http://wccftech.com/amd-vega-10-4096-stream-processors/
>GCN can't scale, max 4096 shaders only on Vega, same as Fiji
IT'S OVER, AMD IS FINISHED & BANKRUPT
>>53703915
So it's a Fury X without the water cooler?
>>53703915
and what if the shaders are hotclocked?
AMD has been trash for a while now.
I don't want them to bankrupt though because competition is good for the industry
>>53703948
probably a more efficient fury x with 8gb.
forever overshadowed by the 980 ti.
>>53704047
>amd in charge of providing competition
Uh huh, sure.
Anyone who truly cares about competition wants AMD dead ASAP because that's what needs to happen so they can be bought out by people who actually know what the fuck they are doing and have an actual R&D budget.
Only fanboys want this POS of a company to be kept on literal pity like it is being kept currently.
If amd keeps around, intel and nvidia will keep not having to even try to do anything because the only thing AMD is currently useful for is serve as pretend-competition so that Intel and Nvidia don't get striked by anti-monopoly laws. It's a glorified convenient shield and nothing more.
>>53703915
>only 4096 shaders
WHY?
>>53704618
Because they couldn't keep the MOAR COARS strategy up forever.
It's finally over.
>>53704618
>two generations newer arch
>trying to compare the number of ALUs when you know literally nothing about the arch
Go back to facebook/reddit/9gag
>>53704047
>AMD dies
>Intel and Nvidia get split up
>Now there's actual competition
No more shitty curry tech competitors, anon
>>53703915
MOAR ASYNC COMPUTE CORES
Vega: 4096 cores + 4096 ACEs
Full Async Compute Powah!
LE LE LE LE LE LE LE LE
>>53703915
> Fiji: 4096 shaders on a 8.9B transistor GPU
> Vega: 4096 shaders on a 15B-18B transistor GPU?
curryniggers can't even keep their speculations consistent.
the 4096 shader chip either isn't Vega or the architecture has change radically and these are all full fp64 ALUs or something crazy.
>>53706919
>all full fp64 ALUs
hot damn, imagine that
>>53703915
It's a leak from AMD's employee, it's not speculation at all
GCN is crap architecture that can't scale beyond 4096 cores
no fucks given, will continue to game at 1080p even longer now.
>>53707010
Remember the 200 W SoC leak from an employee?
Yeah, that it has been in the drawing board doesn't mean it's the actual product
>>53707010
GCN isn't one architecture.
You're a tech illiterate retard, and have absolutely no idea how GPUs work. There is no limit whatsoever to the number of compute units any GPU arch can incorporate on a single die. Massively parallel processors do not have a finite limit to throughput scaling by their very nature.
>>53707048
Fiji can't scale beyond 4096
Vega can't scale beyond 4096
It's crap, period
>>53707010
if the IP v8 was Hawaii, isn't the lead chip of the first generation of IP v9 the Polaris 10?
could the curry niggers really be making so simple a mistake on this?
>>53707076
>It's crap, period
As to be expected from AMD.
>>53707076
That's like saying that the 2500k is crap, period because it can't scale beyond 4 cores, same as the Q6600 and less than the Phenom IIs and FX8350s.
>>53707010
so WHY can't they go beyond that amount?
what did they fuck up so badly that the numbers just can't go beyond that?
are we really looking at some 16 vs 32 vs 64 bit sort of file size limitations or some shit?
even those had workarounds.
You think they just kind of hit a wall? That this is literally the best they can do?
>>53707120
On the little table that CurryTech is using, "GFX" is actually referring to the driver class. An AMD employee on S|A clarified the whole matter.
As always you shouldn't fall for the clickbait generated by street shitting scum.
>>53707236
Don't take shitposting children seriously.
If the process, power envelope, and thermal margin allowed, both AMD and Nvidia could just endlessly scale GCN and Maxwell. Continuing to make no changes to the arch, just adding more units, for as long as bandwidth could feed them.
GPUs do not have any set structural limits. There isn't even any limitation in obscure bite code in the geometry or command processors in the front end that would limit the number of ALUs they could address.
>>53703915
>wccftech
Oh yeah the 380 is going to have 8GB HBM lol
>>53703915
Did you miss the part where it says HBM2?
>>53704047
AMD has been the better choice in the mid range GPU market for a while now
>>53707310
Here:
Posted by bridgman
Couple of things:
1. if you look in gfx_v8_0.c you'll see a lot of chip-specific logic within the GFX 8.0 IP block handler framework, although it mostly affects things like golden register setup
2. AFAICS for this cycle we focused on optimizing HW performance without changing the software-visible behavior more than necessary. If we hadn't just finished restructuring all of our drivers around block revision handlers I guess we might have called this GFX 9, but starting a new set of IP block handlers would have meant a lot of wasted work and running GFX 9 using GFX 8 block handlers would have been a bit wierd too.
