Quantcast
[ 3 / biz / cgl / ck / diy / fa / g / ic / jp / lit / sci / tg / vr / vt ] [ index / top / reports / report a bug ] [ 4plebs / archived.moe / rbt ]

/vt/ is now archived.Become a Patron!

/g/ - Technology


View post   

[ Toggle deleted replies ]
File: 114 KB, 590x332, AMD Vega die shot.jpg [View same] [iqdb] [saucenao] [google] [report]
62170558 No.62170558 [Reply] [Original] [archived.moe] [rbt]

"Infinity fabric on Vega is optimized for server. It's a very scalable fabric and you will see consumer optimized versions of it in future."
https://twitter.com/GFXChipTweeter/status/902899280226131968

So Vega rerelease in 2018 with proper performance? Basically everything Raja Koduri says about this product does nothing but worsening the whole situation. More importantly - have early RX 56/64 adopters been screwed and this Tweet is the evidence for it?

>> No.62170655

>>62170558
Early adopters are almost universally screwed every time. It's not an exception.

>> No.62170763

>>62170558
IF appears to be a pretty minor piece of Vega's architecture, sitting only between the graphics engine and the IO and codec blocks and not in between or inside Shader Engine blocks.

Of all the things I'm interested in or concerned about Vega, IF is not one of them.

> SR-IOV
> DSBR granularity and throughput
> drivers that actually work

>> No.62170805

The more you wait™, the better it gets.

>> No.62171096
File: 54 KB, 643x370, Vega_Fine_Wine.png [View same] [iqdb] [saucenao] [google] [report]
62171096

>>62170558
Here is one good tweet by Raja.

Early adopters can take comfort in knowing that they will be getting significant performance boosts in future drivers.

>> No.62171194

>>62171096
Raja even writes like an "open bob" pajeet, wow.

>> No.62171248

>>62170558
I think it means that navi could be almost double the performance of 64 if it is just two 64 glued together.

>> No.62171330

>>62171096
The delta refers to die size differnce compared to Fiji. Raja made that clear on reddit.

https://www.reddit.com/r/Amd/comments/6x08bf/raja_infinity_fabric_and_many_of_the_new_features/dmcbpci/

>> No.62171347

>>62171248
At 600-700W? kill me pls

>> No.62171381

>>62171330
Delta just means difference. He may be talking about perf/area in the tweet or just raw performance

>> No.62171434

>>62171347
Vega can be tuned down under 200 watts at full load and almost the same performance. It would be a 500 watt card right now, of they shrink the die, optimize the architecture and write drivers, ut can be under 300.

>> No.62171533

>>62170763
It's used between the memory controller and the shaders.

>> No.62171548

>>62170558
Server graphics card, hahaha

>> No.62171562

>>62171548
Distributed rendering

>> No.62171611

>>62171548
The first people to get Vega before the public release was some company runs a cloud run game streaming service. Which might also explain why it has virtualization features

>> No.62171851

>>62171611
So we actually might run mutliple VMs off of one GPU in near future? That could make internet cafees and LAN gaming in general more interesting.

>> No.62171936

>>62170558
>More importantly - have early RX 56/64 adopters been screwed and this Tweet is the evidence for it?
they knew what the card could do and they knew how much it cost

there is no reason to expect anything more than that

>> No.62171951

>>62171851
Maybe yes, maybe no. I heard different rumors about that. Some say that there is no SR-IOV on consumer Vega, others say that a max of 2 clients are supported.

>> No.62171988

>>62171936
But mah FineWine (tm)

>> No.62173029

>>62171096
What exactly is connected by IF on Vega?

>> No.62173182

>>62173029
HBM, Diplay engine, video encoding/decoding and some other stuff.
http://radeon.com/_downloads/vega-whitepaper-11.6.17.pdf
Picture is on page 2.

