AMD Might Replace RX 500 Cards with RX Vega 32 and 28

Card meant for 1080p gaming? 4GB VRAM? No prolem.
There has been only one GPU which did run out of VRAM before horsepower, that's GK104 paired w/ 2GB of VRAM.
Yes, one can use 4k texture pack on 1080p resolution. Placebo works. But it is not question of HW configuration or calling VRAM amount future proof. It is about sanity.
On 1080p, 4k textures improve clarity maybe by 5%. If it chokes memory controller or cause other type of stutter, then it is matter of choice. Secondly those 4k textures have to be processed, so even if there is enough VRAM and memory bandwidth, there is considerable performance impact on average gaming GPU.

 

This would be a very weird update if it will turn out to be true.

 

I'm not sure I fully understand why AMD would do this. If they've got a much better/bigger plan for the RX 600 series then fine, replacing the current 500 series with Vegas might make sense. After all, the whole lineup was a bit boring compared to last year (that being said, I feel that the current 500 series is what the 400 series should've been). But if they don't have anything specifically lined up for the 670 and lower, then shouldn't these be used for that?

If they are in fact releasing more 500 series, I hope they do something to distinguish the differences between the GPUs (unless the performance just happens to be almost exactly the same). This could get real confusing for consumers, and reviews. Maybe these would use the X suffix they like to over-use?

As long as miners are making it difficult to buy GPUs, I find whatever they do to be irrelevant anyway. And perhaps they're releasing this "refresh" for this very reason: they've got all this Vega hardware that isn't all desirable, there was nothing all that remarkable about the RX 500 series, and they don't want Vega tarnishing the next generation. Not enough people own 500s, so thanks to the miners, it's almost as though AMD has a 2nd chance to release the 500 series.

 

Didn't know what resolution you were running, makes sense you're looking for an upgrade.
MSRP = manufacturer's suggested retail price. MSRP is set by AMD, not retailers. At the moment the price is being driven up mainly by low availability and mining craze, not AMD's fault. 1080Ti is the best Nvidia offers and no one can touch it so you're paying a premium for it. That's why I said 1070 on discount or 1070Ti is a good deal for you.
AMD is probably using the same memory controller on Vega 56/64 so it has to have HBM.

My RX560 has 4Gb and maxing out games it doesn't use more than 2.5Gb. Yes, I play at 1080p but if you can afford 2/4k displays, you can afford a 1080Ti too.

 

It is your opinion guys against all devs suggesting 6 gb vram or more for some games playing from 1080p to 2k resolutions or even more. I never said you can't play smoothly, even the benchmarks for these games show otherwise contracting devs suggestions. But I understand that nowadays mainly because of AMD's strategic weakness (Fury series, now furys 2.0 aka vega 11) we have more GPUs with up to 4gb even on the mid/high tier than we ever really needed. If more Vram was useless at first place, games wouldn't utilize it when available, would they? Bad programming? I don't think so. It is like keep saying 8gb ram is enough for gaming but for sure years now it isn't. You need at least 12 to 16gb ram for pure stability. You risk everything from using 3-4 to 6-8gb ram. Electronics technology is advancing with miraculous step-ups.

And no, there are 2/4k displays with half the cost of 1080ti or even less, that was irrelevant tbh. Not everyone needs a top monitor at the same cost of any gpu or even more expensive than the card itself. But everybody is free to make decisions. You love a product, but. Simple as that.

 

Yeah you can get a 1440P monitor for 200 USD

 

Ok at stock vega 56 consumes less (30 watt )than the 580 so that makes me hopeful about consuption, they have to keep the same msrp i believe , guys saying that 4gb are not enough is a bit exesive , news flash gaming is not only max detail or nothing with minimal loss on detail you can squeeze a lop of perfomance out , so many people take review numbers as the absolute a gpu can give in fps and thats it! Well it's a review looking to press the card to the limit does not mean you have to play on these settings . Now about those cards no idea they might surprises or disapointments again

 

I don't know where do you get your info, but 8GB is enough (if you don't have 2 firewalls, 3 different AV apps and bunch other bloatware loaded. But sure, maybe you need all that background crap while you game so 8GB might not be enough for YOU), and 3GB VRAM is enough for 1080p/1200p gaming. Even 1440p in a lot of cases.
Truth is, some games would allocate 100GB of VRAM if you had it. Engine is made in a way that it would try to fill up as much VRAM with assets to shorten loading/swapping times. More often then not, those pre-alloceted assets will not match the needed assets in a very short time, sometimes even after 1 second it will require the engine to dump the previous assets and allocate the new ones. With more VRAM than needed, it will just spend more time filling that much more VRAM. Yea, I would call that BAD codding just because the idea that led to that code is bad.

edit: the only time when you truly need more VRAM is when you enable some setting and notice a sudden stutter, full stops or really high degradation in performance while your "HDD" is going crazy. Obviously you need 4+ GB for a 1440p (to be on a safe side) and up. Mainly because of larger textures and more pixels to render.

