These are the parts that are supposed to ship mid year right?
There still room for a 490 and 490x later in the year.
AMD: Speculation, Rumors, and Discussion (Archive)
- Thread starter iMacmatician
- Start date
- Status
These are the parts that are supposed to ship mid year right?
There still room for a 490 and 490x later in the year.
iMacmatician
Regular
AMD is not going to just lay down and let Nvidia walk all over them. These specs can't be true.
I miss the days of small die strategy when AMD let Nvidia walk all over its shoe laces.
Plus there is the dual Fiji card waiting for some VR wind.
D
Deleted member 13524
Guest
That would be true if the high-end Pascal was coming at the same time or before the first two Polaris GPUs.
According to that interview, Raja Koduri seems pretty convinced they're several months ahead of nVidia with the FinFet chips. There's also the thing with Jensen showing off Maxwell chips as Pascal in PX2, which isn't a good sign at all.
Maybe the Pascal chips will only release in late 2016 with another two Polaris coming out in early 2017, leaving little time for nVidia to keep their high-end chips without competition.
With these two chips against Maxwell, AMD can best nVidia in their most popular offerings:
- Polaris 10 against GM107 for performance notebooks. It might give GM206 performance at a lower price and power consumption.
- Polaris 11 against GM204 (more specifically GTX 970), giving GTX 980 performance at a lower price and power consumption
Neptune is just Pitcairn with a different name.
According to Samsung, "enthusiast GFX" will also use two stacks for 512GB/s and 8GB (it would fit Polaris 11 like a glove):
The cards replacing Hawaii and GM204 will probably have 8GB. Big Polaris and Big Pascal should be the only ones using 4 stacks.
gamervivek
Regular
The supposedly Polaris GPUs going around on Zauba are about Rs. 15-20k and 40-45k. Fiji chips were going around for 75k. So doesn't look like a Fury replacement. It had to beat Fury X by 30% or so just to be on par with 980Ti overclocking anyway.
Nope.
Nope.
D
Deleted member 13524
Guest
If the new cards are coming until summer and dual-Fiji in April, it wouldn't make a lot of sense to cannibalize the later's sales just a couple of months after release.
My friend has a Clevo with a 8970M and I've looked at it and its capabilities in person. There's not a single shred of evidence pointing to the GPU being any other than a Pitcairn with GCN1.0.
There's no TrueAudio and no FreeSync.
AMD may have given it another name for different revisions of the same architecture, but it's essentially the same chip. Same size, same performance-per-clock.
Insured prices based off of import taxes have very little to with retail prices, specially with Indian customs involved and this does change from year to year, just go to any international airport and ask a customs guy, they have a book they get every year from the government that gives them flat rates for different electronic products and what the import taxes should be around. So trying to get make something out of the Rs I wouldn't even try. Also forgot to add exchange rate is quite different now too. Which the Rs is an at an all time low in the past 5 years.
Neptune is Pitcrain, just faster and was made for OEM market.
Last edited:
The days when AMD had the chance to take the performance crown my a mile, but didn't, so Nvidia was still able to sell plenty of GPUs.
Also the days when Nvidia's architecture wasn't very competitive in terms of perf/mm2.
I don't think those days are coming back.
Just what everybody's waiting for!
gamervivek
Regular
https://forum.beyond3d.com/posts/1851261/
Doesn't matter. Hence AMD, or more appropriately RTG, are using Polaris as an umbrella term for the different IPs going into a chip.
Tonga isn't that different from Tahiti in performance either. The M370X in apple's retina macbook shows up as VenusXT from the same family but has been termed as a cape verde rebrand. Too OT for now.
D
Deleted member 13524
Guest
People with a 7970M have successfully flashed its bios to a 8970M and to a M290.
It's the same chip.
Tonga is very different from Tahiti. It may get the same performance as a higher-clocked Tahiti, but it does so with 66% of its memory bandwidth and lower overall power consumption.
There's a new UVD/VCE, color compression, FP16 instructions, TrueAudio, FreeSync, better VSR support and up to 4x better geometry performance on tesselation at >=32x.
