Hopefully you have already read up on my review of the new Radeon Vega Frontier Edition graphics card; it is full of interesting information about the gaming and professional application performance.
But I thought it would be interesting to share the bare card and GPU in its own post, just to help people find it later on.
For measurements, here's what we were able to gleam with the calipers.
(Editor's Update: we have updated the die measurements after doing a remeasure. I think my first was a bit loose as I didn't want to impact the GPU directly.)
- Die size: 25.90mm x 19.80mm (GPU only, not including memory stacks)
- Area: 512.82mm2
- Package size: 47.3mm x 47.3mm
- Area: 2,237mm2
Enjoy the sexy!
Click to Enlarge
Click to Enlarge
Click to Enlarge
Click to Enlarge
Click to Enlarge
Click to Enlarge
Interesting notes:
- There is a LOT of empty PCB space on the Vega FE card. This is likely indicative of added area needed for a large heatsink and fan to cool 300-375 watt TDP without throttling.
- Benefits of the smaller HBM-based package appears to be at a cost of SMT components on the GPU substrate and the PCB
- The die size of Vega is large – bigger than GP102 even, despite running at a much lower performance level. It will be interesting to see how AMD answers the question of why the die has expanded as much as it did.
Feel free to leave us some comments if anything stands out!
Just to be clear, the 564mm^2
Just to be clear, the 564mm^2 figure doesn’t include the hbm stacks, right? (1.35:1 looks much closer to the die itself than including the hbm stacks, but i have to be sure)
Yes, sorry, NOT including the
Yes, sorry, NOT including the memory stacks.
Can you double check the GPU
Can you double check the GPU die-size? I believe AMD claimed much smaller.
As i thought, a load of empty
As i thought, a load of empty pcb space with the length being there to support the cooler. Just sorta flies in the face of the spiel for hbm, part of which was smaller cards. Probably could have made this furyx sized quite easily.
That would be for mezzanine
That would be for mezzanine style passively cooled cards.
Please don’t tell me you
Please don’t tell me you measured the interposer. How can two “tech journalists” attempt to measure the die size and get different results. Either you dun goofed or GamersNexus and since they did it live on stream…
AMD says 484 mm which would
AMD says 484 mm which would make Ryan’s numbers incorrect.
Gamer nexus measured a bit
Gamer nexus measured a bit different numbers….
30mm x 30mm total size of interposer + GPU (does not include substrate)
20.25mm x ~26mm GPU die size
10mm x ~12mm HBM2 size (x2)
~4mm package height (this is the one that’s least accurate, but gives a pretty good ballpark)
64mm x 64mm mounting hole spacing (square, center-to-center)
PCB ~1mm
http://www.gamersnexus.net/news-pc/2972-amd-vega-frontier-edition-tear-down-die-size-and-more
Gamers nexus need to learn
Gamers nexus need to learn how to use those tools. He is like 14 year old kid…
Nah, calm down. Gamer’s nexus
Nah, calm down. Gamer’s nexus is a great site and they do great work. It’s a little ironic since judging someone by how they look is childish in itself.
strangely PC perspective have
strangely PC perspective have updated the number recetly :-))
https://twitter.com/ryanshrout/status/881317151277285376
The thing about GN’s results,
The thing about GN’s results, he did not just write down or just tell you the dimensions. He took a measuring tool and let the viewer see the measurement. Some advantages and disadvantages of that.
The thing about GN’s results,
The thing about GN’s results, he did not just write down or just tell you the dimensions. He took a measuring tool and let the viewer see the measurement. Some advantages and disadvantages of that.
Just so we are clear, as we
Just so we are clear, as we have a lot of space on the card, do you think that RX vega gaming edition will have all of this filled or not ?