>> No.62173208
File: 181 KB, 1600x900, RX_Vega_56_vs_GTX_1070_32_game_aeverage.jpg [View same] [iqdb] [saucenao] [google] [report]
62173208

>>62170558
Then explain pic related.

>> No.62173282

>>62173208
If everything was working properly Vega56 would be ahead of GTX 1080 custom and Vega64 near/above the 1080ti.
Performance as is is an improvement over previous AMD cards, but at the same time its a fucking disgrace for all of the engineers who put work into this chip.

>> No.62173308

how is vega fe with improved drivers?

>> No.62173318

>>62173308
Still slower than RX64, maybe it has another silicon revision.

>> No.62173382

>>62173318
Vega FE shows Revision 00 in GPU-Z while 56 and 64 are Revision C1

>> No.62173526

>>62173208
No, you explain why 10.5Tflop GPU is only 5% faster than a 6.5Tflop GPU while drawing up to 200W more.

>> No.62173587

>>62173318
Did anyone try flashing it with the RX64 bios? It seems to do some black magic on the RX56 cards for some reason. Then again I'm not sure FE has dual bios switch, noone's willing to brick their 1000€ workstation card.

>> No.62173755

>>62173587
It can certainly lead to some "black screen" magic when the 56's HBM cannot do the high frequency. The same effect can be had by modifying power tables in windows registry as "Gamers Nexus" has shown however.

>> No.62173758

>>62170558
> PRO/PRO SSG/and WX are server cards. Vega was an architecture meant for them. In order to cut cost, they took the same die, threw fiji's GPU pipeline on it and went from there.
> HBCC only makes sense with NVME which only the Pro SSG has.
Otherwise, it's just DMA/IOMMU/Memory management unit.
> FP16 hasn't been enabled yet. It will be gimped on consumer cards just like they gimp all the features.
> RX Vega 64/56 are capable of doing: http://gpuopen.com/compute-product/direct-gma/ but this will be outright disabled

Raja is more than happy to reply on these points but he wont because this si what they're busy doing...
> Ensuring all of these features work flawlessly on Pro SSG/WX cards for September launch
> Gimping the shit out of the features in hardware/software
> Waiting for bad Pro SSG/WX dies to toss onto consumer cards

What you are left with is a power hungry consumer card w/ an oversized server compute die with features and subsystems that are gimped to shit yet consuming fuck tons of power.

RTG is more than happy to comment on this and SR-IOV capabilities but they wont. The train already left the station. So, they either enable all of these features or they can kiss the future goodbye.

>> No.62173823
File: 66 KB, 720x707, 1491601973605.jpg [View same] [iqdb] [saucenao] [google] [report]
62173823

>>62171562
You need this for that :
http://gpuopen.com/compute-product/direct-gma/
They've had this feature for some time but gimp it on Consumer cards. It essentially allows GPUs to read and write to each other's memory.

Nvidia pulls the same shit with GPUDirect w.r.t to gimping/disabling it on consumer cards.

>>62171611
That was a PRO SSG or WX card or their specifically designed SR-IOV virtualization card. It wasn't consumer vega. Vega is a micro-architecture w/ multiple different end products. They keep throwing around 'Vega' as if every card has the same feature sets. It's disingenuous and slimmy.

If they were being straight forward, they would say exactly what FP16 will offer for GPU/Compute. They will say whether SR-IOV will be enabled on consumer vega. They will specific if DirectGMA will be enabled. They will specify what other features will be enabled. They aren't but Raja is busy being a slime-ball using Vega as a product name instead of specifying which Vega GPU...

>> No.62173827

>>62173755
So Vega FE is just fucked compared to V64 and 56, despite it being capable of running the same clocks and double the memory?

>> No.62173910
File: 182 KB, 755x1024, lisa_su.jpg [View same] [iqdb] [saucenao] [google] [report]
62173910

>>62173182
Again, the whitepaper isn't even clear on which vega these features are enabled on. The Pro SSG is not consumer vega 56/64. That's what the video editing standards apply to. Vega 56/64 don't have that feature or its heavily gimped.
It'd be equivalent to Nvidia saying Pascal does x,y,z constantly instead of specifying :
GP100, GPU 102,GPU104,GPU106,GPU107,GPU108.

Ultra scumbag to pull what they're currently doing, speak about your top of the line die cut (i.e GPU100) and claim it represents your consumer cards... They know dam well which features will or wont be on Vega56/vega64. They never even did an official launch of their own product ffs. They threw it in the hands of a bunch of bumbling e-celebs then went on vacation? then Raja is playing victim saying the card wasn't showcased properly.. isn't that your job?

I own several AMD CPUs because they delivered a quality product and were up front about it. I almost feel emberassed owning it giving how RTG is behaving. Lisa needs to clean house of this shit and get straight shooters on the payroll. They're manipulating and lying to their customers while stringing them along and the product is already on shelves. WTF are they waiting for w.r.t to detailing the micro-arch and detailing what's on which card? Come September, their future as an organization will be set in stone.

>> No.62173951

>>62173282
I ran into one on a forum during an exchange. I was drilling RTG on their HBCC on Vega 56/64. He gave some details about it but even then was being disengenious and a sleazeball. He gave certain metrics but left out 3/4ths of the round trip pipeline. Absolutely disgusted by how this product has been delivered and detailed and the persistent lies/deceit being pushed. If nvidia wasn't such a p.o.s that gimps the fuck out of their consumer cards too, this wouldn't be worth spending a single minute discussing.

>> No.62174054
File: 161 KB, 720x540, Screenshot_2017-08-30-20-02-49(1).jpg [View same] [iqdb] [saucenao] [google] [report]
62174054

>>62173526
>200W more
wot?

>> No.62174132
File: 298 KB, 2162x1236, 1477263429243.jpg [View same] [iqdb] [saucenao] [google] [report]
62174132

>>62173208
isn't vega 64 meant to compete with 1080 ti anyway and 56 with 1080? just like the fury and fury x, its like the introductory price is cheaper now and when the proper driver is used they eventually go to that level, the illusion of "fine wine", and thus leaving the 1070 with no proper competition unless the nano or other vega variant is introduced.

>> No.62174140

>>62173910
Honestly I believe the Vega consumer lineup is just such a fuck-up, nobody had the balls to go out on stage and say this is the best they could do. They should've jsut harvested the 56s for RX690 and then maybe depending on the yields like a V32 as a RX680 with only 4GB of HBM2, as they clearly presented that they have their streaming into VRAM tech ready and working, it should perform better than then 580's VRAM anyway.

>> No.62174190

>>62174132
As of now, Vega 56 competes would compete with 1070 if it was availible at MSRP (not saying the 1070 is either, but it's still cheaper) and Vega 64 barely competes with the old 1080. Nothing touches 1080ti, not even close, not even in AMD cherrypicked games. Nvidia literally announced they wont release Volta this year, so they have a few more months on contemplating how much to cut down the consumer chips before Radeons get anywhere near to competetive.

>> No.62174227

>>62174190
Consumer Volta was never meant to be released in 2017 anyway, only the hpc chips are in for 2017 as ive read on an article years ago.

>> No.62174295
File: 73 KB, 706x706, 1446786993767.jpg [View same] [iqdb] [saucenao] [google] [report]
62174295

>>62170558
>rerelease
fucking will I have to then wait for re-release partner cards?

>> No.62174373

>>62174295
Either that or see if you can get some information off reddit when the AIB stuff will launch, then F5 spam the shop you want to get them from. This will get you your personal housefire with 90% certainty.

>> No.62174400
File: 343 KB, 640x877, 1493540777490.jpg [View same] [iqdb] [saucenao] [google] [report]
62174400

>>62174140
They could behave like their organization's future depended on it and decide they want market share and not gimp the pro level features that are available in hardware on the 56/64 and instead blow the doors wide open with whats available in hardware. The professionals will still buy vega FE/WX/PRO SSG because it has 16GB HBM2 and NVME. However, they likely wont and that's what will allow Nvidia to do counter come 2018. Whoever opens up FP16 compute, RDMA, and SR-IOV to consumer level cards first with solid APIs wins.

RTG has the potential to do this but they're going to blow it trying to be greedy. They could gain market share across the board but they'll likely gimp the cards instead while allowing the pro/server features to suck wattage w/o full functionality.

The market is not gaymen cards anymore. Whoever begins respecting this and delevering ungimped compute features to consumers wins the next wave. I'm watching these releases closely as I am analyzing the underlying architectures. Both Nvidia and AMD are gimping to protect their margin cow pro/server/workstation cards.

This is a mistake. Whoever breaks the mold first wins.

>> No.62174704

>>62174400
Thing is, Vega consumer cards are probably being sold at a loss, just to not loose face in the gaming market completely. They WANT people to buy the 1000€ FE and the Radeon Instinct cards, because that's where they actually make SOME money. Nvidia can easily outplay them at the money game, you can just take a look at how stripped the Nvidia cards are, they are most likely MUCH cheaper to produce than Vegas, while beating them in almost every task that supports CUDA acceleration.

>> No.62174713

>>62170558
AMD is like fine whine it is shit when you buy it but you have to play the long game when it will be good in 5 years.

>> No.62174797
File: 47 KB, 369x368, 1304938223524[1].jpg [View same] [iqdb] [saucenao] [google] [report]
62174797

>>62174713
>mfw 1050t
W-who's laughing now, i5 fags?

>> No.62174822

>>62170558
>A single feature on a GPU being optimised for server applications means the whole GPU is optimised for server applications
(You)

>> No.62174834

I wonder if ESXi will support these cards for shared graphics. The only supported cards right now are Quadros, FirePros, and GRID cards.

>> No.62174849

>>62174822
Ofc it is, this is 4chan.

>> No.62174890

>>62173758
>>62173823
>>62173910
I despise you and every thread you shit up with your avatarfagging vague innuendos.

No one in the consumer space gives a fuck about pro features, and attacking RTG for disabling them when Nvidia does even worse gimping of their consume GPUs is the absolute height of genuineness.

What consumers care about is RTG actually fucking implementing the core RX Vega uarch features in the whitepaper in the fucking drivers so the card stops performing like an overclocked Fiji from being totally front end bottlenecked.

>> No.62174964

>>62174890
The question is - will they be able to pull that off though? There has been no word on how much performance they think will be able to squeeze out of this arch with drivers.
Ofc they promise improvements, and likely we will see another 15% as with 7000series cards or Hawaii, but I have a feeling that the chip was built for much more and they just won't be able to unlock that extra 30%.

>> No.62174968

>>62174797
>complaining about having a 6 core processor

>> No.62174981

>>62174890
Bottleneck has been somewhat improved however, as clock for clock 1080p performs better than on Fiji.

>> No.62174994
File: 25 KB, 641x530, 1494912778105.jpg [View same] [iqdb] [saucenao] [google] [report]
62174994

>>62174704
More reason to not gimp them and attack the compute market with fury. $1000 FE/Radeon instinct cards aren't going to cut it. They either say they want to dominate compute from the top down and differentiate by more advanced hardware features or they're going to get BTFO. Both Nvidia and AMD gimp the shit out of Consumer cards and are holding back progress. RDMA has been around since forever and it has yet to hit the consumer market. RDMA is possible in the HBCC. So, they're literally disabling it in firmware/software. The same goes for Nvidia. They can either break the mold themselves or someone is going to break it for them.

> you can just take a look at how stripped the Nvidia cards are
RTG does the same stripping my friend.
Nvidia and RTG :
> Disables RDMA on consumer cards.
> Gimp SR-IOV or break it on consumer cards
> Gimp FP16

Instead, they could just unlock the features, see what comes of it and focus on new architectural elements to improve value. Instead, year over year they've held the market hostage by gimping/disabling features to create feature disparity. They both do the same thing. Instead, RTG has taken it one step further and has included a power hungry compute pipeline in consumer cards that they've gimped.
Vega FE comes w/ a fatter memory pipe/bandwidth and double the HBM2. It will sell simply based on that. There's no reason to gimp vega 56/vega 64. The hardware features alone will sell the more expensive cards.
> Nvidia can easily outplay them at the money game, you can just take a look at how stripped the Nvidia cards are, they are most likely MUCH cheaper to produce than Vegas, while beating them in almost every task that supports CUDA acceleration.
You're 100% correct.Thus, if RTG gimps or fucks up this one chance they have to solidify market share in compute, they're toast for the future.

>> No.62175019

It's not like 56 or 64 performance is awful. If you're upgrading from an R9 290 or whatever you're going to be pretty happy.

The price is messed up right now obviously, but you don't want a reference cooler anyway.

>> No.62175049

>>62174968
Yeah, but even at 4Ghz I hardly keep up because single-threaded performance is ass. Thank god I got a 1060 and not the 480, AMD still can't into multithreading drawcalls, even in DX12/Vulkan

>> No.62175060
File: 35 KB, 600x513, heaven.png [View same] [iqdb] [saucenao] [google] [report]
62175060

>>62174964
>The question is - will they be able to pull that off though?
Primitive shaders claims to boost culled polygon throughput from 4 triangles per clock to up to 17. Even realizing half of that claimed benefit would mean that Vega's polygon throughput would double. Imagine the implications for tessellation performance! Vega 64 is currently behind a 1070 FE in tessellation heavy benchmarks.

>> No.62175063

>>62175019
Lol, I am upgrading from an R9 290 and ordered an EK Waterblock with RX 56. I was fast enough to get the card at MSRP and I think I made a good deal on that. Still I do want to hope that the cards aren't hitting their full potential yet and that future drivers will "unleash the beast".
Considering that RTG needed 5 months to fix video stuttering in Firefox, chances are that with enough times the cards will get a lot better.

>> No.62175071

>>62174981
It has primitive discard like Polaris, its just that PDA isn't enough to keep 64 CUs fed, which is why Vega 56 is 100% as fast as Vega 64 at the same clocks.

>> No.62175081

>>62175060
Do primitive shaders have to be enabled by game devs or will the driver do it automatically?

>> No.62175099

>>62175063
Considering how both the R 2xx/3xx cards and Polaris have seem significant performance gains post-release, Vega will likely be the same.

>> No.62175131

>>62175081
They will be automatic. Primitive shaders are one particular preset implementation of Vega new geometry engine (NGG fast path) that supposedly features generalised non-compute shaders. So, eventually devs will be able to have custom NGG implementations, but for the near future primitive shaders will provide a default implementation of the NGG fast path that is designed to sidestep GCN's historical front end bottleneck issues.

>> No.62175189

>>62175131
Now that is some useful information, thank you.
Is there some kind of roadmap or list of stuff that isn't yet implemented with a timeframe of when they will be able to do so? I think if RTG would openly communicate about that, it would positively affect the resonance Vega has recieved lately.

>> No.62175246

>>62175131
Worst case scenario though is that RTG themselves don't know if these features are actually working in hardware, as there is no software implementation for them yet. What if they discover a bug in silicone when writing a driver for it? Early adopters will shitstorm so hard that all of the RTG staff will be fired.

>> No.62175272

>>62175189
You can read the Vega whitepaper ( http://radeon.com/_downloads/vega-whitepaper-11.6.17.pdf ) if you want to get an overview of all the new uarch features in Vega.

As for which have already been implemented, we know for sure primitive shaders and the NGG fast path have not been implemented yet. We know that PDA (primitive discard accelerator) as used in Polaris is working, we don't know to what extent Draw Stream Binning Rasterization (DSBR) is working, and we know that AVFS (Advanced Voltage and Frequency Scaling) is only barely working.

One complication here is that without primitive shaders enabled Vega remains so severely front end bottlenecked that it is difficult to evaluate how impactful already enabled features like PDA and (possibly) DSBR are, because Vega is bottlenecked in front of these features in the rendering pipeline. As for timelines, who knows? Most of us that pay attention to this stuff can't believe they threw the cards out the front door on half finished drivers to begin with.

>> No.62175280

>>62175246
they are working anon, RTG only had time to implement them in the professional drivers though

>> No.62175305

>>62174994
That's a potentially very risky move from RTG, they could fuck themselves royally considering the fact that their margins on consumer vega cards are about as thin as the beef patties in Mcdonnalds.

We all know that the original Titan was the reason Nvidia swore to never bring back uncrippled FP64 performance to Geforce cards. It's not like Nvidia can't take the hit of a few people buying a 1000€ card instead of a 5000€ for select few tasks, but AMD can't afford this shit for sure. Nvidia had no issue releasing a hotfix that enabled some features into the Titan XP back and it suddenly matches or outperforms the FE in most tasks that AMD boasted about in their slides, imagine how fucked AMD would be if they just rolled a next patch, unlocking some of the featues even just down to 1060.

>> No.62175307

>>62175272
>Most of us that pay attention to this stuff can't believe they threw the cards out the front door on half finished drivers to begin with.

AMD releasing cards with unfinished drivers is nothing new but this time they are particularily unfinished xD

>> No.62175478

>>62174890
> I despise you and every thread you shit up with your avatarfagging vague innuendos.
If I spoke any more directly, you'd be silenced immediately. The vagueness and the innuendos are for your protectiion.

> No one in the consumer space gives a fuck about pro features
Are you a visionary who wants to push the envelope of computing and your product or are you greedy myopic business unit that is attempting to hold people hostage to the old wave of computing? Have you even asked your customers what they want? You'd think after a decade+ of whoring a feature as PRO you'd develop new pro features and release the old for consumers.

SLI/Cross-fire both utilize an RDMA capable protocol. Why not open it up to compute? Because you're too lazy to create innovative new architectural features to compel the market and justify your continue feature disparity margin faggotry.

> attacking RTG for disabling them when Nvidia does even worse gimping of their consume GPUs is the absolute height of genuineness.
I attacked both. They both pull the same shit. Nvidia just is clear about it. RTG seems to want to play games w/ declaring vega does x,y,z and not declaring which features are available on the pro cards. Pay attention faggot.

> What consumers care about is RTG actually fucking implementing the core RX Vega uarch features in the whitepaper in the fucking drivers so the card stops performing like an overclocked Fiji from being totally front end bottle necked.
Well, they took Vega (a compute first architecture) and slapped a fiji gpu pipeline on it.
Right now, they're busy writing drivers and code for vega compute for their pro card release in September. They don't give a fuck about Vega 56/64 and it shows. Btw, vega 56/64 are nothing more than Pro/WX card dies that didn't make the cut. So, until they begin mass producing pro/wx cards, they have no dies for the Vega 56/64

We both have valid critiques....

>> No.62175494

>>62175305
>It's not like Nvidia can't take the hit of a few people buying a 1000€ card instead of a 5000€
Alot of companies straight up ignored the Quatro lineup because the OG Titan provided the same functionality to them and better performance thanks to higher base clocks. And I mean ALOT. Nvidia will NEVER give us workstation features that could be actually useful to Geforce cards. Even with the Titan XP patch, the card still lacks performance in the same tasks as it's quatro twin.

>> No.62175508

>>62175019
This is why gaymers are not respectable drivers of the market anymore. You sit here tolerating this power inefficient garbage that hasn't improved performance in any notable sense and say : yeah man, time to buy another viydya card. The future of GPUs is compute because they demand and push the market. Meanwhile you shill on here for a $500-800 card and say (it's not awful). Do you go out and buy a CPU that provides marginal performance and say the same thing? wtf man.