It's probably fair to say that the new driver architecture ends up pushing back a bit on how new HW generations are named... IOW we only bump the GFX version number if the programming model is significantly different rather than making the decision based only on internal implementation.
>>53707357
And yet NVidia outsells them 4 to 1.
>>53707357
They been dominating the low and mid end since the HD 5000 series, with a few exceptions for niche cases or power efficiency
>>53707398
That's Apple-like fanbase 4u
>>53707310
>>53707363
good to know.
>>53707236
because that's what the chip is designed as. That's like asking why there's only 2048 shaders on Tahiti.
GCN at its core is just a 16 wide vector SIMD cluster. It can scale arbitrarily provided enough cooling and power.
>but why not 65536 shaders!!!
Because that requires a massive die, and typically yields on a die decrease exponentially as size increases. Wafer costs are fixed. Anyone thinking they were going to build a 600mm2 chip right out the gate is an idiot. It'll be around 400~mm2 max. As wafer costs decrease and yields increase as the process is refined, they'll be able to make bigger chips.
>>53703915
which means amd is investing in quality shaders rather than quantity shaders right?
it's why nvidia has less cudacores but equilant performance
>>53707468
>AMD
>quality
considering their strategy has been "MOAR COARS" for a while now, not likely
>>53707310
>"GFX" is actually referring to the driver class. An AMD employee on S|A clarified the whole matter.
did he happen to clarify what the other rows meant as well?
>>53707482
Quite likely with Keller making Zen and taking lead of the CPU division, along with Koduri doing the same with the GPU division
>>53707491
UVD is their SIP video decoder, VCE is their SIP video enconder
>>53703915
>GCN can't scale, max 4096 shaders only on Vega, same as Fiji
>same number of shaders, vastly smaller die and power consumption
>big dies not expected til next year
No, they're just following in Nvidia's footsteps and releasing a full range of small dies before releasing the real performers.
>>53707468
nah if you equalize clocks, Nvidia's Maxwell cores are at best 5% more efficient per clock. Advantage probably will go away in DX12.
Kepler was actually more inefficient than GCN
Tesla/Fermi had hotclocked shaders, so were more efficient, sorta, than GCN. If you downclocked the shaders they'd shit the bed performance wise.
VLIW5/4 were horribly inefficient per core but it didn't matter because the cores were hilariously small. GCN is something like 50% more efficient core per clock than VLIW for example.
>>53707491
>>53707501
poo-in-loo article actually has a key:
> GMC – Graphics Memory Controller
> IH – Interrupt Handler
> SMC– System Management Controller
> DCE – Display Controller Engine
> GFX – Graphics IP Revision
> SDMA – System Direct Memory Access
> UVD – Unified Video Decoder
> VCE – Video Coding Engine
the problem is that some of the things I either don't immediately understand the significance of (IH/SMC/SDMA) or just seems waaay broad (GFX).
things from other block diagrams that are marked as changed don't appear to necessarily be represented (geometry processor, L2, CUs, GCP)
>>53704253
$0.05 has been credited to you account
keep up the good work Pajeet
>>53707649
Nice rebuttal, amd fanboy.
>>53707619
All of those things, are designations for the drivers and how they handle the hardware.
They are not listing individual hardware changes.
Fiji, Carrizo's IGP, and Tonga are all technically Volcanic Islands IP, third generation GCN. Between them however they are all very different in what is actually on die.
Polaris is 4th generation GCN, the CU themselves are totally different from every other iteration of the GCN family, but you see all of the above listed under "8.0"
AMD's driver stack sees them as being the same. Thats all it means.
The other things like memory controller, video decoder/encoders are much simpler to change, and are updated more often as features are added more frequently.
>>53707357
>AMD has been the better choice in the mid range GPU market for a while now
Hell, they have been the better choice since the Radeon 9600, and that was what, 13 years ago? Only the GeForce 8600GT managed to beat them in price/performance in the mid-range.
>>53707714
preach
>tfw shitposting on /g/
>>53703915
>wccftech
pajeet pls
keep the clickbait in youtube
>Rumor
GTX 700 series
>2880:240:48
GTX 900 series
>3072:192:96
hurdurr Maxwell and Kepler are same because not much chance in numbers!
Retarded halfwit.
>>53711314
>change*
What's the problem? The ALUs are probably notably larger to compensate for the no increase in count, this is good, meaning the architecture is completely different than what we had until now.
