AMD: Speculation, Rumors, and Discussion (Archive)

[pharma]

AMD Reveals Fury X2 Compute Performance at VRLA Winter Expo keynote - Has 12 TFLOPS of single precision compute

Now according to Roy, the Fury X2 has around 12 TFLOPs of SP. In contrast, the Radeon R9 295X2 has around 11.5 TFLOPs of SP performance compute. The major difference of course is, dass die 295X2, is a 500W unit, while as the Fury X2 is a 375W unit so there is a massive power efficiency difference here. The Fury X2 is Roughly 40% more power efficient than the 295X2. Since SP compute Usually has a high correlation with gaming performance, this bedeutet, dass the Fury X2 will perform slightly more than at R9 295X2. Needless to say, this number is lower than what we expected, Although we do not know what clocks the card which running on and what the cooling solution which for that matter (to air-cooled solution created for the Tiki would have to be Significantly under-clocked than a liquid cooled solution). I have attached the full video presentation below, but here is a transcript of the critical moment: In contrast, the Radeon R9 295X2 has around 11.5 TFLOPs of SP compute performance. The major difference of course is that the 295X2, is a 500W unit, while as the Fury X2 is a 375W unit so there is a massive power efficiency difference here. The Fury X2 is roughly 40% more power efficient than the 295X2. Since SP compute usually has a high correlation with gaming performance, this means that the Fury X2 will perform slightly more than an R9 295X2. Needless to say, this number is lower than what we expected, although we do not know what clocks the card was running on and what the cooling solution was for that matter (an air-cooled solution created for the Tiki would have to be significantly under-clocked than a liquid cooled solution).

http://wccftech.com/amd-fury-x2-12-tflop-compute-vrla-winter-expo

 

[pTmdfx]

Regardless, Jawed was the only one expressing certainty that those are 4TFLOPS DP because the slide is aimed at a HPC audience (I don't see HPC anywhere in that fragment of a slide though, I guess he took it from the quad-channel RAM).

This piece of slide from FudZilla succeeded the previous story on a 16-core Zen APU with huge amount of I/O and also quad-channel memory. Also considering AMD has an HPC APU in its 2016-17 plan, and has claimed APUs for HPC of 200-300W TDP are in the works (in its HPC event in JP a few months ago), it is hard not to conclude it's the HPC APU.

 

[Deleted member 13524]

The clue is in the non-PCI Express link twixt CPU and GPU, which, apart from the fact that a MCM of CPU and GPU is never going to be a consumer product, is a pure HPC solution.

Why not? Intel offers a socketable MCM with their desktop Broadwell 4 core + GT3e, and will do so with Skylake's 4 core + GT4e. That 100GB/s link could be within the MCM.

What was once Greenland is now Vega, isn't it?

I don't think I've seen the name Vega outside rumor mill sites like wccftech. According to Raja Koduri, the 2016's new Radeons are all under the Polaris name, and GPU families have traditionally been kept for longer than a year (GCN1, GCN2, GCN3, Maxwell, Kepler, etc.).
It could be that Vega is 2017's GCN5 architecture, which is why they're not Polaris, but it sounds like a long shot IMO.

 

[Deleted member 13524]

AMD Reveals Fury X2 Compute Performance at VRLA Winter Expo keynote - Has 12 TFLOPS of single precision compute
http://wccftech.com/amd-fury-x2-12-tflop-compute-vrla-winter-expo

For that to be true, each Fiji chip has to be running at 730MHz. That doesn't make much sense for a 375W TDP and dual Fiji.
If the Nano with a 175W TDP manages to keep above 830MHz at all times, why would a setup with over twice the TDP of the Nano need lower clocks for each chip?

Besides, a dual-Fiji at 730MHz would probably perform worse than a single Fiji clocked at 1150MHz even in most AFR scenarios, since it would require a scalability of over 78% [= 1150/(730*2)]. It would also perform worse than a single 390X.

If those numbers are true, then the card is completely pointless for anything but VR at a low-ish power consumption.

 

[function]

Perhaps the Nano has different binning parameters?

