AMD: Speculation, Rumors, and Discussion (Archive)

Status
Not open for further replies.
Ryan Smith said:
For what it's worth, when Tonga was released I was explicitly told that it supported FP16. So if that's not the case, that's a change in what AMD is saying.
Click to expand...
Yeah... That's odd. Newest marketing material clearly states that GCN3 only supports fp16/int16 on APU.

I have only personal experience about low level optimizing/profiling of GCN1 and GCN2. Never tested GCN3 personally. Tonga was a mid-tier card, so I never had it in my workstation. Fiji on the other hand was a 4 GB card. Powerful, but not really suited for driving game development tools on 3x high resolution multimon setup. Many devs actually bought GeForce Titans solely because of the memory amount. Game development tools tend to need more GPU memory than games, as you need to be able to open multiple worlds/scenes simultaneously. Radeon 480X is actually a nice product in this sense. Cheapest 8 GB graphics card, pretty fast and latest GCN4 architecture. I could see it becoming quite popular among game developers, especially if the forthcoming console refreshes are based on GCN4 architecture.
 
From pcper interview with raja. Polaris uarch was designed around Polaris 11. Polaris 11 is mobile focused and design decisions made limited clock vs what they would be if the target was higher pref/tdp.

I guess Depending on oem and apple demand for mobile polaris 10 parts we could also be getting the dregs on desktop as well.

From phone blah blah blah
 
Frenetic Pony said:
Feh, in terms of die space and tdp AMD was close onto Nvidia back with the Fury X/980ti. Now silicon for silicon and watt for watt the Rx480 is significantly behind. But that doesn't seem to matter when all you can see are "out of stock" signs all over the place https://www.nowinstock.net/computers/videocards/amd/rx480/
Click to expand...
Maybe Overclockers has them all :)
More seriously maybe logisitcal-distribution issues.
Still has a lot of stock over here in UK: https://www.overclockers.co.uk/pc-components/graphics-cards/amd/radeon-rx-480
https://www.scan.co.uk/shop/computer-hardware/all/gpu-amd/radeon-rx-480-pci-e
Cheers
 
Ryan Smith said:
For what it's worth, when Tonga was released I was explicitly told that it supported FP16. So if that's not the case, that's a change in what AMD is saying.
Click to expand...
At least the driver still reports FP16 Precision still to DirectX:
„Minimum Precision 16/32-bit“

I think the news here is, that's applying to INT16 as well.

Gipsel said:
The outer perimeter is probably the only die shot part in it. And they probably used a weird mix of a floor plan and some kind of a schematic block diagram (as nV usually does as well in these fake die shots) "to highlight certain things". A floor plan shows the actual layout of the stuff on the die but I wouldn't believe that the SPs look like these small squares and a vector unit like a 4x4 assembly of such squares without any registers nearby. The layout of the SIMD units appear to be fairly different judging from a comparison with actual die shots of GCN-GPUs.
Click to expand...
Yeah, probably. I look forward to the first photos without probablys and maybes where everything checks out to theorycrafting guidelines. :)
 
Last edited:
CarstenS said:
I think the news here is, that's applying to INT16 as well.
Click to expand...
Yeah, they didn't talk about int16 before GCN4. Could be that someone in marketing just mixed up fp16/int16.

Int16 is important. Most address calculations (raw/structured/typed buffer load/store) can be done with int16, especially for 2d buffers (65536x65536 address space). I would also love to see native 16 bit LDS storage types. Pack/unpack to 32 bit is extra ALU.
 
Updated RX 480 TDP issue ...
Good morning. Many people are asking me to add other cards that are out of spec to the OP like the Asus Strix 960.

There are a few fundamental differences. Most cards that are out of spec over-draw from the external PCI-E power connectors.

The Asus Strix 960 has issues with power spikes on the PCI-E slot according to Tom's Hardware. Other 960s do not.

Having said that, the reason I do not add these to OP is because the spikes are the not the issue; the average is. A power spike can last 1ms, if the average is taken over say 1 minute it will be lower than the maximum 75W. With the RX480, both retail and review samples from different board vendors have had this issue identified AND THE AVERAGE POWER DRAW EXCEEDS THE MAXIMUM SET BY SPEC

It is not about power spikes, it is about sustained power draw above the 75w limit from the PCI-E slot. If anyone can find another card that exhibits this issue I will gladly add it to the OP.
Click to expand...

UPDATE 10 (MILESTONE!) - RETAIL CARD
Hardware.fr confirms this too. They also have a retail version (Sapphire 480) which exhibits the same problem. They also confirms the power usage going over 150W with both a review and retail version of the card.

They added also something interesting, they removed power and temperature limits and tested the card with no OC. The card pulled almost 200W in Witcher 3.
Click to expand...

https://www.reddit.com/r/Amd/comments/4qfwd4/rx480_fails_pcie_specification/
 
Last edited:
@sebbbi I haven't seen anything mentioned, but is FP16 support of polaris also double throughput of FP32, or merely double storage density? Thanks! :D

3dilettante said:
Could the way it's being measured be throwing off the boot-time calibration?
Click to expand...
What one-time calibration are we talking about...?

Esrever said:
People buy like 6 cards to mine and the mobo has to supply like 80W to each gpu. Pulling near 500W through the pcie means 500Ws from a 4 pin or 6 pin on the mobo. Thats a lot of power through the mobo circuitry.
Click to expand...
If a mobo has 4 PCIe 16x slots, then traces and passive components are sized to handle current demands. Btw, most mobo power comes through the ATX 24-pin connector, although I've seen that boards with a lot of PCIe 16x slots often have one (or even two) auxiliary molex 4-pin power connectors as well near the PCIe slots.