Die area is different, transistor count is different... It's not the same chip.
It's the same chip.
Tonga is very different from Tahiti. It may get the same performance as a higher-clocked Tahiti, but it does so with 66% of its memory bandwidth and lower overall power consumption.
There's a new UVD/VCE, color compression, FP16 instructions, TrueAudio, FreeSync, better VSR support and up to 4x better geometry performance on tesselation at >=32x.
Die area is different, transistor count is different... It's not the same chip.
Taking AMD's marketing numbers for power efficiency as accurate (not the best odds) a Hawaii replacement at 2.5X power efficiency is going to be a threat to Fiji even at 150W unless AMD does something more to hobble it. Perhaps cut the memory bus down enough to harm its performance and capacity?
Assuming Hawaii operates at 300W, the 390X gets something that was uncomfortably close to Fury Nano, which was surprisingly close to the rest of the Fury range in a number of scenarios.
Then we get the architectural changes for Polaris, which could bump up performance within the constrained hardware and power environment.
If the 480X doesn't opt for HBM, there could be a memory power consumption penalty, although if we take 2.5X as gospel there will 30 or so Watts to play with just for that. It would take a cut-down bus to keep GDDR5 from beating Fury in capacity and hurt the 480 enough to allow the limited benefit Fury got from its supposed bandwidth advantage over Hawaii to be compelling.
A third-party custom board could have quite a bit of breathing room to really cut into the justification for most of Fiji's range.
D
Deleted member 13524
Guest
Maybe a safe(r?) bet would be to have low clocks (<1GHz) and pair it with power distribution circuitry and a cooling solution that won't allow for substantial overclocks but brings higher power efficiency and a smaller PCB.
AMD lowering the price of Nano can be an indicator that the 480X will be going right at its place in power consumption, form factor and performance.
How awesome would it be if the 480X was a 130W card with a performance between the 390X and the Fury, with 8GB HBM2, the form factor of the Nano and a very quiet cooler?
With HDMI 2.0 and full HEVC decoding/encoding, it would be the perfect card for many in the HTPC crowd.
That would help contain the silicon on a reference design. A third-party custom board would have more leeway unless additional measures are taken with the GPU or AMD restricts that path.
I was thinking earlier in one of the Fury threads about what would have happened if there were a modest boost in CU count while the graphics pipeline were allowed to expand past GCN's limits on shader engines and backends--which the speculative numbers given earlier permit.
Possibly, it's not just the Nano that would worry.
It might be a maybe on HBM2 at that capacity for what is ostensibly a Hawaii replacement.
You could also fit 3 of them in the power budget of an X2 card. Getting three work with the current switch setup could be problematic, unless the new interconnect claims mentioned earlier in the thread could point to a new way of handling communication.
Otherwise, it's a Gemini board coming in under the Fury and FuryX power budget. We'll have to see how much time the delayed Gemini product will have, since it should have been out earlier and the silicon that stands to replace it is beginning to march.
D
Deleted member 13524
Guest
If Polaris 11 really has 48 CUs and significantly higher geometry output than Fiji, then Fiji had better go through a significant refresh (I'm thinking higher GPU clocks and upgrading to 8-HI HBM stacks).
Otherwise the Fury cards could be threatened too, of course.
Then again, if these 48 CUs top at 850MHz with little overclock headroom, there's no reason to worry.
AMD's statements so far point to Gemini being marketed as the best-possible card for VR. That means using LiquidVR for one GPU per eye and lower latency by not having the same GPU alternate between one viewpoint and another.
For VR, a 3-GPU solution would probably leave the third GPU aside for anything but pure compute jobs (overkill?). Using it for rendering would probably bring in a lot of unwanted latency.
iMacmatician
Regular
On the subject of the memory bus, perhaps Tonga is a good comparison point for the Polaris chip with a rumored 3072 SPs, since both should have the improved memory compression over Hawaii and earlier chips (I would expect Polaris to improve further if anything).