AMDKEKS ON SUICIDE WATCH
NO HBM2
NO GDDR5X
256BIT BUS KEK
ONLY 4096 SHADERS
STAGNATING AND DEAD
0% MARKETSHARE
BTFO BTFO BTFO BTFO BTFO
>http://wccftech.com/async-compute-boosted-hitmans-performance-510-amd-cards-devs-super-hard-tune/
MUUUUH ASYNCH-SHADERS
AMD-fans are literally retards.
>>53707010
>GCN
>crap architecture
>>53711411
>>53711314
>>53710104
>>53708193
>amcuck damage control
>>53711455
It's crap.
>"Most of the performance gains in AMDs case are due to CPU driver head reductions."
>>53711314
780 Ti and 980 performed roughly equal on release.
2048:128:64 vs 2880:240:48
1126/1216/1266 vs 876/928 /1019
On shaders, GK110 has 40% more however GM204 is in general clocked 30% higher; this shows in a general advantage in compute scenarios. The GK110 also has an advantage in the TMU's, having 25% more than even GM200, though Nvidia apparently modified those quite a bit. The real difference is in the ROP's, GK110 was a horrendously ROP limited chip, something GK104 didn't suffer as much under, this lead to the scaling issues in games that allowed GCN cards (namely Hawaii) to outperform it (once drivers were refined) despite hilariously lower specs. GM204 has 33% more ROP's in addition to the 30% higher clocks. That's the reason for a large portion of the performance discrepancy in newer games.
Basically bad design for the GK110, whoever bought that really bought a lemon. Fiji suffers from the same issue but less so, it's the reason it's not scaling properly. It just can't output enough pixels. The design is pretty much 2x Tonga with HBM.
>>53711483
>780 Ti and 980
You should be comparing the Ti's vs Ti's, their shader/TMU/ROP numbers are much closer to each other.
GM204 is a much smaller chip than GK110, also it's much more power efficient.
>>53711472
It's a 16 wide vector SIMD, Nvidia uses a 32 wide scalar SIMD
>>53711529
AMD OWNED AGAIN
>>53711529
I thought AMD was superscalar
ITT: Techilliterate gaymen are shouting about how much they hate technology
>>53711517
It's a difference in types of loads it's designed for mate; along with clock speeds. GK110 is still FAR FAR superior to even GM200 for compute purposes, hell for FP64 an old GF110 outperforms the GM200. There's a reason GK110 is still in production for workstations.
It honestly reminds me of what happened with R580, where AMD broke from the tradition of putting equal numbers of pixel shaders, TMUs, and ROPs; by quadrupling the pixel shader count but keeping everything else the same as R520. This was the result.
>>53711595
ITT: AMDcucks getting murdered by superior Nvidia intelligence and architecture
>>53711602
>1920x1400
Those were the times
>>53711613
yeah that was a weird resolution, some CRTs ran it
>>53711602
The only reason Maxwell has shit FP64 because there's no room for FP64 units, they're not part of the CU's like AMD does, they can be bolted on at will in Nvidia's case, assuming enough die space to spare.
>>53703915
WCCFTech should be bannable. Not even kidding.
>>53711655
yeah and that was Nvidia's choice
>>53703915
Nah if they keep at it with good cards like the 380/x in the midrange and fury against the x80 card from nvidia they should do alright.
>>53711659
Also this
>>53711655
While this is nice for power consumption on mobile and consumer chips, for professional markets it requires a new die or two, which is a massive cost AMD can't afford to make, so their choice is powering off the FP64 units instead of removing them.
>>53703915
>GCN can't scale
The source doesn't say that
>4096 shaders
>Vega
The source doesn't say that
>Implying more shaders is the only way to increase graphics performance
Tech illiteracy should be bannable on /g/
>tfw don't know enough about GPU tech to understand what the fuck that even means
I just came here to shitpost and spam dank memes about house fires and wood screws
>>53704126
Lol fury x is better you cuck
>>53711664
this right here.
the amount of people who're willing to spend 600+ bucks on a GPU is tiny even if it feels different on /g/ where we have a high concentration of benchmark autists.
even 1200$ prebuild gaming systems use GTX 960's or 970's at best so it's this price range that matters most.
the only thing AMD truly sucks at is sucking.....the balls of prebuild manufacturers and ofc the whole marketing shit.
>>53711935
'no'
>>53703915
>2.3x times denser logic than on 28nm
>same amount of shaders
This doesn't sound suspicious to you at all? Wouldn't you think the shaders are now much larger?
Kill yourself, subhuman.