The 4/8-pin CPU AUX connector is probably reserved for CPU use I would think, as you wouldn't want sudden power demands from other devices to potentially jeopardize CPU stability. ;)

Ext3h said:
Second, current follows the path of least resistance. If the card draws a majority of the power from the motherboard, it's because the resistance is marginally lower on that path, and AMD isn't actively balancing the power draw.
Click to expand...
I wouldn't expect mobo to offer a lower resistance path than a fairly high-gauge cable coming straight out of the PSU to the graphics card, not when mobo power first has to go through significant distance of paper thin traces and then up through the additional PCIe connector, which has teensy-tiny pins one might add... :)

What you *should* be worried about is spikes in power draw due to insufficient capacitor capacity on the inside, disrupting the system stability. But I don't see that happen, given that the VR setup of the 480 is pleasantly oversized.
Click to expand...
Yeah, the video linked previously in this thread of the guy explaining RX 480 board layout went over this. Power spikes hit an on-board 12V capacitor bank + inductors, and should thus barely affect 12V voltage stability. Also, there's additional capacitor filtering on the motherboard, and power supplies today are very good at handling power surges/voltage dips, especially advanced power supplies with digital control of their VRMs (like Corsair's i-series for example.)

This should not be anything to worry about.

pharma said:
AMD Radeon RX 480 4GB Video Cards At-Launch Have 8GB of Memory
We asked AMD if they would allow us to host the BIOS to allow the 4GB cards to be flashed to an 8GB card, but they said absolutely not.
Click to expand...
Click to expand...
Lol! When hell freezes over, they'd allow something like that. ;) :LOL:
 
Last edited:
Kaotik said:
I can't even begin with how f#cked up that post has become. The OP for example is going on and on about 110 w 150 TDP, not realizing that he's talking about 2 different things, GPU power and board power.

I'm quite sure PCI-SIG didn't just get one perfect sample and approve the product line on that
Click to expand...
If they got a sample at all ... I just heard about "internal testing" and meeting PCI-SIG standards. As another Rx480 user posted:

This isn't anywhere near being scientific, but I think it's accurate enough to confirm the problem. Running stock clocks with stock voltage while running ethereum mining = 83w from the 6 pin connector, and 88w from the PCI-e slot. That's a violation of both ATX and PCI specs. I don't particularly mind it violating the ATX spec as a quality 6 pin connector can provide 200w without issue. The PCI-e slot, on the other hand, is an issue. I bought 4 of these cards today, and intend (intended?) on setting them up on a Rampage 5 motherboard. I don't think even a top end motherboard like that will be able to supply 352w to the PCI-e slots, even using the 4 pin Molex. Wish Asus had used a 6 pin instead..

If AMD can provide a BIOS update for the cards that forces 75% of the current through the 6 pin, problem solved. If that's not possible through software, then these cards should be recalled or they should have a warning label on them about possible motherboard damage when using crossfire.
If anyone is interested, I can test other GPUs as well with my setup. Either Hawaii or Tahiti.
Click to expand...
 
Ryan Smith said:
For what it's worth, when Tonga was released I was explicitly told that it supported FP16. So if that's not the case, that's a change in what AMD is saying.
Click to expand...
According to the GCN3 ISA manual, it does have (non-packed) FP16 arithmetic instructions. But I do not see any U16/I16 instructions except those datatype converting instructions. So native 16-bit integer is probably new, I guess...

Oh, just checked again. There are 16-bit integer instructions available in GCN3...
 
Grall said:
I wouldn't expect mobo to offer a lower resistance path than a fairly high-gauge cable coming straight out of the PSU to the graphics card, not when mobo power first has to go through significant distance of paper thin traces and then up through the additional PCIe connector, which has teensy-tiny pins one might add...
Click to expand...

But then why the card consumes more through the Mobo than the 6pin? wouldn't be the opposite? Also would the interference make an audible effect on the audio in low tier Mobos(less filters)?
 
Wondering why OCUK is continuing to advertise 110w TDP? I assume they are following included manuf. specs.

Specification:-
- XFX Backplate
- GPU: Polaris
- Stream Processors: 2304
- Core Speed: 1120MHz
- Boost Speed: 1328MHz
- Memory Speed: 8000Mhz
- Memory interface: 256-Bit
- Memory capacity: 8192MB GDDR5
- PCI-Express X16 lane required
- 350W or greater PSU required
- 110W TDP
- Power Connectors: 1x 6-pin required
- Display Outputs: 1x HDMI, 3x DisplayPort
- Warranty: 2yr
Click to expand...
https://www.overclockers.co.uk/xfx-...ddr5-pci-express-graphics-card-gx-238-xf.html
 
Status
Not open for further replies.

Similar threads

hoom
AMD: Zen 4, Speculation, Rumours and Discussion
8 9 10
Replies
199
Views
34K
MfA
M
Rys
Reopening Architecture and Products + rules and guidelines changes
3 4 5
Replies
90
Views
13K
Shifty Geezer
Shifty Geezer
D
AMD: Zen 2 (Ryzen/Threadripper 3000?, Epyc 8000?) Speculation, Rumours and Discussion
16 17 18
Replies
340
Views
147K
PSman1700
P
Shortbread
Next-Generation NVMe SSD and I/O Technology [PC, PS5, XBSX|S]
105 106 107
Replies
2K
Views
189K
Karamazov
K
L
A rambling history of AMD and nVidia's rise to the top
2
Replies
20
Views
5K
Leovinus
L
Back
Top