Different combinations of PCI-E power + external create the 175 / 375 limits, and within those and given market factors it might make sense to target a slightly different threshold (of frequency and power) for Nano than for Fury X2.

 

[iMacmatician]

For that to be true, each Fiji chip has to be running at 730MHz. That doesn't make much sense for a 375W TDP and dual Fiji.
If the Nano with a 175W TDP manages to keep above 830MHz at all times, why would a setup with over twice the TDP of the Nano need lower clocks for each chip?

Besides, a dual-Fiji at 730MHz would probably perform worse than a single Fiji clocked at 1150MHz even in most AFR scenarios, since it would require a scalability of over 78% [= 1150/(730*2)]. It would also perform worse than a single 390X.

If those numbers are true, then the card is completely pointless for anything but VR at a low-ish power consumption.

Yeah, that number didn't make a lot of sense to me either.

Is it possible that the Fury X2 in the Tiki has a TDP of ≤ 300 W, because either that's the actual TDP of the Fury X2 or it's underclocked in the Tiki for some reason? Unless I'm missing something, the 375 W number for the Fury X2 seems to come from the two 8-pin power connectors, and I don't see why those automatically imply a 375 W TDP. But even 12 TFLOPS at 300 W doesn't seem impressive against the Nano with its 8 TFLOPS at 175 W.

My other thought was along the lines of function's post above. If one ignores the Nano then 12 TFLOPS isn't unreasonable assuming a 375 W TDP.

 

[function]

Yeah, that number didn't make a lot of sense to me either.

Is it possible that the Fury X2 in the Tiki has a TDP of ≤ 300 W, because either that's the actual TDP of the Fury X2 or it's underclocked in the Tiki for some reason? Unless I'm missing something, the 375 W number for the Fury X2 seems to come from the two 8-pin power connectors, and I don't see why those automatically imply 375 W. But even 12 TFLOPS at 300 W doesn't seem impressive against the Nano with its 8 TFLOPS at 175 W.

I might be mistaken in this, but isn't 375 likely to be 150 + 150 + 75 (PCI-E)?

Interconnect and cooling can take additional power (e.g waterpump + fans or triple fan vs single), and again, if the best returns for optimal PPW bins are on Nano then that's where it makes sense to send them.

 

[iMacmatician]

I might be mistaken in this, but isn't 375 likely to be 150 + 150 + 75 (PCI-E)?

Sorry, I was unclear. The second mention of 375 W in my post is supposed to refer to the Fury X2's TDP. So the Fury X2 could hypothetically have a 300 W TDP, and the clocks that give 12 TFLOPS were set with 300 W in mind, but its power connectors support 375 W for overclocking.

 

[Jawed]

Why not? Intel offers a socketable MCM with their desktop Broadwell 4 core + GT3e, and will do so with Skylake's 4 core + GT4e. That 100GB/s link could be within the MCM.

300W+ of MCM on consumer mobos. Yeah, right.

 

[3dilettante]

For that to be true, each Fiji chip has to be running at 730MHz. That doesn't make much sense for a 375W TDP and dual Fiji.
If the Nano with a 175W TDP manages to keep above 830MHz at all times, why would a setup with over twice the TDP of the Nano need lower clocks for each chip?

Is there a source for the 175W TDP?
I've seen a 175 Typical Board Power figure, which isn't the same thing.
From that same source, we see Furmark numbers where a system with a Fury board with 275 TBP is 83W higher than the Nano system, but 175 is not 83 less than 275.
Since both rely on TBP, both are suspect regardless of interpretation.
The Fury X also has a 275W TBP, and that pulls 216W more than Nano.

The PCIe bridge burns ~8W or so, making the budget at least a little less than 2x even if TBP were trustworthy.

If those numbers are true, then the card is completely pointless for anything but VR at a low-ish power consumption.

If those numbers are true, then a Hawaii replacement in the form of Polaris 11 is again shown to be dangerous, because this is really not giving much margin between what Hawaii can do on a per-GPU basis before cutting the power draw in half.

 

[CarstenS]

AFAIR AMD gave Typical Board Powers only for quite a while now. At least Hawaii-timeframe, probably earlier.