Question for gurus, maybe AMD has measured that a number of applications have problems utilizing more than 4096 ALUs efficiently and are deciding on making them individually more powerful instead of increasing their count?
>>53711983
They just need to continue what they're doing with midrange and they should be good.
Just like with their fx processors, never understood the hate for those. Amazing multithread performance compared to Intel.
>>53712100
Unlikely that they aren't utilized, the scheduler itself decides over how much shaders will the workload be spread over, be it 20000 or 4000, it would make little difference.
What AMD decided to most likely do is not keep the 1 ROP and 3 TMUs per 64 ALU, but increase the ROP and TMU count per 64 ALU.
Which would no doubt increase the CU size quite a lot.
>>53712100
No. AMD's long term target is actual moar cores with multiple smaller GPU dies replacing a single big GPU.
OP is basically full of shit. That 4096 shader chip is either Polaris 10 or part of an APU. It's extremely unlikely to be Vega.
>>53712206
We're a long way off from multiple chips on a single interposer, that's not happening.
What's so weird about keeping the shader count? Why does it matter? The transistor count goes up by double, what's important is how much transistors are there in a single CU
Everyone ITT defending AMD but nobody has a good explanation why keeping the same shaders is a good thing.
tl;dr it's not, you're a fucking fanboy
>>53713021
Keeping the same amount of shaders for now is a smarter move since the kinks in 14nm haven't been worked out as much as 28nm.
Since they're not increasing the number of shaders on die, this has a two-fold benefit:
1. Dies become much smaller, about 2-3x smaller depending on how much more dense the transistors are
2. Smaller dies are vastly cheaper and have a higher retention rate over large dies. You get to pack more dies onto the same wafer, and get more working dies out of it.
So, since AMD is getting these benefits, they just need to work on 2 things: individual core performance (for probably a small uplift in performance desu) and the power consumption, which moving to a smaller node will help with anyways.
They'll likely release much larger dies at a later date when 14nm is known better.
Genuine question here.
How can AMD compete with no R&D money? Didn't they sell most of that shit? Do they plan to get grants? Technology i.e. more patents from Intel?
>>53714607
They are hoping ZEN and Polaris are going to be big hits, AMD is knees deep in debts.
They can't really afford to make a new architecture like Nvidia does with almost every generation.
>>53714682
>>53714607
The good thing for them is that we just happen to be at the point where both Intel and Nvidia can't afford to make new architectures as fast as they used to.
>>53714819
Maybe Intel because they are stuck with 14nm, but Nvidia has the money and still has to move to true 14nm same with AMD but they don't have the money.
>>53704253
>I want a monopoly
AMD is still going MOAR COARS
NVIDIA can't into Asynch Compute
...Looks like we're waiting until 2017 for new GPUs then?
>>53714927
>AMD is still going MOAR COARS
>OMG! COMPANY IS MAKING MORE CORES FOR HIGHLY PARALLELIZABLE TASKS
>>53714927
Old meme from when software hadn't caught up.
>>53703915
So Nvidia and AMD new cards are shit?
>>53714927
dude i can't wait I'm playing games from 2012 on a 570...
>>53714861
Nvidia and their Gameworks API sure did a good job of ruining their image.
>>53715080
Sure, if you have an AMD card.
>>53712068
even 390x is often better lmao
>>53715115
I own a Kepler card and Nvidia just told me to buy a new GPU every year or gtfo.
>>53715178
That what you get for being a retard and not waiting for Maxwell.
>buy now idiots BFTO
>>53715165
>being this much of a retard
https://www.youtube.com/watch?v=pDxTcvZ4T2k
>>53703915
>wccftech
lol
Asynch Shaders are not worth the trouble for programmers.
It's too hard to programm them and integrate it into the whole package.
Anyways, AMD better release a card with the performance of a R9 390X with the power consumption of the GTX980 for less than 300€.
Anything else and it will be a major dissapointment.
>>53715243
Lmao butthurt nvidiot is butthurt!
>>53715339
>people actually believe this
kek
protip: shaders are already async, always have been. That's how GPUs work.
So should I just go ahead and just upgrade my 7950 or should I still wait?
>>53715354
>http://wccftech.com/async-compute-boosted-hitmans-performance-510-amd-cards-devs-super-hard-tune/
>Devs Say It’s “Super Hard” to Tune
Fuck off
>>53715342
>this damage control
>>53715404
>currytech says india is retarded
WOW!
>>53715178
Designing an NVIDIA graphics card costs about a quarter billion dollars. Planned obsolescence is a thing.
>>53703915
Why the fuck do people post wccftech articles? Do you nvidiots not know they're fake? Nice shilling fags.
>>53715432
Amd's naming scheme is based on time and not some performance thing like nvidia.
