AMD announces Radeon M300 series notebook video cards

Significantly more performant than the R9 290X using half the wattage. After looking at the R9 Nano kinda do want to ask the question, couldn't they make a cutdown version of it and place it in a laptop.

I am very curious as to the cost and actual performance of the R9 Nano.

 

We'll find out later. As the nano is being marketed for their quantum machine (and potentially other small form factor gaming machines like steam, etc.), it seems possible to release a cutdown version for mobile. At 175W, it is a small jump to get it mobile ready...

 

Tbh I think the people that don't have money burning holes in their pockets and aren't in dire need of a computer should wait 1 more year for 2016 for both Pascal and R9 400 series. Simply because of the huge die shrink of 28-16/14 nm nodes, which I'm willing to bet will have massive accompanying performance upgrades. Not to mention HBM2 etc. I think that will be the prime time to upgrade and get your moneys worth. Seems like we've been in 28nm for ages..

 

Sadly I've fallen into the chasm of the great eternal cash inferno, where I was bitten by the annual upgrade bug. A slave to its whims and desires, I have become.

Great to hear about the Nano TDP! AMD could obliterate NVIDIA in notebooks with their HBM, with it being so energy efficient and with 512GB/sec bandwidth. Now all they need to do is release it.

Sent from my Nexus 5 using Tapatalk

 

No uses half as much power. But even that's not true anymore now that Fury Nano is 175W, which is more than half of the 290-300W that 290X is.

 

Well, the R9 M295X wasn't a bad move for AMD, it's just too little too late for what the GTX 900M series offers. I mean, no one in their right mind would go for the M295X when the 970M soundly beats it and costs the same/less. It's just a no-brainer.

 

Cost isn't the biggest issue, heat is. At 125W the M295X is hard to cool even in an 18.4" Alienware. Remember the furnace that was 880M? Yeah that was a 120W card, and the M295X has 5W extra TDP.

 

Isn't mxm3.0b limited to 256gbs?

 

No, only GDDR5 is limited to 256-bit. HBM takes up much less space and as a result can go much wider. To get 256 GB/s on a 256-bit bus requires 8 GHz GDDR5, which doesn't even exist on desktop cards unless you overclock.

 

Well, as much flak as the Alienware 15, the only other computer with the m295x, gets (and some of it is well deserved), it can cool the R9 m295x just fine. (Images courtesy of notebookcheck)

Spoiler: AW m295x temps

78 C. Higher than Maxwell 2, but it's below 80 C.
I'm thinking Apple overclocked their R9 m295x too much (+127 MHz on Core, + 112 MHz on Memory). Combined with Apple's poor cooling solution to let thermal throttling happen, then end result is... horrific on their end.
Maybe the R9 m390x will do better, but that has yet to be seen.

 

To address this, I want to clear up confusion in this thread, mxm3.0B is a form factor interface. What is connected to it? Usually either a number of pcie 2.0 or 3.0 (in laptops it varies but most high end have 4+ lanes). The memory had its interface to the gpu. Most on this form factor are 256-bit for higher end cards (as of the last PUBLISHED spec more than 3 yrs ago, without knowing of other revisions). HBM has MUCH higher bandwidth between the memory and the GPU. This DOES NOT effect how much data can be sent over the pcie lanes UNTIL they are saturated and become a bottleneck. Just because a spec limits a specific part of a card DOES NOT mean it limits the pcie lane bandwidth!!! It's something that I've seen in this thread and others that I want to dispel! So HBM on mxm cards IS possible! Now, revisions on the spec may be needed so that they can sell it as an mxm3.0b card, but that is an issue of revising a standard, NOT a technical impairment!!! enjoy!!!

 

You know, it is kind of those things like the eye can only see 30fps, that has a strong prevalence in the face of actual information or critical thinking haha Keep at it. It's just like the lack of knowledge regarding TDP regulation on haswell CPUs, that drive people crazy thinking many different things.

 

I'll be honest, for the longest time I believed the eye can "only" register 30FPS myth having been proved wrong. Personally I've found that my sensitivity between 30FPS and 60FPS isn't that bad. I can play at 30FPS and not rage about how it's not 60FPS. It really has to drop down into the teens before I start to get annoyed.