With GPU and memory fully
With GPU and memory fully integrated on the interposer & substrate, and most of its surroundings taken up by the VRMs, there’s nothing left to “fill” the rest of the board with, unless you want some extra-sick under-the-fan lighting effects. 😉
I suspect that most of the
I suspect that most of the custom designs will be all over the place. Some will be short, some will be just as long with 3 fans. I believe AMD was limited by the blower styled cooler.
what on earth did they put in
what on earth did they put in there to get it that big. some stupid math:
480 is 232mm^2 thats 9.9 shaders per mm^2 (yeah gpu not all shaders) blah blah.
vega at that rate would have 5600 shaders, they have almost 1600 shaders worth of extra stuff. or they dropped the density of their GPUs to get higher clocks.
yes dumb math.
compared to nvidias GPUs its seriously out of whack.
compared to polaris again its more than double the size for less than double the shaders.
should be interesting when all details about it are out. a die shot would be nice. final rx vega performance as well. whatever they put in there better have been worth it.
I have a sneaking suspicion
I have a sneaking suspicion that this super high end HBCC they put in there ballooned the die with no discernable benefit for gaming applications. There is probably a reason that no other gpu in the world has basically a cpu style memory controller. It’s not worth the die space trade off for gaming when you have enough vram sitting close to the chip.
Well for some non gaming
Well for some non gaming forms of graphics processing that HBCC/HBM2 appears to be working just fine, and it’s not like any current games are fully optimized for Vega. There will be developers getting Vega FE and Tweaking Games to take advantage of all the new IP in Vega. And who knows what extra shaders/ACE/NCU are on this die to improve Yields.
It’s amazing the sense of entitlement that many gamers have towards AMD and Nvidia, and Nvidia makes billions off of its Pro GPU business unit. I guess that once AMD gets a larger share of that Professional market then AMD/RTG will have the funds to produce a gaming specific GPU SKU with lots of extra ROPs to up the frame rate and sell GPUs on gaming metrics alone.
AMD will more than likely make much more off of its CPU business and semi-custom operation and AMD needs to get more of that high margin Professional GPU market share to have some GPU product to offer alongside its Epyc line of Server/Workstation/HPC offerings.
RX Vega is what you are looking for and that’s another 4 weeks off. SIGGRAPH will see a real Radeon Pro WX version of Vega with the real certified Drivers for professional graphics workloads and AMD will be releasing RX Vega for gaming so who knows what that RX Vega Die may look like.
All of these negative reactions after Raja repeatedly stated that this SKUs is not for gaming.
Nvidia has all those billions to focus on GPUs/GPU driver development while AMD has its currently limited resources to manage both a new x86 Zen micro-arch and also a new GPU Vega Technology for more than just gaming workloads. AMD’s focus has been on the Professional Markets with its Zen micro-arch and that modular Zeppelin die that is used across all of AMD’s current line of Zen based products. I think that AMD is doing the best with its limited resources and that focus is on the professional markets where the margins are higher.
AMD needs to get its overall gross margins above 40% to begin to turn a profit or maybe for just breaking even while having the revenues to increase resources for things like driver developers and other development expenses.
“so who knows what that RX
“so who knows what that RX Vega Die may look like”
Everyone with working eyesight reading this post. It’s literally / up there.
I think he means a block
I think he means a block diagram…
So you have proof positive
So you have proof positive that the Radeon Pro FE die is the very same die that AMD will be using for RX Vega. It may be so, but no one knows that until the gaming RX Vega SKUs arrive. This early Radeon Pro FE SKU could be/is using an earlier variant/stepping of the Vega micro-arch and AMD is selling that production for lower clocked Semi-professional usage where overclocking/gaming performance in not usually needed.
I’d expect that AMD may have a newer Vega stepping from an improved 14nm die process node fabrication run from GF with better electrical/leakage characteristics and maybe higher clocks/Less thermals than this Vega FE SKU.
Raja Explicitly stated that Vega FE is not for gaming so maybe he knew something about this SKU’s stepping and die fabrication run a little in advance. I think that the Radeon Vega Pro FE branded SKUs will be great for their intended developer base to use for graphics software, as well as games, development(targeting the RX Vega SKUs). So the new Vega Micro-Architecture and its new IP(Primitive Shaders, NCUs, Etc.) and Cache/Cache subsystems can have a software ecosystem that is more release ready for the gaming focused RX Vega SKUs, in addition to any professional graphics usage.
It may very well be the same size die, then it may not be when RX Vega gets here, but I bet that the Vega FE SKUs are using an earlier stepping/run that is a good enough bin for its intended usage.
Are you seriously trying to
Are you seriously trying to sell me on a sane company launching two significantly different silicon implementations of the same variant of the same mirco-architecture within a month?
Cause that ain’t flyin’ with me.
Who said significantly, I
Who said significantly, I said stepping and this process takes place a good while before final product release. So AMD was getting their non updated and updated silicon some time back(More than one month ago for sure) and making their tests and looking at the testing results and making their finalized choices months ago. AMD already had sufficient wafers/dies from probably more than one revision to make some binning choices and any lower binned parts that could not be stably clocked the highest were set aside to build up stock to create Radeon Pro FE SKUs while maybe the latest GF wafer/dies runs(decided on months ago) where all checked with those top bins reserved for RX Vega.
GF should have been producing Vega wafers for at least 6 or more months before AMD collected enough samples to create its possible final binning metrics, and those final binned metrics where decided upon more than 3 months ago, if not much longer. So for simple steppings that’s always an ongoing process, as are GF’s fabrication process node tweaks to get better leakage/electrical results. The entire binning process makes use of testing done on many wafer runs before there is enough data to run the results and figure out what the bottom to top testing results range becomes. And that all involves statistical analysis software where all those wafer/dies runs of working dies’ performance metrics are all fed into the statistical analysis software, specialized software that is used to create the binning regimens for GPUs, as well as CPUs and other processors.
The Vega FE product you are looking at was probably from a run produced some months ago, up to the point where AMD felt comfortable with the results and that process will continue until the last of the Vega production is replaced with Navi production and it all continues from there with the newest designs.
AMD will be producing some bins of the large Vega die with lower shader counts, and then AMD will have already been working on the more affordable mainstream Vega tape-outs(already decided) that will eventually replace the Polaris RX 570/580 mainstream SKUs. None of this happens in one month the entire process is staggered and the final Release To Market(RTM) product came off the fab lines months ago to be binned and made into several different classes of final product.
Any new Vega Dies(Less than a month old) coming off the Line after the final binning metrics where decided may even have some dies with performance figures that far outperform the top binned metrics used for the first RTM SKUs and those may be held in reserve for product updates/products refreshes in the months after RTM.
Vega has 49bit virtual
Vega has 49bit virtual addressing and 48bit physical for huge HSA installations. HBCC would be very useful for managing near memory and far memory pools in that case.
Since this GPU goes into their MI25 HPC accelerator, it will probably get a lot of use.
Also, i dont know if Vega Nova will have 8 or 16GB RAM. Might be useful there too.
Does the heatsink incorporate
Does the heatsink incorporate a vapor chamber? That kinda looks like one, though could just be a copper slab as well.
That’s a vapor chamber.
That’s a vapor chamber.
Hmm what is that unsoldered
Hmm what is that unsoldered connector thingy on the backside? Placeholder for ssd maybe or, what? Anyone have any bright idea/knowledge what is that.
There are no mounting holes
There are no mounting holes for an M.2. My guess it’s a debug port that they would solder on.
So what is the IQ of this
So what is the IQ of this beast?
Something is missing, where
Something is missing, where is vrm cooling(did you take it off, and did not take a picture of it or?)?
300 watts on those and no heatsink on them, what kind of power fets are they? Well it has 12 -phases so maybe they can stay cool enough.
The VRMs are cooled by the
The VRMs are cooled by the fan and the black full-cover base plate seen in the fifth picture. (that picture is a bit confusing because it has the vapour chamber and fin stack turned upside-down)
Yeah, indeed. Were there any
Yeah, indeed. Were there any problems with vrm temps anyway, that vrm design looks very solid.
There is still a little bit
There is still a little bit of thermal paste on the VRM components, so they must mate with the copper base plate.