>> No.62175529

>>62175280
^this. They're busy making sure they work on PRO/PRO SSG/Workstation pro driver packages. They'll chuck them over to the consumer group in September whereby they'll gimp the shit out of the features for consumer dies.

>> No.62175559

>>62171951
AMD's papers said 16 clients

AMD still hasn't made it clear whether or not SR-IOV is coming down the pipe for Vega or not. It's not enabled on FE or RX.

>> No.62175593

>>62175272
The real question is how come Radeon is so shit at driver developmen It's their achilles' heel ever since AMD acquisition. Just throw more money at it! Over the 10 years a 1000 monkeys with typewriters would do a better job.

>> No.62175648

>>62175305

[1/2]...
They aren't. All companies lie about this and its tied heavily to volume. Margins are shit on many products but, when you move volume, this changes the financials.

Furthermore, lets be real.. They're using bad dies from PRO/PRO SSG/WX cards. So, supply isn't hear yet because they haven't spun up production and haven't gotten enough bad dies from their pro line. Their margins may be shit on having to ship/deliver consumer vega w/o shit dies but, given that they were going to toss the bad pro dies away, there's nothing hurting their margins. Going for $0 to +$ helps margins. They do CPUs the same way.

> FP64 is a meme
Were talking about FP16 which is becoming a standard for gaymen/compute.
If they gimp it for compute on Consumer vega, you can kiss their ass goodbye

> It's not like Nvidia can't take the hit of a few people buying a 1000€ card instead of a 5000€ for select few tasks, but AMD can't afford this shit for sure.
RTG has shit market share. Nvidia has gaymen and compute cornered. RTG can't afford to try to get Nvidia's margins because they're an underdog w/o market share. What they can try to do is ship volume instead and cut their margins and come out on top. Once you've captured the market, then you can start talking about margins. As for $1k card vs $5k card, Vega has 8GB HBM2. WX cards have 16GB HBM2. Double the bit width and mem. That alone differentiates them. PRO SSG has NVME storage on board. Professionals will buy them. The question is, wtf brings consumers to Vega 56/64? 1070/1080 performance? GTFO. Nvidia only need cut the MSRP again and they can.. They already have. You notice the 1080s dropping in price? The fucking is coming once Vega56/64 actually start to ship in mass in October. So, either RTG differentiates and shows why they are worth more than 1070/1080 or they're gonna get blown out and have no budget to realize their roadmap (AGAIN).

>> No.62175667
File: 41 KB, 420x420, 1494911910249.jpg [View same] [iqdb] [saucenao] [google] [report]
62175667

> Nvidia had no issue releasing a hotfix that enabled some features into the Titan XP back and it suddenly matches or outperforms the FE in most tasks that AMD boasted about in their slides, imagine how fucked AMD would be if they just rolled a next patch, unlocking some of the featues even just down to 1060.
Like I said, whoever breaks the mold first wins.... RTG can be idiots here and Nvidia is going to blow them the fuck out (guaranteed)

CUDA 9 is just around the corner ...

>> No.62175841
File: 23 KB, 578x490, 1496266894043.jpg [View same] [iqdb] [saucenao] [google] [report]
62175841

>>62175494
GPUs are now a mature product that are used for general purpose compute. They are CPUs. Intel and AMD CPU divisions have always been mature w.r.t to adhering to x86 and differentiating based on hardware features. Meanwhile, you have GPU companies playing games with GPU pipelines creating arbitrary lines between pro/pro-sumer/consumer. They're setting themselves up for a 3rd party group to cut their asses at the knees by focusing on affordable compute. Google and many other companies are working on TPU modules and intel is furiously pursuing new highly parallelized compute units. ARM is doing the same. Both RTG and Nvidia can keep playing these stupid games or they can blend the two, ship volume, and carve new markets for their products.