 

Oh, there is always different preference for any fps. It's not that people get crazy or not, its just that it is very different to claim you can't see above/below something, compared to saying you don't mind certain fps ranges.

Now that I am gaming at 96hz, even 60fps seems like the bare minimum now haha. This doesn't mean I can't game 30fps. Some games just don't run well enough, specially on consoles and I still enjoy them. But if possible, I would always prefer higher fps, regardless of the game.

 

oh man, I have the most elite desktop and I still love my consoles. Call me a child of the generation but I will never give them up. I love how standardized it is and how when you buy a game on them you generally know it will play well.

 

I'd say, you know it will play. You can't have any guarantees it seems, as I have experienced anything from bad performance, to game breaking glitches haha

I play them for exclusives and because it doesn't hurt to have a different system to relax and game on. But one of the compromises you get with a consoles, is that you can't determine your performance on your own, so you will have to make do with whatever is given to you. And lets face it, even with old consoles, we could finish games that often ran at 10 fps....

 

Same here.
Anything above 25 FPS for instance seems to be fine with me (28 fps is basically the same like 30).
You might be able to notice some more 'smoothness' to fast pacing scenes at 60 fps (compared to 30), but to me, its hardly noticeable (actually, I don't think I can detect it to begin with).
Certainly not anything to get upset about or having a ridiculous demand that anything below 60 fps or even 35 fps is just 'inadequate' for playing.

 

When it comes to fps in games there is no magic number above which you will get the best possible experience. There are games which will run just fine at lower fps while others will feel worse even though the average fps is higher. Generally minimum fps matters much more than the averages but this is still different from game to game. Some game may randomly drop fps from time to time which can be a bit annoying, another may run so consistent at 28 fps it may even feel smooth just because it doesn't fall below 26 and doesn't go above 30, and some games you may get average 45 fps but it will be so terribly optimized the frame rate would just drop to 15 and jump to 90 and end up a stuttery mess. It's obvious that while none of the three cases above are optimal, the third case is the worst by far while the second is arguably the best. Also a locked 30 fps may look smooth in some cases and not very smooth in others. For example fast paced games like Racing sims will feel smoother the higher the fps is and competitive shooters will benefit too because you turn the camera so much while RPGs and most MMOs are generally fine at lower fps if you can get consistent framerate, not all of them though, Action MMOs like Vindictus still need high framerates.

 

http://www.forbes.com/sites/jasonev...-new-fiji-graphics-card-beats-nvidias-980-ti/
An article showing AMD released benchmarks on game performance with settings. Granted reviews are two days out, but decided to share.

Also, got banned for trolling in Anandtech's forum for supposedly trolling trying to explain how the increased bandwidth allows for faster in and out of data from the HBM frame buffer allowing 4GB of HBM to function as well as 6-8GB of GDDR5 because they put the same amount of data through in the same amount of time... Pretty hilarious being called a troll for dispelling misinformation...

 

Surely that would mean that flagship GPUs would require less memory than low end GPUs? Because they have the faster memory. But that's not the case.

What am I missing?

Sent from my Nexus 5 using Tapatalk

 

OK, a couple factors play into the necessary memory. Not all ram is created equal. You have to look at the effects of clock speeds and bus width to understand the speed, or bandwidth, of the ram to understand what effects it has on the frame buffer. Examining the two standards, GDDR5 has a 32-bit bus width and is clocked at up to 1750Mhz (7GBps). HBM has a 1024-bit bus width with a clock speed of 500Mhz (1GBps) on gen 1 HBM. What this means is GDDR5, per chip, can do a max of 28GBps, while HBM can achieve 100-128GBps per stack. That is much faster! Not only this, by shortening the distance and better cache implementation, latency is significantly reduced between the GPU and the ram.