 

[function]

Is there a source for the 175W TDP?
I've seen a 175 Typical Board Power figure, which isn't the same thing.

I like techpowerup's reviews, they give a number of power draw figures.

http://www.techpowerup.com/reviews/AMD/R9_Nano/28.html

They show 195W "maximum" (Furmark) and 209W gaming "peak" (Metro: Last Light) for Nano. Their Fury X review shows a crazy high "maximum" of 432W and gaming "peak" of 280W.

Doesn't seem unreasonable that an X2 card might have to drop those Nano clocks a little.

 

[Deleted member 13524]

300W+ of MCM on consumer mobos. Yeah, right.

300+? If Polaris 10 consumes 50W or lower, I doubt Polaris 11 will consume more than 100W if the chip gets clocked to ~660MHz (3072 ALUs @ 660MHz to reach 4 TFLOPs).
And I don't think Zen is going to consume 200W, not even in its 16-core iteration. The 8-core version will probably stay in the ballpark of 100-130W, like Intel's current Xeon E5 offerings.

200-230W doesn't sound too unrealistic anymore, especially if you consider that AMD is selling a 220W CPU right now.

Doesn't seem unreasonable that an X2 card might have to drop those Nano clocks a little.

I wouldn't call going from 830MHz minimum to 730MHz a little drop...

 

[Jawed]

3072 ALUs @ 660MHz to reach 4 TFLOPs

HPC people really don't care about SP FLOPS.

 

[Deleted member 13524]

HPC people really don't care about SP FLOPS.

Yes, but my argument is that this could be a solution for gamers, not HPC people. A single socketable MCM that would blast away any Intel solution with eDRAM and that could fit inside relatively small box while offering >2x console performance.

 

[Kaotik]

Yes, but my argument is that this could be a solution for gamers, not HPC people. A single socketable MCM that would blast away any Intel solution with eDRAM and that could fit inside relatively small box while offering >2x console performance.

The difference is that such APU (even if it's MCM) hasn't been on AMD's roadmaps, HPC APU has.

 

[Deleted member 13524]

The difference is that such APU (even if it's MCM) hasn't been on AMD's roadmaps, HPC APU has.

You're right, I didn't remember these supposedly leaked slides.

The fudzilla slide clearly says 1/2 DP ratio, so this Greenland is probably something between ~4096 SP / 64 CU @ 1GHz and 6144 SP / 96 CU @ 667MHz.
Meaning it's most definitely not Polaris 11.

 

[gamervivek]

And there it is.

code customs values deliveries statement
C993 ~ Euro 1500 2 pieces for February 1, 2016 Polaris 11
C981 ~ 650 Euro 2 pieces 28 December 2015 48 pieces for January 30, 2016 Polaris 10
C980 ~ 560 Euro 7 pieces 1 December 2015 2 pieces of 8 January 2016 24 pieces for January 14, 2016 Polaris 10
D000 ~ 420 Euro 15 pieces for 14 January 2016 Polaris 10
C924 ~ 240 Euro 8 to August 13, 2015 test chip
C913 ~ 200 Euro 30 pieces for January 6, 2015
test chip

https://translate.google.com/transl...neration-mit-einer-1500-euro-karte&edit-text=

 

[Kaotik]

And there it is.
https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=en&ie=UTF-8&u=http://www.3dcenter.org/news/ein-weiteres-lebenszeichen-von-amds-polaris-generation-mit-einer-1500-euro-karte&edit-text=

Speculation is, you mean?
So far pretty much everyone else has agreed that C91x and C92x are Polaris 10, C98x's are Polaris 11 and based on price C993 would be dual Polaris 11. No clue on D000 though

 

[AnarchX]

Below the table is said: the right column is speculation. Afer C-PCB-numbers are "full" AMD has probably go to D.
It could be a dual Polaris11. This will probably blame Dual-Fiji, AMD could aim for 2x 12GiB GDDR5X 12Gbps @ 256-Bit, but would allow a nice ~$999 enthusiast board at a lower BOM than the HBM-experiment.