Monday, July 3rd 2017

AMD Raja Koduri Confirms RX Vega Die Size at 484 mm²

AMD's Raja Koduri, leader of the company's Radeon Technologies Group, has somewhat informally confirmed on Twitter the overall die size of AMD's Vega chips. After PC Perspective updated their prognosis regarding Vega's die-size to a beefier 512 mm², Twitter users plied Raja Koduri with questions regarding this subject. Koduri declined to answer directly, actually opting for a somewhat cryptic response, in that " (...) the answer [to Vega's die-size] is the closest perfect square number actually:)".

For the math-savvy around here (or even just for those of you who have read the headline), that particular equation should solve towards a perfect 484 mm² die area. Good news for AMD: this isn't the company's biggest die-size in consumer GPUs ever. That dubious honor goes to the company's Fiji XT silicon which powered the company's R9 Fury X, coming in at a staggering 596 mm² in the 28 nm process. For comparison, AMD's current Polaris 20 XTX-based RX 580 chip comes in at slightly less than half the confirmed RX Vega's die-size, at a much more yield-friendly 232 mm². NVIDIA's current top-of-the-line Titan Xp comes in at a slightly smaller 471 mm² die-size.
Source: Raja Koduri @ Twitter
Add your own comment

23 Comments on AMD Raja Koduri Confirms RX Vega Die Size at 484 mm²

#1
PerfectWave
Die size is bigger cos include memory chip?
Posted on Reply
#2
uuuaaaaaa
PerfectWave, post: 3687233, member: 165997"
Die size is bigger cos include memory chip?
No this is just the GPU part, it does not include HBM.
Posted on Reply
#3
TheLostSwede
PerfectWave, post: 3687233, member: 165997"
Die size is bigger cos include memory chip?
No, the memory is mounted on an interposer, it's not part of the die. That would waste too many chips and would make the chips prohibitively expensive.
Posted on Reply
#4
xkm1948
Here is to hope that big die gets some good performance (gaming, neither mining nor "Prosumer use") later down the road.
Posted on Reply
#5
ernorator
This brings back (not so good) memories about 2900 xt, 420mm2


The only good thing about it was HD4xxx after
Posted on Reply
#6
MxPhenom 216
Corsair Fanboy
TheLostSwede, post: 3687239, member: 3382"
No, the memory is mounted on an interposer, it's not part of the die. That would waste too many chips and would make the chips prohibitively expensive.
He might be talking about the high bandwidth cache which would make for large die sizes.
Posted on Reply
#7
Joss
I wonder if they could put two chips together working as one (not crossfire) using the Infinity Fabric, somewhat like they did with Threadripper, would that be feasible?
Posted on Reply
#8
uuuaaaaaa
Joss, post: 3687262, member: 152251"
I wonder if they could put two chips together working as one (not crossfire) using the Infinity Fabric, somewhat like they did with Threadripper, would that be feasible?
Vega supports infinity fabric, however I think that is what they will be doing with Navi.
Posted on Reply
#9
zo0lykas
can some1 please explain me what means "Die Size"
Posted on Reply
#10
Hugh Mungus
xkm1948, post: 3687247, member: 50521"
Here is to hope that big die gets some good performance (gaming, neither mining nor "Prosumer use") later down the road.
We've seen an all-in-one card with older drivers (main improvement will be gaming), wx9100 will be for professional use only with certified drivers or something and rx vega will be for gaming only and its die is smaller than vega FE's die for no obvious reason. Maybe vega 10 has much lower latency or something because of the shorter distances between its components or something. Interesting to see the wx9100 and a "small", yet conplete vega 10.
Posted on Reply
#11
Basard
zo0lykas, post: 3687276, member: 152086"
can some1 please explain me what means "Die Size"
The die is one of the many silicon squares that they cut out of the wafer during production.
Posted on Reply
#12
zo0lykas
so basicly if number lower its better?