Polaris 10 came first, then 11.
So in theory vegas 10 is almost a die shrink of fiji to get the hbm memory controller down to scale. Leaving the vegas11 chip we know about to be a higher tier, again 14nm yields being a factor in all of this.
Intel supports Feminist Frequency. Just saiyan.
>>53711911
You're a few years late m8
>>53704253
>If amd keeps around, intel and nvidia will keep not having to even try to do anything
>serve pretend-competition
>pretend-competition
/g/'s version of economics
>>53703915
AMD esta acabado y en bancarrota!!
>>53716033
>drakeposting
pls no
>>53716033
Pretend is better than full-on rapefest on our wallets. Get redpilled.
Anyone know if the Athlon X4 880k or the A10-7890K will be any decent for PS2/Dolphin emulation?
>>53704618
>more efficient shader ISA and design
>larger, more powerful geometry unit/tessellator
>enhanced (probably larger as well as faster) L2 cache
>less shaders needed
>can clock significantly higher
>etc.etc.
Nobody complains that Nvidia outperforms AMD using less shader ALUs...
>>53716565
Considering that top binned 4.5Ghz Nehalem is roughly a little faster at PS2 emulation than top binned 5-5.2Ghz Piledriver, and 4.5Ghz Nehalem still has issues with PS2, I'd say no.
>>53712158
because a fuck ton of shit relies on single core still.
where preformance matters, intel beats out amd in multicore, and where single core matters, intel curb stomps amd.
if mantle/vulkan/dx12 came out around the launch of the bulldozer line, it may be completely different.
>>53712074
i would personally think, with no more info then the die shrink, that on the higher end cards, amd is forgoing the increased power efficacy and is instead going full performance because the data sheet said something along the lines of 40% more powerful just due to the process shrink alone or 30% less power for the same performance.
>>53704253
if amd gets bought by someone, then the they WILL NOT sell chips anymore, they will sell all on one solutions.
>>53705178
nvidia has competition by intel, because nvidia is the VAST minority on the graphics market share.
intel can realistically say their competition is arm and ibm
shit would not get broken up.
>>53718184
>ibm
Not since IBM ragequit most of their fabs.
>>53718384
they would still be a competitor like apple was to Microsoft when Microsoft had to bail them out, and with our current government who can barely keep timewarner Comcast from merging, do you really want to trust they will break up intel/nvidia?
>Sitting on Hawaii beach for the next 3-4 years until I can go to space
Not so bad,if AMD only like making small incremental upgrades to current performance fine but don't except people to jump on board when they are comfy.
>>53718606
Those who bought hawaii early on are left laughing - it looks like its going to usurp tahiti as the most future proof gpu ever made.
Honourable mention: 8800GT.
>>53715784
Who cares. I'm not going to purposely by considerably worse hardware just because of some stupid agenda. Most Walmart products give profits to jewish companies but I'm not going to suddenly stop buying my bag of sweets because of that.
>>53712225
not a gpu, but amd is doing this with their 32core zen cpu.
>>53713152
not to mention, even if they just die shrink with no changes (besides to get the fucker to run) they can be up to 40% faster just due to the shrink alone and depending on how much power efficacy you forego.
>>53714682
correct me if i'm wrong, ubt isn't the cpu side the one thats dragging amd down and if it was just the gpu, they would have profits?
but on the zen, amd has written off the laptop and pc market as non profit areas, they are trying to get back into servers, where even a 10% market share would pay off their debts, and if zen is good, is a realistic goal.
>>53703948
>>53703915
"Which sounds identical to the R9 Fury X but are far from it. The 4th generation GCN cores are ~50-55% faster than their predecessors"
maybe actually read the article idk
>>53719167
cont
"
Considering that the 14LPP process clocks considerably higher than 28nm we’re looking at upwards of 60-70% performance improvement with Vega 10/ Greenland over the R9 Fury X."
>>53718692
>>53715784
if i have the option realisticly i take the one that makes me feel better.
amd gpus are good enough to be competitive, their cpus aren't right now.
good thing outside of games and pro software, nothing requires more then a dual core to run smooth, and if you have any quad core you are set...
granted, my quad bottlenecks games around 50fps, and because its slow it drags my gpu down a bit too... but i can deal with it, at least till early 2017, where i will have a new computer weather it be intel or amd cpu is all on amd, zen JUST needs to be as powerful as sandy bridge and under 500$ for an 8 core and ill take it, and more or cheaper is just icing on the cake.
>>53716841
ps2 emulation requires a 100% rewrite of the core code to do it, but the only person who knew the code good enough to do it left the project.
