UPDATE (5/22/15): Johan Andersson tweeted out this photo this morning, with the line: "This new island is one seriously impressive and sweet GPU. wow & thanks @AMDRadeon ! They will be put to good use :)" Looks like we can confirm that at least one of the parts AMD is releasing does have the design of the images we showed you before, though the water cooling implementation is missing or altered.
END UPDATE
File this under "rumor" for sure, but a cool one none the less…
After yesterday's official tidbit of information surrounding AMD's upcoming flagship graphics card for enthusiasts and its use of HBM (high bandwidth memory), it appears we have another leak on our hands. The guys over at Chiphell have apparently acquired some stock footage of the new Fiji flagship card (whether or not it will be called the 390X has yet to be seen) and it looks…awesome.
In that post from yesterday I noted that with an HBM design AMD could in theory build an add-in card that is of a different form factor than anything we have previously seen for a high end part. Based on the image above, if this turns out to be the high end Fiji offering, it appears the PCB will indeed be quite small as it no longer requires memory surrounding the GPU itself. You can also see that it will in fact be water cooled though it looks like it has barb inlets rather than a pre-attached cooler in this image.
The second leaked image shows display outputs consisting of three full-size DisplayPort connections and a single HDMI port.
All of this could be faked of course, but if it is, the joker did a damn good job of compiling all the information into one design. If it's real, I think AMD might finally have a match for the look and styling of the high-end GeForce offerings.
What do you think: real or fake? Cool or meh? Let us know!
actually nice design, clean
actually nice design, clean sleek..
Inconspicuous is nice when
Inconspicuous is nice when done right but this looks like it was thrown together without much thought. Incredibly unattractive style imo.
Surely their powerpoint design team could’ve done better.
If the top end cards are
If the top end cards are water cooled then what happens (rare, I presume) when someone wants to put more than two cards in their machine .. even two might be quite bothersome?
ideally you’d have them on
ideally you’d have them on the same loop, just like water cooling sli/crossfire as it is today.
Hopefully this marks a change
Hopefully this marks a change and we move away from “Gamery” looking enthusiast parts from AMD.
Do i need a personal Hoover
Do i need a personal Hoover Dam to power it?
I hope it does not require
I hope it does not require water cooling.
why not?
why not?
Looks about the same size as
Looks about the same size as a lot of the ‘short’ 970s (which were dropped onto 760 PCBs).
This is NOT the HBM card, if
This is NOT the HBM card, if you follow the chiphell link they show the card with the cooler of and you clearly see traditional memory chip layout.
This is one of the rebranded older card fit into a very small form factor and given its only one 8 pin connector probably not one of the highend cards.
I saw that as well but I was
I saw that as well but I was wondering, the main complaint people have about the proposed spec is the lack of memory. Is there any technical reason they couldn’t use traditional memory as something like a texture cache?
That would be pretty huge. Solves the one issue people seem to have.
That would be a much more
That would be a much more complex memory controller than what AMD has hinted at with HBM. It is a good idea, but really tough to implement.
Yes and they rejected it for
Yes and they rejected it for now.
“- Could we see a mix of HBM and GDDR5, sort of like how a SSD and HDD would work?
– Mixed memory subsystems are to become a reality, but nothing yet, main goal is graphics cards.”
Having any GDDR5 would defeat
Having any GDDR5 would defeat a major advantage of having HBM, which is offering high bandwidth at low clock rates, plus having a much simpler PCB design and not having to run the memory traces out of the interposer module. GDDR5 at high clock rates is a power eater. The extra latency from having to go off module for the added memory space is just not worth it, as proper staging and compression of textures from main memory into the HBM will probably be enough until HBM2’s introduction. There is no beating 4 channels at 1024 bits wide per channel for HBM, and once the larger HMB stacks start arriving, texture compression will have to make do. With all that bandwidth a lot more extra processing can be done per frame and not have any adverse latency consequences that cause frame rates to suffer, that the older memory technology has to deal with. Expect HBM and the GPUs memory controller to be able to handle many more CPU requests for all processor cores/threads pushing data over more of the 16x PCIe’s available bandwidth, as HBM’s more lanes at higher available bandwidth should allow for better peak PCIe transfers. I hope that there is some benchmarks that can properly measure the PCIe bandwidth usage for these first HBM enabled SKUs, as I would love to see how much those 16x connections are going to be utilized compared to the non HBM cards.