So how does this translate to HBM being able to perform the same function using less ram? Well, we have to look at what the industry has done to this point. As cards where being developed, the processing power required more bandwidth between the ram and the GPU. So, how can you increase that bandwidth if you are stuck with 32-bit bus width? You run them in parallel and slap more chips on the card!!! That's the solution, but has the cost of heat. It sped things up while costing less than developing a better solution. Why bandwidth? BECAUSE THAT WAS THE BOTTLENECK!!! After that solution was implemented, coders got lazy because they had so much ram to play with, there was no need to optimize the frame buffer.

But that still doesn't explain why HBM can perform the same function on less ram... But it actually does. By increasing the bandwidth by about 52% over the Titan X (336.5GBps vs 512GBps) along with caching improvements and reduced latency and a couple other tweaks, the data is transferred in and out of the frame buffer in HBM FASTER than GDDR5!!! So even if filled to capacity, it is functionally as if you had MORE ram if it was on the slower standard of DDR5! This also has to do with improvements to the ROP (which has only 64, far fewer than Titan X and 980 Ti) and an additional 800 shader cores to process the data quicker. Does this mean that it doesn't have limitations? NO! At 4K the performance is right on par. When 5K and 8K are used, due to larger frame buffer sizes, the speed advantage breaks down and capacity, once again, is king. It still MAJORLY ON SYNTHETIC BENCHMARKS outperforms the 980Ti, but gets owned by the Titan X! Why? Because now the larger capacity becomes necessary because of the resolution requiring larger size frames. This is where the 12GB of GDDR5 is able to buffer better and make the card shine. The 4GB on each standard IS NOT EQUAL and 4GB OF HBM HAS THE CAPABILITY OF PUTTING THROUGH AS MUCH DATA AS 6-8GB OF GDDR5 IN THE SAME AMOUNT OF TIME ALLOWING FOR EQUIVALENT PERFORMANCE! Even this is an oversimplification of all factors that take place.

Please excuse part of the above explanation, copied from what I posted in Anandtech...

As to why previous jumps in standards didn't yield the same results, it didn't have the bandwidth increase to the same degree, latency usually went up, not down, etc. It is a combination of the increased processing with the HBM that gives the same results on 4K. It is possible that 4GB is limiting the potential of the card to a degree, which is why I also had said:

"Now, as for the reasons for only 4GB instead of 8GB cards. It is a limitation of current interposer tech. Currently, for Gen1, interposers only allow a stack of 1GB of memory to rest on it. To do more, you must use a dual interposer which adds significant costs to production. It is not a lack of memory available, it is that the costs don't justify the return for the initial release!!! As this will only be needed for competing with the Titan X at 5K and 8K gaming, it can wait for the time period of releasing a dual Fiji card this fall without sacrificing sales."

So it is a mix of things to allow the utilization of the faster bandwidth, keeping frames in the frame buffer shorter time (but where it is, overall, giving the same amount of data put through the ram in the same amount of time), better caching, lower latency with much more powerful processing that makes it equal. That doesn't mean that a little more ram would not give any boost, but the gains may be minimal (compare the 980 Ti with 6GB vs Titan X with 12,while factoring in fewer ROPs and slightly fewer shader cores (300-400) giving approximately the same performance). In addition, there is more to be said on driver optimization and AMDs standard 3+ month optimization window that they have greatly increasing performance. As time moves on and the new APIs are introduced, the performance will grow...

If I made a mistake in here, please correct me...

 

Thanks for that explanation.

So I take that it's due to reduced latency that HBM can do more with less? All GDDR5 modules have similar latencies so that's why the same doesn't apply there? Right?

Sent from my Nexus 5 using Tapatalk

 

Well, read/write operations will be done faster due to lower latency. Additionally, HBM simply moves data considerably faster.

 

This is an area I am still learning and I'm applying general principles... I think I stated the reasons for the equal performance correctly. Is there anything you would recommend changing in the explanation?

 

Do you mean bandwidth? Because the bandwidth gap between low end GDDR5 GPUs and high end GDDR5 GPUs is greater than the difference between the top GDDR5 GPUs and the AMD Fury X's HBM 1 implementation?