 

Vega 56 doesn't consume less than 580, although they are pretty close.

https://www.pcper.com/files/review/2017-04-17/power-rise.png
https://www.pcper.com/files/review/2017-08-13/power-rise.png

 

My thoughts exactly.

Recently (few months back) War Thunder engine was coded to do this: Flying would use 700/800mb and now uses about 1.5Gb because it keeps the old files instead of clearing memory between matches. The idea is: loading is faster if the match has same kind of units.

If you're playing on a 4k monitor with 4k textures and everything maxed out I'd say 8Gb are good enough.

 

IIRC AMD's actual MSRP was only $90 less (here), which still puts it in full retard territory. As Agent-A01 said, 1440p monitors can be found at $200. I don't remember what I paid for my monitor but it was possibly less than half of what a 1080 Ti costs.

And affording an item isn't everything, it feels so stupid to pay that much for something that would cost a fraction of that if there was real competition in the industry. The 1070 is not an option at all, that's one of the cards heavily affected by miners, it's in the $600 range. If you're going to spend $600 you might as well spend a little more for a 1080. It would feel dumb paying $400 for the meh level of performance of a 1070, let alone $600 + tax.

At least the price of a 1080 would be easier to swallow if it wasn't absolutely ancient history by computer standards.

 

A single stack of HBM2 can be up to 8GB. I doubt AMD will ONLY release 4GB variants.

HBCC should do a good job making up for those few scenarios where you need > 4GB anyway. X-Plane 11 uses the most of anything I use and I bet a good 30% of its asset in VRAM aren't called on very often.

 

I can not say i looked a lot at it i remember hardware unboxed on his review found out that v56 was consuming less on tottal system consuption so after that i was with that impresion, the numbers from your link also indicating a 2/3 vega 56 should be lower than 580 on consumption , except if something goes wrong ... again !

 

Interesting..... I am concerned about the price with HBM2 RAM, as well as whether it's enough for 1080p maxed out gaming or not. I'll wait for some more definite news etc... but at this point I'd also be concerned about whether or not they can make enough to stock the shelves considering the lack of Vega 56 or 64 on the market.

 

I can see the Vega 32 and 28. What I can't see is only 4GB HBM @ 1024Bit. That will be less shaders than the RX580 and less memory bandwith than the 8GB RX580. There is no way in hell it will out perform a RX580 with those specs.

 

To have tried GTX960 2g vs GTX750ti 4g, because a friend was thinking that this "upgrade" was not too much an "upgrade", i can say that now 4g is the limit to have for play.
With most actual not too demanding game the GTX750ti do the same as GTX960 due to memory lack despite being more powerfull and having vram compression.

About HBM and HBM2 they are really fast vram don't blame the ram but more the GPU. On green (pro) side it work pretty fine (but damn expensive).

I agree that with HBM2 you need less vram because it is very fast but when texture pack is already around 2g in actual game, you already need the next step...

 

My old Fury Nitro held its own ground decently well in WQHD.

That said, the Fury series had higher bandwidth than Vega does. In how far Vega's improved (?) compression helps, I do not know.

Generally speaking however, never once have I ran into a VRAM bottleneck using the Fury Nitro for a year in WQHD gaming - and we're talking '16 to '17.

Those cards will more than likely be fine for where they'll wind up at.

Price : Perf-wise, well, that's left to be seen.

 

Now it makes sense that NVIDIA launches the GTX 1060 TI...

 

illustrative, not intended as rude to poster, blah blah blah:

"but 8GB is enough (if you don't have 2 firewalls, 3 different AV apps and bunch other bloatware loaded. But sure, maybe you need all that background crap while you game so 8GB might not be enough for YOU), and 3GB VRAM is enough for 1080p/1200p gaming. Even 1440p in a lot of cases.
Truth is, some games would allocate 100GB of VRAM if you had it. Engine is made in a way that it would try to fill up as much VRAM with assets to shorten loading/swapping times. More often then not, those pre-alloceted assets will not match the needed assets in a very short time, sometimes even after 1 second it will require the engine to dump the previous assets and allocate the new ones. "

Once the limited 200GB/s+ gpu ram is exceeded - then what?

What must it read & (god forbid) write to, to refresh its workspace?

It can read any presciently preloaded game material from spare system memory at approaching 16GB/s, but spare ram is limited and games & some apps are large.

It can read a top nvme ssd at ~3.2GB/s,

or sata sd at ~500MB/s (slower sata is ~no cheaper, it uses ~same nand as way better nvme bus ssdS)

it can read from HDD at ~100MB/s (incredibly, gamers do it seems - not all can afford large enough ssdS)

There is a huge chasm between vram and the next fast and spacious option once vram is exceeded.

