Bondrewd
Veteran
Nah, it's due to all the MAC piles required for GEMM going brrrrr at alarming speeds.
See any other ML part; they all hit very nice density figures for their respective nodes.
Nah N7+ has all the same options as vanilla N7.
Nvidia Ampere Discussion [2020-05-14]
- Thread starter Man from Atlantis
- Start date
-
- Tags
- nvidia
See any other ML part; they all hit very nice density figures for their respective nodes.
Isn’t cache also relatively dense? A100 has a load of that too.
Bondrewd
Veteran
Kinda, depends on the implementation.
Zen2 CCD is mostly SRAM, but is nowhere near the peak theoretical density (or what mobile SoCs achieve for that matter).
Tensor cores take only about 10% of the die in Turing, even if that would be 20% for Ampere, that can not explain the high transistor density.
If that is the case, N7+ is a more plausible explanation for the high density compared to vanilla N7.
Last edited:
Bondrewd
Veteran
They're xtor-dense MAC piles.
Nah, see RDNA2 stuffz and Zen3 and Kirin 990 5G.
Just read about this new Graphcore AI processor, eclipsing the A100, with 59,4 billion transistors on 823 mm2.
That is a transistor density of 72 MT/mm2.
As this is more than the max of 66,7 MT/mm2 for vanilla 7nm, this also must be using N7+.
The high density is a result of the 900 MB on chip memory this Graphcore has.
1 bit of SRAM requires 6 transistors, so the 900 MB translates to 43,2 B transistors...
That is a transistor density of 72 MT/mm2.
As this is more than the max of 66,7 MT/mm2 for vanilla 7nm, this also must be using N7+.
The high density is a result of the 900 MB on chip memory this Graphcore has.
1 bit of SRAM requires 6 transistors, so the 900 MB translates to 43,2 B transistors...
Last edited:
It doesn't say it's using HP cells, so doesn't need to be N7+ necessarily.
Last edited:
If it would use HD at 91.2 MT/mm2 (which is only suitable for mobile SoCs AFAIK), the 900 MB would fit in 473mm2, leaving the remaining 16.2 B non SRAM transistors on 350 mm2, for a density of 46.3 MT/mm2. I don't think so.
And in case you would further doubt, SRAM is the densest way you can pack transistors.
Last edited:
Or then TSMCs estimates on transistor density aren't counted on just SRAM cells, but assume certain proportion of memory vs logic vs phys and whatnot.
For what it's worth, Hexus says it's N7
https://hexus.net/tech/news/cpu/144154-graphcore-ipu-machine-m2000-1u-blade-capable-1petaflop/
No that would be ridiculous. The first thing tried on a new process is always some SRAM and based on that density is specified.
Graphcore says: "...Graphcore Colossus™ Mk2 GC200 IPU. Developed using TSMC’s latest 7nm process"
Though they don't tell anywhere it is N7+, it doesn't need a genius to figure that out.
If it proves not to be N7+, I'll buy you a beer, or maybe something stronger
Last edited:
D
Deleted member 13524
Guest
900MB of SRAM.
Wow.. imagine having that as framebuffer.
Wow.. imagine having that as framebuffer.
I'm not an expert on these on any level, but let's seeNo that would be rediculous. The first thing tried on a new process is always some SRAM and based on that density is specified.
Graphcore says: "...Graphcore Colossus™ Mk2 GC200 IPU. Developed using TSMC’s latest 7nm process"
Though they don't tell anywhere it is N7+, it doens't need to be a genious to figure that out.
If it proves not to be N7+, I'll buy you a beer, or maybe something stronger
https://en.wikichip.org/wiki/7_nm_lithography_process#Industry
N7 with HD cells is supposed to offer around 91-92 MTrans/mm^2. We also know that with HD cells one SRAM cell is 0.027 µm^2, which fits into 1 mm^2 37 million times.
Since 1 SRAM cell is actually 6 transistors, that 37 million times turns into 222 million transistors.
So if the density was reported on just SRAM cells, it would be 222 MTrans/mm^2, not 91-92 MTrans/mm^2 for the HD cells, right?
(For what it's worth, wikichip did feel the need to point out that N7 can offer really dense SRAM cells)
Apparently you can't just compute density from SRAM cell size.
Like in this table a 14 nm SRAM cell is 0.05um2, which would be 6x20 MT/mm2, but as you see in the table it is actually only 6x13.7 MT/mm2.
In practice it must be even less, as for the 7nm Samsung chip with 0.026um2 cell size on that same page, the 256 Mbit SRAM is 69.3mm2, which equates to a mere 22 MT/mm2. If anybody has access to this paper, there might be more information there.
Megadrive1988
Veteran
Holy shit that's a lot!
A1xLLcqAgt0qc2RyMz0y
Veteran
This is the "Nvidia Ampere Discussion" thread.
Can the mod please remove the off topic posts.
Can the mod please remove the off topic posts.
D
Deleted member 2197
Guest
"A100 has now become the fastest GPU ever recorded on #OctaneBench: 446 OB4* #Ampere appears to be ~43% faster than #Turing in #OctaneRender - even w/ #RTX off!"
Is Cuda 11 also used for the Turing results? Nvidia has a history of not using newer versions of Cuda for older GPUs because it zaps some of the performance deficit.
Yes 50% faster what ive been reading on many places now. Quite good i think.
D
Deleted member 2197
Guest
The OctaneBench results are comparing A100 (no RT cores available ) vs the best Turing result (with RT cores).
Last edited by a moderator:
Does this benchmark even use RT cores?
Looking at Octanebench results page, the 1080Ti is tied with the 2080 and the Titan V is easily beating the Titan RTX.
https://render.otoy.com/octanebench/results.php?v=4.00&sort_by=avg&filter=&singleGPU=1