Sent from my Nexus 5 using Tapatalk

 

So my logic to get 4x4 gb ram is all in this too I knew i was not crazy in someways splitting data on smaller part with the highest bandwith is what Hbm is all about

not 10 chip with low bandwith

2-4 chip with very high bandwith and no latency

 

No latency because it's close to the core... Unless they figure out how to get the heat to not wreck from the core to the memory, we can't ever see this design on mobile.

 

The only limitation why it wont fit in MXM format is that the chip wont fit between the mounting holes, or it will be A REALLY tight fit (at 45º angle). Most likely wont fit. Take a look for yourself - http://www.bit-tech.net/hardware/graphics/2015/06/24/amd-radeon-r9-fury-x-review/2

EDIT: Just measured my Quadro 5000m's chip (the largest chip ever fitted to an MXM board (same as 480m), as far as I know) - it is 43x43mm and is quite literary over the clear zones of the mounting holes! If you rotate the chip 15~20º you could fit 45x45mm chip, but that's about it.

 

Are you talking about heatsink mounting holes or the video card mounting holes to hold it to mb/chassis?
Edit: reason I ask is you could just move the holes. It would not be the first time an mxm heatsink was separately designed dependent upon whether you bought amd or Nvidia....

 

The hole spacing is there by standard, the design of the heat-sink does vary from GPU to GPU, but the hole spacing is the same. Don't know how much retooling would be involved if they decide to redesign the MXM module. The connector should be the same, I see no reason to change it. OK, maybe they can drop the LVDS, since most of the new displays are eDP anyway and replace it with two more DPs. That would be sad for those with older machines (me). Time will tell how it would play-out.

 

MXM is a loose standard... So they'd need new X brackets. That's really not a big deal. My concern is moving the heat fast enough not to make the HBM artifact and eventually fail. The common heat pipe design is just not going to work, its going to need active cooling old school fan sitting on heatsink Alienware style.

 

I'm wondering how a modified P7xxZM would handle a fury nano... Same hbm, but made for a single fan... I think, with minor modification, you could handle the heat...

Edit: by modification I meant design mod to run the pipes directly over the vram and vrm on a card like that.

 

Those pipes can't move the heat fast enough...you have to remember HBM gets hit with the heat next to the core...you'd need a mod that puts the fan right on top of the heat pipes... I'm guessing we will see AMD release a Fury Nano with a 30% drop plus GDDR5 at this point... It's hard enough to release HBM but its even worse when you go to ODMs that already dislike you for a number of reasons and make them make a new cooling design.

I would love to be proven wrong but I'm tired of putting faith in AMD and so are the ODMs.

 

Welp here I go

Spoiler: I'll save you the long post on the page XD

No, it doesn't. I'll elaborate a bit further soon.

This line in bold is incorrect. The 336.5GB/s is the mathematical bandwidth of the Titan X. I.E. a Titan Black would have this. But as we know, Maxwell has memory bandwidth optimizations that even showed up in that cuda benchmark around the 970 vRAM fiasco. I'd put the Titan X's bandwidth closer to 400GB/s; maybe somewhere around 380GB/s at MINIMUM. Assuming 380GB/s to be the effective bandwidth, this means only a ~35% improvement is gained over the Titan X/980Ti (132GB/s) at minimum. This also means that the Titan X/980Ti are 75% as fast as HBM on the Fury X. It's a large jump, sure, but it isn't a massive, end-all one allowing vRAM size to be stretched indefinitely. If you were comparing it to the 4GB on the 980, with a ~257GB/s effective bandwidth (224GB/s mathematical one) versus the 512, I would have wholeheartedly said that it would have allowed that 4GB to stretch further. But against something with even higher vRAM limits? It isn't happening. Even if the cache improvements are great and the latency is reduced.

I think this has less of an impact than you imagine. This would fall flat under a card's ability to render what's on-screen, and does not automagically mean that more cores would equal better performance with this. As far as I can see, the workload processing ability of the card as a whole is what matters. AMD happens to have a massive number of cores to get a certain processing power. One maxwell core is stronger than one Fiji core at the same clockspeeds. If someone wants to counter my statement on this, please do so.

