iMacmatician
Regular
Don't worry about it, I didn't even realize it the first time I read your post.I'm sorry, you're absolutely right. I saw the 9% error and didn't even stop to think about the sign
AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread
- Thread starter iMacmatician
- Start date
-
- Tags
- amd
gamervivek
Regular
420 also looks good for some other reason. Anyway, it would be grand if Fiji were only 420mm2 and smoking 290X, which would leave another gpu above it since the synapse leak did confirm that AMD have a gpu over 500mm2. AMD are still quite competitive against nvidia on die size but per shader efficiency and power efficiency has them lagging quite a ways behind. The latter also seems to translate into lower clocked cards which might also decide the Titan X vs. Fiji contest.
gamervivek
Regular
There were leaked slides from AMD around Titan X's release. Nothing about power except that WCE will have two 8-pins.
http://hexus.net/tech/news/graphics/81697-slides-amd-radeon-r9-390x-wce-presentation-leaked/
I forgot that I ran the numbers from it,
First numbers are for TitanX over 290X, TP for techpowerup, TR for techreport. Then how much faster 390X was in the slide, followed by how much faster/slower it is than Titan X. I didn't calculate some of them.
bf4 - 52.6 TP 53.4% TR-65.5% 390X - 60% ~4%
far cry 4 - 29.1 TP - 31.5 TR- 25% 390X - 55% 20.1% 17.8%
alien isolation - 38.7% TR - 36.4 390X - 65% 18.7%
tomb raider - 47.6% 390X - 50% ~
So it could be quite a bit faster depending on the game and about equal and slightly slower in others.
And its samples are still landing in India for about ~1250$.
http://hexus.net/tech/news/graphics/81697-slides-amd-radeon-r9-390x-wce-presentation-leaked/
I forgot that I ran the numbers from it,
First numbers are for TitanX over 290X, TP for techpowerup, TR for techreport. Then how much faster 390X was in the slide, followed by how much faster/slower it is than Titan X. I didn't calculate some of them.
bf4 - 52.6 TP 53.4% TR-65.5% 390X - 60% ~4%
far cry 4 - 29.1 TP - 31.5 TR- 25% 390X - 55% 20.1% 17.8%
alien isolation - 38.7% TR - 36.4 390X - 65% 18.7%
tomb raider - 47.6% 390X - 50% ~
So it could be quite a bit faster depending on the game and about equal and slightly slower in others.
And its samples are still landing in India for about ~1250$.
This is the insurance cost for send the packages, not the price of the pieces inside the packet. Nor the retail prices ..
gamervivek
Regular
Sounds possible, though 980 was going around for its retail price.
http://i.imgur.com/ox37hOP.jpg
I'd rather have fiji at 20% faster and 25% costlier than 0-5% faster and 20% cheaper. I was thinking that it would be at least 70% faster than 290X in a few games at 4k, sounds like a distant possibility with the vram limitation.
http://i.imgur.com/ox37hOP.jpg
I'd rather have fiji at 20% faster and 25% costlier than 0-5% faster and 20% cheaper. I was thinking that it would be at least 70% faster than 290X in a few games at 4k, sounds like a distant possibility with the vram limitation.
Three things:
- The slides were obviously faked.
- The latest info I heard about puts it at 90% of Titan X, which incidentally fits in with the sub GHz core clocks the circulating Fiji samples have.
- Fiji is going to be a prestige priced, stopgap technology demonstrator, no point in building up inflated expectations.
- The slides were obviously faked.
- The latest info I heard about puts it at 90% of Titan X, which incidentally fits in with the sub GHz core clocks the circulating Fiji samples have.
- Fiji is going to be a prestige priced, stopgap technology demonstrator, no point in building up inflated expectations.
gamervivek
Regular
What was so obvious? Not even 70% faster than 290X in choicest games, the slides were a downer.
The 8 GB Dual-Link Interposer, obviously.
70%? What were you expecting? Hawaii is already 438mm², there is no new process yet and no significant architecture changes that could provide higher efficiency. 4096 SP's provide a maximum theoritical uplift of 45% from Hawaii at the same clock. Fiji is likely to clock lower to save power and the scaling from more execution units is rarely linear. That gives Fiji a realistic boost of maybe 36 to 40% from Hawaii.
gamervivek
Regular
DLI sounded much like GDDR5's clamshell mode. It was quite a well thought up fake then.
At 4k the 8GB could come in handy alongside the tessellation and memory bandwidth improvments, granted 285 didn't set the world on fire. Them using reference 290X could affect a few percentage points as well which is also reflected in different reviews of Titan X with different 290X cards. As for clocks, the slides were for the WCE clocked at 1050Mhz and they do give a fine print for the performance improvement slide. Expected improvement there as well.
At 4k the 8GB could come in handy alongside the tessellation and memory bandwidth improvments, granted 285 didn't set the world on fire. Them using reference 290X could affect a few percentage points as well which is also reflected in different reviews of Titan X with different 290X cards. As for clocks, the slides were for the WCE clocked at 1050Mhz and they do give a fine print for the performance improvement slide. Expected improvement there as well.
Jawed
Legend
I reckon there are no game tests that are ALU limited on Hawaii. If there were, it's unlikely Titan X would be 35% faster than Hawaii at 8MP, since Titan X is only 12% faster in math.
Assuming the same level of ALU utilisation between GCN and Maxwell, yes, but that's a pretty big assumption.
GCN seem to have the more complete instruction set, especially integer. Almost all instruction are maximum speed, 1 cycle. How do you propose that Maxwell, with the same shader, occupies less ALU? You think they did go back to VLIW and execute multiple instructions in 1 cycle?
What I meant is that GCN isn't necessarily as good as Maxwell at keeping its ALUs consistently fed.
gamervivek
Regular
Hawaii(290X) is ~40% faster than Tahiti(280X) at 4k in TPU's review of Titan X. At worst they're using an overclocked non-reference cooler 290X which maintains a higher clock speed than 280X's and at best they're using a reference cooler 290X that is throttling below the 1000Mhz clock.
https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan_X/29.html
In the recently released Witcher 3, it's often >50% faster.
I don't see why Fiji couldn't see a 50% improvement at the same clocks compared to 290X once coupled with Tonga's improvements. And are we sure it would support all dx12_1 features like conservative rasterization and ROVs? That might cost a bit of die space too.
https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan_X/29.html
In the recently released Witcher 3, it's often >50% faster.
I don't see why Fiji couldn't see a 50% improvement at the same clocks compared to 290X once coupled with Tonga's improvements. And are we sure it would support all dx12_1 features like conservative rasterization and ROVs? That might cost a bit of die space too.
Love_In_Rio
Veteran
Then Nvidia must be full of magicians.
In R&D you may not always get what you pay for, but you definitely don't get what you didn't pay for. Nvidia has been spending roughly the same amount of money and engineers on hardware and software development for graphics that AMD has for all of its product development. In the same time AMD has canceled about four CPU projects, several GPU developments, lost several teams and had to allocate most of their physical design group on semicustom development. Nvidia's lead is bound to increase further in the future, because they are making significant investments, while AMD is shrinking.
Nvidia is still limited by the same fab process and thus limited by the same laws of physics. Unless AMD goes belly up I don't see performance gap between the two increasing.
Also, what is Nvidia doing with all that R&D money? Last time I checked, AMD was (always?) first to adopt a new fab process. First to have GDDR4, GDDR5 and now HBM.
The only real problem AMD has right now is competing with Nvidia at throwing $ at game devs.