Basard, post: 3687306, member: 33749"
The die is one of the many silicon squares that they cut out of the wafer during production.
Posted on Reply
#13
Hugh Mungus
zo0lykas, post: 3687348, member: 152086"
so basicly if number lower its better?
More surface area is more expensive, but also provides better cooling, so to a point, yes.
Posted on Reply
#14
uuuaaaaaa
zo0lykas, post: 3687348, member: 152086"
so basicly if number lower its better?
In terms of performance bigger (within a gpu generation) usually means better.
Posted on Reply
#15
Hugh Mungus
uuuaaaaaa, post: 3687608, member: 98273"
In terms of performance bigger (within a gpu generation) usually means better.
But smaller with the same amount of streamprocessors is generally better.
Posted on Reply
#16
uuuaaaaaa
Hugh Mungus, post: 3687624, member: 172152"
But smaller with the same amount of streamprocessors is generally better.
That would require a smaller process node that is why I stated "within a gpu generation".
Posted on Reply
#17
Hugh Mungus
uuuaaaaaa, post: 3687626, member: 98273"
That would require a smaller process node that is why I stated "within a gpu generation".
Not necessarily. Optimizations can make almost as much difference as a smaller proces node sometimes if older versions were barely optimized.
Posted on Reply
#18
uuuaaaaaa
Hugh Mungus, post: 3687631, member: 172152"
Not necessarily. Optimizations can make almost as much difference as a smaller proces node sometimes if older versions were barely optimized.
Sure, that is true, but different uarch, would not qualify as the same GPU generation in my book.
Posted on Reply
#19
medi01
So bigger than latest titan although nowhere as fast and quite a bit more power hungry.

Wake me up when Navi is there, as Vega woes are somehow not Raja's fault.
Posted on Reply
#20
Hugh Mungus
medi01, post: 3687683, member: 158537"
So bigger than latest titan although nowhere as fast and quite a bit more power hungry.

Wake me up when Navi is there, as Vega woes are somehow not Raja's fault.
Titans are like bad geforce cards with bigger dies. At least vega FE has some workstation optimizations. Only useful for the budget pro or rich arabian oil guy's son that wants everything computer related, but that's why there will also be rx vega and the wx 9100 for gamers and professionals with a decent budget respectively. Rich arabian oil guy's son will buy those as well, because why not?!
Posted on Reply
#21
Darmok N Jalad
ernorator, post: 3687251, member: 162422"
This brings back (not so good) memories about 2900 xt, 420mm2


The only good thing about it was HD4xxx after
Actually, AMD followed up the vacuum-loud 2900XT with the highly successful 38x0 series. The 3870 was one of my favorite cards, as it could easily run Bioshock with DX10 goodness, and it matched up with the 8800 GT quite well.
Posted on Reply
#22
Basard
zo0lykas, post: 3687348, member: 152086"
so basicly if number lower its better?
For die size, if you're talking about compute power, a bigger die will give you more computing power always, unless cavemen designed the die.
If all things are equal--same transistor/trace size inside the die (same manufacturing process) then bigger is always better.... The bigger the die, the more transistors you can cram into it, therefore, the more you can process in one cycle.
When they shrink down to a new process, say from 24nm to 14nm, the die shrinks to about half it's original size and keeps the same amount of transistors inside. Both chips can process the same amount of information in one cycle, just that the 14nm process die will do it with less power consumption.
There's die size and process size. Normally when they shrink down to a new process, they will just add more transistors to make up for the freed up power consumption.
For apples to apples comparison, a smaller die and process version of the same chip will be a 'better' chip.
So, yes, a smaller die is better. And no, a smaller die is not better.
Posted on Reply
#23
Vayra86
Joss, post: 3687262, member: 152251"
I wonder if they could put two chips together working as one (not crossfire) using the Infinity Fabric, somewhat like they did with Threadripper, would that be feasible?
That's probably what Navi will look like.

And it is the path forward for almost everything. Nvidia has NVLink, Intel has Quickpath, etc.
Posted on Reply
Add your own comment