There is no way that 4GB on a Fury X can put through 6-8GB of data that a 980Ti would in the same amount of time, going by bandwidth. And there is something else that is an issue: not all data is cached and swapped out. For example, a friend of mine with a 970 was explaining to me that while he sees his vRAM hitting the limits often, it does not trigger the 970 problem in every game. This is in line with behaviour that some games work just as fine with cards that have more vRAM as with cards that have less, except for the fact that the cards that have more vRAM see much heavier vRAM utilization. It's as if they just toss everything into cache for no reason. These kinds of games will not be hindered by the Fury X's 4GB limit. But the other kind; games that follow in the vein of Wolfenstein: The New Order where it cannot compress data? Those will falter. What happens with W:TNO is that HardOCP tried to force "ultra" textures on a 2GB GPU when it requires a 3GB GPU or better to turn them on. The game ran at a near-standstill and crashed with out-of-memory errors. This could happen for games that truly need (still dev unoptimization, but it could happen) over 4GB of vRAM at 1440p and higher resolutions.

Actually, we didn't have bandwidth improvements for quite a while. The GTX 280 and GTX 285 had huge amounts of memory bandwidth at the time (160GB/s on the 285; it had a 512-bit memory bus with GDDR3, which is EQUAL to a 256-bit memory bus on GDDR5 at the same memory clocks) and didn't actually get much faster until the GTX Titan/Titan Black/780Ti (192GB/s on GTX 580 and 680, 178GB/s on AMD 6970 etc). AMD actually sort of led the charge with the 7970; it having 288GB/s bandwidth, but its power would probably never be known because it took so long for its potential to shine with drivers. It's only lately that there's been a real big surge for them, with both AMD and nVidia trying to crack the 300GB/s barrier by default with their top tier cards, as well as large amounts of vRAM being a thing.

This is the real issue. nVidia is slacking lately, but we still get multiple drivers a month in most cases. AMD does not. Not even close, as far as I've seen. They need to polish things up here. I'm certain the Fury X can beat things. I've seen multiple benchmarks from multiple people using outdated drivers for nVidia, and when PCGamer produced their benchmarks with the absolute latest drivers (353.30 and whatever AMD's own is; it was a beta) the nVidia cards got significantly better scores in their game testing; something they even made note of in the tests as being a result of the recent drivers (with regards to the Titan X and 980Ti especially). So most synthetic benchmarks run by various websites including Tom's Hardware may need to be taken with a HUGE grain of salt. I decided I'm not accepting any benchmarks at all with drivers that are older (and perform worse than any existing newer driver) for nVidia cards, because then the results are skewed to AMD regardless of what the test produces, as the nVidia cards are not running as well as they could be at the time of the testing. And the same goes for the AMD cards, mind you. I want each card to be tested at the best they can be whenever the tests are done. But yes, 3 months down the line is going to be the best time.

 

And new heat-sinks with those brackets As for the pipes, they transfer the heat pretty well, not to say better than a fan blowing on top (take a look at all of the top-tier desktop air coolers). What is lacking is the "meat", so to say at the end of those pipes. You'll need a hell of a radiator (and fan) to dissipate the heat as fast as possible.

 

My M17xR1 worked around that problem by sitting a fan underneath the card... It didn't have pipes running everywhere. It didn't need to. Those fans did their job. The way most designs (I'm looking at you Clevo) work now is the rely on pipes to take the heat to the radiator and a poorly insulated fan to take over to cool the radiator. This is horribly inefficient.

 

I ust noticed the M395X @ notebookcheck and was amazed to see 2048 gcn shader cores in an amd mobile chip. When'd they launch this?

 

LOL sangemaru. More than 2 years ago haha! M395X is a rebrand of the M295X with a minor bump in clocks.

Hotter (or as hot) as the 980M with performance in the 970M ballpark but not as OCable due to the heat. Had it arrived alongside the 780M things might have been different...

 

Actually, it's just a straight rebrand of the R9 m295x. Notebookcheck's R9 m395x is the Apple m395x, which has higher clock speeds.

 

I assume this design of heatsink may be popular if the HBM is getting too hot: