PCIe-SSD-Review/Thermal-Throttling-Conclusio

 

 

As you can see, you would have to write nearly 150GB at over 1.5GB/sec to get a 950 PRO to warm up enough to throttle, and when it does, the throttling is very minor, dropping to only 1.2GB/sec intermittently. The slightest airflow prevents this from happening at all, and even if there was zero airflow, the chances of maxing a 950 PRO out on writes for that long of a burst is extremely unlikely in even the most demanding consumer usage scenario.​

 

In fact they later talked about the airflow from a hot graphics card would prevent any and all throttling. So "horribly susceptible" is just outright FUD. Also lets look at their pros:

• Best performing M.2 SSD Tested
• Best performing SSD Tested (period)
• Power Consumption is half that of the closest competing NVMe SSD (that would be the Intel 750)
• M.2 form factor fits in more places (with adapter)

So the Intel really only has size and power failure as wins against this.

​

As for knowing their market. If we are talking workstation users, where any form of data loss is unacceptable, the user will have a UPS anyway.

 

http://www.newegg.com/Product/Product.aspx?Item=N82E16820147467

http://www.newegg.com/Product/Product.aspx?Item=N82E16820167359

$20 difference, but the Intel is Intel and there is always NAND over-provisioning on that end, where as Samsung is always performance based.  More space to the Samsung, but more reliability to the Intel.  The 800GB is $700 now, but a week ago it was $569.

 

Well that Intel SSD is 112 GB smaller in size at 20$ less. That's a lot more to pay for some capacitors. Either way, I would not buy either as they are too expensive. I'm sure we will see faster and much cheaper nvme drives this year.

[patrickjp93]

It's competitive with both Kepler and Maxwell mid end cards. Not sure why you say otherwise.

 

Lol it doesn't. Let's look at PCPer's conclusion: http://www.pcper.com/reviews/Storage/Samsung-950-PRO-256GB-and-512GB-M2-NVMe-PCIe-SSD-Review/Thermal-Throttling-Conclusio

 

 

In fact they later talked about the airflow from a hot graphics card would prevent any and all throttling. So "horribly susceptible" is just outright FUD. Also lets look at their pros:

• Best performing M.2 SSD Tested
• Best performing SSD Tested (period)
• Power Consumption is half that of the closest competing NVMe SSD (that would be the Intel 750)
• M.2 form factor fits in more places (with adapter)

So the Intel really only has size and power failure as wins against this.

​

As for knowing their market. If we are talking workstation users, where any form of data loss is unacceptable, the user will have a UPS anyway.

 

 

Well that Intel SSD is 112 GB smaller in size at 20$ less. That's a lot more to pay for some capacitors. Either way, I would not buy either as they are too expensive. I'm sure we will see faster and much cheaper nvme drives this year.

 

Anandtech came to a completely different result, so who's right? http://www.anandtech.com/show/9702/samsung-950-pro-ssd-review-256gb-512gb

After 6 seconds, without a direct fan, it's throttling, even with no dGPU nearby.

 

The 970 fights perfectly well against a 290/X and 390/X. The 980 punches at and above. The 980TI has nothing to worry about from an aftermarket Fury, and even the Fury X doesn't catch it in most cases.

 

AMD isn't as competitive against Nvidia as its loyal fanbase believes.

[Notional]

Anandtech came to a completely different result, so who's right? http://www.anandtech.com/show/9702/samsung-950-pro-ssd-review-256gb-512gb

After 6 seconds, without a direct fan, it's throttling, even with no dGPU nearby.

 

The 970 fights perfectly well against a 290/X and 390/X. The 980 punches at and above. The 980TI has nothing to worry about from an aftermarket Fury, and even the Fury X doesn't catch it in most cases.

 

AMD isn't as competitive against Nvidia as its loyal fanbase believes.

 

I don't see that anywhere (you need to link directly or quote). But The difference in testing is that PCPer wrote 150GB to the drive in one go, where Anandtech wrote data at full speed for one hour straight. Normal airflow in the case should eliminate all that. IF for some reason, your case is crap and your use case is extreme, you can always put a tiny heatsink on the controller like puget did for Barnacules computer.

 

What I can see is in that test however, is that the 950pro beats the crap out of Intel in pretty much all the tests.

In other words, Hawaii is plenty competitive to the 970, thus midend Maxwell. 390x trades punches with 980 in all the tests we see. NVidia only tends to win properly on GameWorks titles.

 

Noone talked about Fiji, although that is better at DX12 (when asynchronous compute is used).

[That Norwegian Guy]

My dream: a single-slot graphics card using a small amount of HBM at a price under $250. Render farm go. Something about 750Ti-ish power, but with like 2GB of HBM.

[gamervivek]

780 wiped the floor with 290 so much that 290 ended up being about 20% faster at 1440p.

 

http://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_Waterforce/23.html

 

The problem for AMD has been clockspeeds and the resulting power usage, once they correct it then nvidia are going to fall quite a ways behind.

[SurvivorNVL]

My dream: a single-slot graphics card using a small amount of HBM at a price under $250. Render farm go. Something about 750Ti-ish power, but with like 2GB of HBM.

My dream is a single slot GPU around $450 with the performance of two 980ti's and at the power of a single 980ti.  So, basically--four generations from now.

[Notional]

780 wiped the floor with 290 so much that 290 ended up being about 20% faster at 1440p.