If you are using the extra
If you are using the extra off package, PCB mounted memory as overflow, then it doesn’t actually need to be that fast since you would not use it often. The heirarchy goes from small and fast out to big and slow. If they did run some external memory connections (unlikely, and probably unnecessary for a GPU), then it would probably be on a much narrower, slower, and lower power interface compared to the interface you would see on a non-HBM gpu. If we do see off-package memory with HBM equiped APUs and server chips, I wouldn’t expect it to be connected on a 512-bit interface.
They will want to implement
They will want to implement secondary memory for APUs or server CPUs such that the HBM would act like a cache. I am not sure how they would implement this on a gpu though. Setting it up like a cache may run into some complexities. I suppose they could have it on a completely separate memory controller off the gpu and have the driver manage it rather than a transparent CPU-style cache.
Anyway, capacity issues, if they really are an issue, will disappear with HBM 2. If HBM 2 takes too long, it may be worthwhile to add more HBM stacks to GPUs rather than attempting to add secondary memory. We will see controllers with secondary memory with HBM 2, if it is used for server chips and APUs.
It will be interesting to see how that will be implemented. It may be possible to put the secondary memory controller(s) off the stacked DRAM logic die rather than off the main chip which would have advantages. Given the sizes involved, the caching mechanism may need to look more like paging system memory out to disk rather than a traditional CPU type cache. If that is the case, then you would probably want to implement it as part of the virtual memory system, which would not be simple. You would need to add another level with some pages in HBM, some in slower memory on the PCB, and possibly some out on a storage device. Although given the sizes involved, going all the way out to disk may no longer be necessary if you have huge amounts of RAM. If we have memory modules with stacked memory chips, then the capacity could be huge. You probably could have a tiny but powerful system with one APU with HBM, one so-dimm sized memory module, and one m.2 ssd.
Men this is what I´ve been
Men this is what I´ve been wating for… small, just one PCIe port when water cooled with a water block and powerful, if all of this is true, definitely my next GPU with a freesync monitor.
I wonder if they are gonna do
I wonder if they are gonna do something silly, rebrand all the old cards to the rx-3xx and name the flagship “Hercules” or “Perseus”, like “A TITAN KILLER!!!!”
Actually that type of
Actually that type of marketing is done in all kinds of markets, new features are introduced at the high end, and then trickle down to the lower tiers as time goes forward and newer models are released, and those models that were once highend are rebranded with newer model numbers as lower tier products, happens all the time.
Can’t really say anything bad about AMD doing it, especially since this new memory tech is brand new and the manufacturing of it is not mature yet, so yields are lower and cost is higher.
well yea, and using a lower
well yea, and using a lower number is fine (still dont like it, understand it) but is seems the r9-270 is going to be renamed r9-370, and it;s still an 7xxx part (7870 i think, dont quote me)
So if it was the new 250 or 260 I would be ok, this is just a need for new/rebrand
how come no one get this the
how come no one get this the newer cards are morelikely to have the same chip thats in the r9 285 mean great performance with less power
Correct me if I’m wrong, but
Correct me if I’m wrong, but isn’t the r7-230 from the 5xxx gen?
Close. The R5-230 is a 6450
Close. The R5-230 is a 6450
Actually I think that would
Actually I think that would be great, hehe…
Zeus? Olympian? A reader
Zeus? Olympian? A reader mentioned "Uranus"…
We need to get over that old
We need to get over that old joke and rename that planet “Urectum”
Maybe they could call it
Maybe they could call it “Rambo”
I love the chrome finish.
I love the chrome finish.
The tension…the MOZAFAKKING
The tension…the MOZAFAKKING TENSION…IT’S KILLING MEEEEH!!!11 TWO WEEKS!!! TWOOO WEEKSSSSSSSS!!!11
I really want to see the
I really want to see the board itself. It should be super simple with just the interposer package and the power delivery system. The traces should also be super simple: just PCI-e link and display port, both of which are relatively low pin count. You also have the power and ground from the VRMs, but with no memory interface, it should still be a super cheap board. The interposer isn’t cheap though.
I have a feeling, I have
I have a feeling, I have nothing to base this on, but I have a feeling that the board will be INSAINLY simple. Just a guess tho, but I have a AM1 system and the MB has nothing too it, I have a feeling that this card, IF it’s real, IF it looks like this, I think it’s gona be a card of connectors and not much else.
just release the cards
just release the cards already AMD!!!
deadlines only lead to beta’s
deadlines only lead to beta’s becoming products! idiot