Lmao, it's 2-3 years ago all over again. Then the benchmarks that shit on nvidia will arrive and everyone's gonna be posting on the NVIDIA DEAD thread.
/g/ never change.
*sigh*
http://vocaroo.com/i/s1o1SLVjM02v
>>53720339
actually, on nvidias part, they foresaw 20nm not happening so they instead made process level optimizations for 28nm, amd did not do that, it largely why there was such a huge difference between 7XX and 9XX
>>53720569
Fury makes much more sense when you consider it was originally a 20nm part. Hawaii too to a lesser extent (given AMD's historical lead on node shrinks). I would wager maxwell is literally pascal lite to keep power draw sane while on 28nm.
>>53720613
no, Maxwell was its own thing, it looses much of the optimization if the node changes, and things were implemented in such a way to take advantage of the quarks of the 28nm process. on a different node, it may be still slightly better then the 700 line, but not by much.
though everyone better hope it wasn't pascal lite, otherwise nvidia is in for a world og hurt when it comes to dx12 and vulkan.
>>53720665
> it wasn't pascal lite, otherwise nvidia is in for a world og hurt when it comes to dx12 and vulkan.
Given For a while Nvidia had Volta planned as the next big leap I wouldn't pin your hopes on pascal.
What if the cores are even more wide?
>>53720613
>>53720665
the progression is basically:
> Fermi - strong GPGPU but hot as fuck
> Kepler - stripped out dynamic scheduling, shader granularity, and most fp64 to save power
> Maxwell - stripped out virtually all the rest of fp64 capacity to save even more power
Pascal is rumored to be bringing back fp64 performance, but this would only make sense for the highest end chips that would be sold in professional/enterprise/HPC versions.
>>53720813
Given most of Nvidia's workstation cards are still kepler based (and the inroads the hawaii based firepro cards have made) it would make sense for FP64 to come back with a vengeance in pascal.
>>53720866
Hawaii and Fiji are native 1/2 dp performance, Maxwell is 1/32.
Maxwell was *never* intended as a professional/HPC GPU, just for gaymen cards while Kepler held the line in that market segment.
Things would presumably have been much different is TSMC 20nm wasn't a complete fucking flop.
>>53721052
Hawaii is 1/4 and fiji 1/32 iirc. Its why tahiti still reigns supreme on the AMD side for double precision.
>>53703915
You realise that that those 4096 on Fiji were overkill, right?
Even with Async-shaders they're still not always entirely fed work.
With the new architecture, they actually have dedicated hardware to better feed work to all the units.
As shown by Fury VS Fury X, they've more than enough CU's to go around, what they need to is to work on the weakest parts of the pipeline, which is elsewhere.
>>53721236
Tahiti is natively 1/4, Hawaii/Grenada are 1/2.
The Grenada XT-based FirePro S9170/W9100 are around 5.2 sp TFLOPS, 2.6 dp TFLOPS, where Tahiti tops at around 3.2TFLOPS/800GFLOPS.
I'm not actually sure about Fiji, nor can I explain how the 290X is beating the w9100 in a fp64 benchmark, given that consumer Hawaii should be gimped to 1/8th for dp.
>>53721496
could be the hardware isnt physically disabled and they just installed firepro firmware to trick it into acting with full DP? its the same GPU but the 290X is clocked higher than the firepro hence the grater performance if the DP isnt gimped
>>53721496
>nor can I explain how the 290X is beating the w9100 in a fp64 benchmark, given that consumer Hawaii should be gimped to 1/8th for dp.
Higher clocks perhaps? Especially given the notorious throttling of the 290x reference cooler.
>>53721541
>>53721553
clock and looser throttling would clearly help, but it's hard to believe that AMD wouldn't fuse of the unlocked fp64 capacity in a way that a firmware flash couldn't touch.
FirePros cost thousands of dollars, and a 390X is like $400.
>>53703915
>>53721471
>...the number of compute units has been reduced to 56, thus 3584 shader processors and that shows in performance. However with its 64 ROPs in tact overall these cuts do not see to make a huge difference...
https://www.guru3d.com/articles_pages/asus_radeon_r9_fury_strix_review,37.html
The fact that disabling those compute units changed the performance by a smaller margin than the % of total units disabled indicate that improvements can found by focusing on the other parts of the GPU, which is what they've been doing AFAIK.
Remember that a GPU isn't JUST it's shader/compute processors, the most important work is still being done by things like the ROP, et al.
>>53721663
>FirePros cost thousands of dollars, and a 390X is like $400.
Driver validation, ECC memory and support make up for the huge cost of those cards. Thats why Nvidia makes such a killing on quadros - the silicon is the same (more or less) as GTX cards but its the extra shit they charge an arm and a leg for.