So even if we can afford todays thirst for ram (we feel much poorer this upgrade than last), we move the problem, not remove it.

Vega can narrow the above chasm a lot.

HBCC allows a nvme raid array to be used as extended gpu cache.

It uses the sub optimal pcie bus, but maxing out the available 16 lanes is simply a matter of the cluster size of the array.

I speculate an affordble 3x:

"Samsung SM961 Polaris 256GB M.2-2280 PCI-e 3.0 x 4 NVMe Solid State Drive SSD"
Sequential Read : 3,200 MB/s

Sequential Write : 1,800 MB/s

Random Read : 450K IOPS

Random Write : 400K IOPS

http://www.ebay.com.au/itm/Samsung-...547148?hash=item46657f38cc:g:9y0AAOSw8HBZLt79

or maybe even:

Samsung SM961 128GB OEM MZVPW128HEGL M.2 2280 128G PCIe PCI-Express 3.0 SSD

http://www.ebay.com.au/itm/Samsung-...763720?hash=item211c7b4e08:g:f1wAAOSwJ7RYUQNh

Sequential Read : 3,200 MB/s

Sequential Write : 1,800 MB/s

Random Read : 450K IOPS

Random Write : 400K IOPS

with:

Most TR mobos have 3x classy nvme m.2 format ports, & TR will within days, will support raid in bios for nvme free - no fancy controller costs (find another way to boot imo - leave as a non boot swap/temp file/cache drive).


Dearer 500GB samsung 960 pro raid benchmarks (yes, there are faster more expensive options) i have seen, show striped raids are pretty seamless up to a sweet spot of 3 drives, which conveniently come free on most TRs, as above..

adding a fourth drive added substantially to the read speed of the array, but way less than previous drives added. Inconclusive. Just saying, 3x seems a cost effective sweetspot

Here is a curious thing tho. It seems to me, using "storage" to simulate "ram" in this manner, requires a lot of writing also, and fast read is no good if its tied up with, or waiting on, a write.

As u see above, write on nvme lags reads by a big factor - 1800 vs 3200MB/s

HOWEVER, write speeds improve disproportionately well from raid. Raid narrows the gap a lot - almost to parity. This seems a hidden and desireable balancing benefit. As above, whilst adding the 4th drive was least cost effective for read speeds, it added as much write speed to the array as the first drive added.

For simplicity lets pretend these indicative claimed bandwidths are correct, and later discount them by whats fairish.

If we get lucky and using such small 128GB drives is ok & using the oem version isnt a problem, THEN ~$110 ea., thats 384GB of theoretically 9.6GB/s sequential read and 5.4GBs write for $330US.

My gut feeling is 20% of read speed is lost to raid overheads, and maybe as little as 5% to write speeds, or simply 8GB/s & 5GB/s

Yes samsung''s claims are for ideal reads and write, but i figure if hbcc does its job right, then the big slabs of page file being swapped, will be ~ideal reads and writes. HBCC intelligently manages a pool of resources (ram,vram,storage including raid nvram) with varying properties, and will not expect the array to perform tasks more suited to DRAM component of the pool.

Tho this small raid array may not saturate the 16x pcie gpu lanes alone, hbcc(ontroller) may use the remaining bandwidth to serve the gpu concurrently from the system ram component of the managed cache resource pool.

Raw bandwidth of any component cache resource cannot be considered in isolation. by teaming, pattern recognition, pre-empting and even learning, the "whole" of hbcc is far greater than its component parts.

I am dying to know user options within vega hbcc setup -any links anybody?

But wait, there are steak knives too.

We also see, in production by amd, a futuristic vega gpu card with its own nvme raid - the radeon vega ssg. Its as above using vega hbcc, but it side steps the pcie bus altogether & interconnects direct to the gpu, with minimal latency and demands on pcie bus bandwidth or cpu.

I cant believe the hostility to this hbc concept. There is no reason it cant work in theory (it does in practice now in servers, big time), and its essential that it DOES work.

The ram shortage is insoluble for some time. We can dream about our ideal system, but we can no longer afford them. Only AMD offers some relief for big memory, or getting by with less dram memory.

fyi, i heard 3d nand making can be done on recycled old fab gear. Simple manufacture and the rapid paced advances of adding more layers, make 3d nand seem relatively supply unconstrained near term.

From another perspective, as i recall an answer of Raja's? re hbcc, in a context of having sufficient PC system memory usable by the hbcc pool, it could simulate 2-3x more gpu memory than actually installed, w/o it being discernable to the user.

Maybe nvme raid can help out a bit w/o being too discernable also?