Code:
Memory Bandwidth
Chip SP GFLOPS Configuration (GB/s) FLOPS/byte
HD 7970 GE 4313 384-bit, 5.5 Gbps 264 16.3
R9 380X 3973 256-bit, 5.7 Gbps 182 21.8
R9 390X 5914 512-bit, 6.0 Gbps 384 15.4
R9 Fury X 8602 4096-bit, 1.0 Gbps 512 16.8
R9 480X 6144 2048-bit, 2.0 Gbps 512 12.0
R9 480X 6144 384-bit, 6.0 Gbps 288 21.3If the name is indeed 480X instead of "490X," then that could be reason for stock configurations to be 6 GB (the 380 series have 4 GB) which would limit its memory capacity increase over the current Fury lineup. If, in addition, the speculation of a 8 GB Fiji becomes reality, then the 2016 Fiji will have more memory than the 480X. That being said, I agree with the above posters that the 480X needs low clocks or the Fury parts need a boost, otherwise there is likely to be no meaningful performance gap between them.
Code:
R9 290X 137%
R9 390X 150%
R9 480X 850 MHz 132%, 146%, 159% [+0%, +10%, +20% from architecture] [estimates]
R9 480X 1000 MHz 156%, 171%, 187% [+0%, +10%, +20% from architecture] [estimates]
R9 Fury 172%
R9 Fury 1050 MHz 181% [estimate]
R9 Fury X 187%
R9 Fury X 1100 MHz 196% [estimate]
D
Deleted member 13524
Guest
Hum.. a 384bit bus would mean:
1 - That AMD wouldn't launch a single new solution with HBM throughout the whole 2016 (Hynix and/or Samsung probably wouldn't like that)
2 - At least 12 GDDR5 chips (AFAIK each GDDR5 chip can only connect through a 32bit bus max)
3 - A large PCB to house 12 VRAM chips
4 - A lot of power consumption coming from GDDR5, specially compared to HBM
It doesn't really fit AMD's statements about all Polaris solutions being very power-efficient.
Unless there's a severe shortage of HBM production or some other terrible event, I don't think AMD will go back to using GDDR5 in configurations that would need to go over 256bit wide. Or probably even less than that.
Polaris 10 had to be very cheap so it's probably coming with 128bit GDDR5.
The only way I could see GDDR5 being used in Polaris 11 is if it was the new GDDR5X memory at 10Gbps in a 256bit configuration. That would somewhat undermine points 2, 3 and 4, but point 1 still stands.
Samsung and Hynix didn't start volume production of HBM2 without having customers already signed on to purchase the chips. And I doubt Fiji alone (even a refresh using HBM2) would be enough to justify the production.
A more probable solution IMO is Polaris 11 / 480X using two stacks of HBM2 with a clock that is lower than the maximum rated 2Gbps for increased yields.
1 - That AMD wouldn't launch a single new solution with HBM throughout the whole 2016 (Hynix and/or Samsung probably wouldn't like that)
2 - At least 12 GDDR5 chips (AFAIK each GDDR5 chip can only connect through a 32bit bus max)
3 - A large PCB to house 12 VRAM chips
4 - A lot of power consumption coming from GDDR5, specially compared to HBM
It doesn't really fit AMD's statements about all Polaris solutions being very power-efficient.
Unless there's a severe shortage of HBM production or some other terrible event, I don't think AMD will go back to using GDDR5 in configurations that would need to go over 256bit wide. Or probably even less than that.
Polaris 10 had to be very cheap so it's probably coming with 128bit GDDR5.
The only way I could see GDDR5 being used in Polaris 11 is if it was the new GDDR5X memory at 10Gbps in a 256bit configuration. That would somewhat undermine points 2, 3 and 4, but point 1 still stands.
Samsung and Hynix didn't start volume production of HBM2 without having customers already signed on to purchase the chips. And I doubt Fiji alone (even a refresh using HBM2) would be enough to justify the production.
A more probable solution IMO is Polaris 11 / 480X using two stacks of HBM2 with a clock that is lower than the maximum rated 2Gbps for increased yields.
- Status