A firepro/quadro has to run perfectly 100% of the time. GTX/R9 cards are often left to crash their drivers when shit goes wrong.
>>53721713
Its a shame that a combination of architecture and 28nm didn't let AMD scale fiji to 96 (or 128) ROPS - such a card would massacre the 980ti and most likely the titan x.
>>53721797
I think the kind of ROP and TMU' for Polaris is a brand new one, if I remember the interview videos correctly.
>>53721844
Its the new command processor i'd focus on - that will fix a lot of AMD's issues with DX11 overhead as well as tessellation culling (the primitive discord accelerator).
GCN4 is looking to be a beast no matter whats thrown at it.
>>53719220
In any case it's hardly useful to program for anything more than 3 cores of usage - and that last core would be running a very low intensity thread.
AMDs module design with the gimped FPU units will never work well for PS2
>>53721748
>Driver validation
yes
>support
true
>ECC memory
bwahaha
professional/enterprise GPUs don't use ECC memory, they just "support" is at the datastructure level, which translates to nobody uses it.
system DDR DIMMs have 72 bit interfaces (only 64 used for non-ECC), populated by 9 chips with 8 bit wide interfaces.
GDDR chips are only 32b wide, where you'd need something like pairs of 36b chips for ECC, but no GPU or board manufacturer wants to support all that extra complexity for a smaller market segment that doesn't seem to actually truly care all that much.
Is it bad that my retirement money is all in amd?
>>53721663
>>53721748
it use to be you own a pro card, you got a pro driver, this driver and your card, in specifically tested systems would guarantee no crashes and if you did run into a crash, you got same day dev time and new custom for you drivers.
that's where the cost goes for these cards, you are paying for support you hope you never need.
>>53703915
I will literally celebrate when AMD dies.
>>53722000
AMD looks to surge upwards so you might make a killing. As ever strocks are voltaile (see /biz/).
>>53722005
True. The tech support is where the money goes. If shit goes wrong you can (and no doubt will) be on the phone to AMD/Nviodia demanding some sort of solution within 24hrs.
In a wider context thats where a huge amount of tech companies make their money - offering on demand support for their products in the event of failure. Random example: ubuntu - the money canonical makes (as limited as it is) is based upon tech support rather than directly selling its shit.
>>53703915
>wccftech
>>53719183
SWEETTTT
>>53722206
EAT THE EGGS RICCOLA
>>53704253
>AMD dying ever
>>53722000
Definitely a bad decision. You'll probably lose most of your money, but it looks like you just started saving and if you don't have too many liabilities it isn't a big deal if you lose most of your money
>>53718679
Tahiti will also gain a nice boost in dx12 you dip.
7970 will overtake a gtx 970. It's already pretty close
>>53722778
>Tahiti will also gain a nice boost in dx12 you dip.
I dunno man, only two ACE leaves limited room for gains compared to the 8 that hawaii, tonga and fiji have.
In practical terms (ignoring the unique nature of the HWS in GCN 1.2) its borderline impossible to saturate GCN with compute tasks while keeping a sensinsible frametime. By the time you do saturate the rest of the architecture can't keep up and thus frametimes rocket.
>>53722899
aren't ACEs literally completely unused in games w/o async?
they seem like mostly a GPGPU thing with a slight potential advantage for games.
>>53722778
thats like saying the 280x will take the 970,which current hitman dx12 benches says it doesn't.
>>53723005
>aren't ACEs literally completely unused in games w/o async?
More or less.
>they seem like mostly a GPGPU thing with a slight potential advantage for games.
Its not a slight potential for games - its a huge potential. ACE in a nutshell let you coverup stalls with more compute work. The simplest comparison (though not entirely telling the full story) is hyperthreading for GPU's. Async compute is a seriuous game changer if leveraged as it allowed for the gpu to always be one step ahead of the cpu - the ideal scenario for graphics rendering. In a perfect world a gpu will always be pegged at 100% while a cpu is as close to idle as possible since a gpu cannot do shit unless it gets the nod from the cpu.
Async compute is entirely about achieving maximum gpu utilisation without any additional cpu overhead.
>>53723107
though every AMD gpu gets gains not enough to over take another tier unless at the higher end
>>53722616
people said the same thing about betting against the housing market in 2007/8
>>53703915
SEE YA FAGGOTS
AMD IS THE WORST.