 

http://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_Waterforce/23.html

 

The problem for AMD has been clockspeeds and the resulting power usage, once they correct it then nvidia are going to fall quite a ways behind.

 

Now you know why NVidia are so fixated on GameWorks and proprietary tech. They know they need vendor lock in to keep consumers. I don't envy NVidia's position at all. Pascal better be the second coming of jebus. It's pathetic to see how far the 780 series has fallen. That's planned obsolescence for you. I would be pissed if I had one of those cards today.

[patrickjp93]

I don't see that anywhere (you need to link directly or quote). But The difference in testing is that PCPer wrote 150GB to the drive in one go, where Anandtech wrote data at full speed for one hour straight. Normal airflow in the case should eliminate all that. IF for some reason, your case is crap and your use case is extreme, you can always put a tiny heatsink on the controller like puget did for Barnacules computer.

 

What I can see is in that test however, is that the 950pro beats the crap out of Intel in pretty much all the tests.

In other words, Hawaii is plenty competitive to the 970, thus midend Maxwell. 390x trades punches with 980 in all the tests we see. NVidia only tends to win properly on GameWorks titles.

 

Noone talked about Fiji, although that is better at DX12 (when asynchronous compute is used).

 

Actually Nvidia wins at a couple Gaming Evolved titles ironically enough.

 

No it isn't. Fury X loses to 980TI handily now at AOTS and Fable.

[SurvivorNVL]

-Snippity-

Well, in terms of more NVME goodness - http://www.tweaktown.com/news/49399/zotacs-new-pci-ssds-capable-more-2gb-sec/index.html  That's a thing.  Something tells me.. we're gonna start seeing NVME adoption from companies -really- fast.  Corsair NVME drives.. that'll be a thing.  "Neutron GTX XT-N"  Or something.

[Notional]

Actually Nvidia wins at a couple Gaming Evolved titles ironically enough.

 

No it isn't. Fury X loses to 980TI handily now at AOTS and Fable.

 

Sure, and AMD wins in AC:U which is a gameworks title. Point is that Hawaii beats the crap out of the 780 series today and is competitive to 970/980 cards.

 

None of those titles are out yet. But we know the dev of AOTS had to outright make a specific code path just for NVidia for it to work. Fable uses very little to no concurrent async compute. But I digress, it has nothing to do with Hawaii, although it will be interesting to see how much Kepler will get destroyed by Hawaii. I kinda expect Kepler to be better at DX12 than Maxwell.

 

Well, in terms of more NVME goodness - http://www.tweaktown.com/news/49399/zotacs-new-pci-ssds-capable-more-2gb-sec/index.html  That's a thing.  Something tells me.. we're gonna start seeing NVME adoption from companies -really- fast.  Corsair NVME drives.. that'll be a thing.  "Neutron GTX XT-N"  Or something.

 

Hot damn. Yeah, SATA is dead. Probably why we see ASUS boards have 1-2 U.2 ports on their new boards. I think NVME and M.2/U.2 will get a lot of focus this year. I'm not so interested in speed increases above 2GBps. I'd rather see massive price decrease. A 950 Pro should not cost 3x the price of their Sata drive with the exact same NAND. A controller is not that expensive.

 

So more competition on that area is very welcome. After all the NAND is ready and has been for a while. I will say though, that I don't like PCIe cards, as they take up too much space. I'd much rather see U.2 become standard, so we can get SATA SSD sized nvme beasts.

[SurvivorNVL]

-Snippity-

Sata-E was dead on arrival, and I still don't know why they are including it, and Sata in general is definitely dead.  Hopefully they swap over entirely to 1-4 U.2 per board, and the prices go down quickly.  Hopefully as well there is just more 1tb+ NVME ssds out there.  I will shell out $1000 for a 2tb NVME drive if they can deliver it.  I'm not that against PCIE, though.  Something about PCIE drives that just--it seems nice having as many slots populated.

That and I will never go over two GPUs anyway, and triple slots are just not my thing.

[Notional]

Sata-E was dead on arrival, and I still don't know why they are including it, and Sata in general is definitely dead.  Hopefully they swap over entirely to 1-4 U.2 per board, and the prices go down quickly.  Hopefully as well there is just more 1tb+ NVME ssds out there.  I will shell out $1000 for a 2tb NVME drive if they can deliver it.  I'm not that against PCIE, though.  Something about PCIE drives that just--it seems nice having as many slots populated.

That and I will never go over two GPUs anyway, and triple slots are just not my thing.

 

Sata-e is the dumbest connector I've ever seen, and I used to deal with ribbon cables and pata. Seriously you take 2x6GB connectors, and add them together with a 3 smaller connector, and you get 10GBps. A internal raid controller in a standard SSD with 2 Sata connectors would be better. Then they turn this frankenconnector into one ultra wide connector with 3 cables sticking out. At lest stack the sata connectors, similar to U.2. Ugh Sata-e must be developed by an idiot.

 

For sure. Samsung has already launched a 2TB sata drive, and will have larger drives this year. Popping an nvme U.2 controller on it, should not be an issue. Hopefully nvme prices will fall a lot.

 

That sounds more like an issue with ATX being too large and redundant in PCIe connectors. Most people, even in here don't need more than ITX, and for those who does, mATX should be more than enough for their needs. ATX needs to die as a mainstream size.