PLAGUES PREBUILTS AND TRICKS PEOPLE WHO DON'T KNOW WHAT THEY'RE DOING INTO BUYING THEIR SHITE PRODUCTS
PLEASE GO BANKRUPT NOW, LET THE CPU MARKET BE BETWEEN INTEL AND NVIDIA
>>53723108
hyperthreading is lite context switching triggered by cache read misses, but I was under the impression that async computer was for using the ALUs more while a shader was waiting on results from fixed function units than for cache stalls.
how effective is juggling dozens of contexts on cache misses when L1/L2 sizes are pretty fucking small?
>>53723184
dont forget to make another thread after this one dies
>>53723224
>how effective is juggling dozens of contexts on cache misses when L1/L2 sizes are pretty fucking small?
Speed yo. GCN takes almost zero hit when context switching as everything is decoupled whereas kepler/maxwell requires and entire cache flush.
>>53722616
>You'll probably lose most of your money
You're wrong, AMD will not dip below $2 any time soon. Personally, I am waiting for it to drop to 2.20 so I can buy 3k worth of shares and sell them after they release their new CPU and GPUs which will both be released in q4 2016.
>>53723440
I expect AMD to hit 4 burgers/share within the next 18 months and continue upwards from there. Right now is the perfect time to invest if one is a gambling anon.
They might tank, but personally i'd bet otherwise.
>>53723440
what if it doesnt drop anywhere close to 2.20
>>53723107
Agreed. AMD cards will get some gains, but not that much. Although I guess it depends how much nvidia decides to gimp the 900 series when pascal comes out.
>>53712074
Less silicon? Polaris probably smaller die sizes
Density = shaders / area
>>53723724
Historically AMD/ATi has used smaller dies than Nvidia and achieved performance parity.
Even now iirc Fury x is the densest die ever made.
>>53703915
Oh look it's this meme again
>>53723763
>Historically AMD/ATi has used smaller dies than Nvidia and achieved performance parity.
this is b8, all of AMD's dies are larger than nvidia's currently, which the exception of fiji which is 600mm2, the same size as the gm200 die. hawaii is ~30% larger than gm204 and tonga is nearly 50% larger than gm206.
>>53722778
>7970 will overtake a gtx 970. It's already pretty close
wut? the 970 destroys the 7970 in every single game, if AMD didn't have shit drivers and fixed the massive CPU overhead in dx11 games cards like the 390 and 390x would be competitive, maybe even better than the 970. instead they end up being close to half the performance without an OCed i7 paired with them.
>>53703915
>Which sounds identical to the R9 Fury X but are far from it. The 4th generation GCN cores are ~50-55% faster than their predecessors according to the SiSoft Sandra entry of Polaris 10 that we spotted recently. Which had a 2304 GCN 4.0 core GPU running at 800mhz beating a GTX Titan black and an R9 290X which has 2816 GCN cores running at 1000Mhz.
>>53720339
Fuck it's that guy who sounds kinda like me again
spooked.
>>53725460
did a post from 2014 somehow find its way in this thread
/v/ here. Why is everyone hate curry WCCFTech so much?
>>53727217
curryniggers that post flase rumors and clickbaity bullshit
its the buzzfeed of tech news sites
>>53726713
nigga you retarded
>>53728166
Isn't HBAO an nvidia technology?
>>53728187
HBAO is compute and bandwidth heavy , it actually runs better on AMD cards.
>>53722206
>You don't have to tell me what happened to you but you do have to eat all the eggs.
>>53728187
>>53728228
what's the high-level overview between SSAO and HBAO?
is the difference in quality really that appreciable?
No one here knows shit about chip construction why do we even bother to pretend like we do?
>>53732306
HBAO is more precise but at the same time looks darker.
DESIGNATED
GRAPHICS
DEVICES
>>53716108
That's not even a red pill for Christ sake get your facts straight
>>53732502
isn't all AO just a post-processing hack approximating the appearance of self-shadowing geometry given a very rough description of ambient lighting?
>>53722010
Why? It would fuck up the market you stupid shill, I perfer Nvidia, but do not stand by shilling.
>First 28nm GPU: February 2012
>First big 28nm GPU: February 2013 (Titan)
>Big consumer 28nm chips: November/October 2013 (780ti/290x)
History repeating itself but /g/ is too 12 year old to remember
>>53734378
it's not a matter of memory, it's just that the bulk of /g/ is too retarded to realize that new semiconductor manufacturing nodes always start with lower overall yields and that chip yields scale inversely with size.
> but, yeah, most here won't remember when Nvidia's 40nm Tesla shrink had initial yields at ~1.7%.
>>53734378
7870 was the best. I hope polaris 11 will be THAT glorious as well.
>>53734529
1.7 is just a meme you dip, their real yield rate (for a full 480) was 15-20% in the beginning.