Both RTG and Nvidia can get blown the fuck out if they keep playing their respective games. Most of these 'pro features' are enabled by interfacing to 3rd party asics using open standards. There's nothing amazing about RDMA... It's decades old and is being held hostage. If needed, capable people can directly bypass software/firmware/driver locks and target the hardware directly. Bypasses already exist.By centering on ancient standards and penny asics to differentiate one's product, they lock themselves out of consideration for future hardware paradigms. They either (both) cut the shit and start giving these features to consumers in this generation or they open themselves up to get BTFO in the upcoming generation by new entrants to the market. I'm not talking about gaymen cards. I'm talking about going directly at their necks in their high margin compute markets.

So, the choice is there's. Whoever cuts the shit first wins the next generation of compute.. They literally have until the end of the year (max). After which, new entrants are coming for them both.

>> No.62175919

>>62173318
It just throttles

>> No.62175941

>people are responding to the frogposting meme brazilian subhuman
/g/'s gone down the drain

>> No.62175948

>>62175667
no one wants to start that because its a race to the bottom profit wise.

>> No.62175991

>>62170558
Ryzen is also a server optimized CPU, funny how AMD marketed them both as consumer but it infact was the opposite in both cases, almost like they sandbagged to purposely mislead Intel

>> No.62176118
File: 27 KB, 560x315, 1179.jpg [View same] [iqdb] [saucenao] [google] [report]
62176118

>>62175948
If they wont start it, a 3rd party will. The future of compute hasn't been defined nor has its architecture. Certain players are being given an opportunity to show good faith in pushing the envelope of computing from their respective markets and segments. If they don't, another party w/ no interest in preserving the old will push through ever stagnant segment like a disruptive freight train.

All of the features I have commented on are enabled by 3rd party asics and open standards. They have been available for decades. What's simply going to happen is someone is going to package it and target it better and its going to cut all of those stalling at the knees.

>> No.62177383

>>62174400
I can agree with this

I actually have no intent of using vega as anything but a gaming gpu. Nevertheless the SR-IOV is a make-or-break feature for me, because I want to run a windows virtual machine for gaymen on my Linux desktop. If Vega does that for me, it is an instant buy.

So here I am, not even a professional, just regular gaymerfag and I want SR-IOV.

If Nvidia delivers SR-IOV that works on linux, I'll buy them over AMD.

>> No.62177460

>>62175593
No, the real question is why RTG management and design leads would target an architecture that relies so heavily on driver development to get the best performance when their driver teams are historically completely shit at their jobs.

If I knew my teams' weakness was driver development, I would've worked on an architecture that requires less driver development, not more. RTG upper management are a bunch of clueless morons. The only thing saving them their jobs is their experience and connections with other people inside RTG. From a technical competence point of view these people have been failures for almost a decade now.

They need to be let go at some point. RTG is driven by a team of brainless zombies just pulling in paychecks. What happened to the core CPU teams needs to happen to RTG.

>> No.62177462

>>62174994
>why are they intentionally gimping car market?? why they wont give me a new Ferrari for 4000$? ;'(((((
- t. you.

Get a grip on reality. NVidia and AMD need to earn back money spent on R&D. The only way of doing this is selling some of cards at much bigger margins to those willing to pay for pro features.
Without product segmentation you'd paid 2-3x to just MUH GAYM on their cards. But hey, you'd get your FP16 at least! Totally worth it!
Fucking retard.

>> No.62177626

>>62175667
god, you are even worse than photoshop retard from march
if your aim to be insufferable, well done

>> No.62177671

Don't care ordered a 1080 and I'm sick of waiting for stock and driver's

>> No.62177833
File: 1.49 MB, 1688x1916, 1495228655847.png [View same] [iqdb] [saucenao] [google] [report]
62177833

>>62177383
Yep, the tech (protocol/asics) are very old that allow for this. Thus, it's literal greed and a lack of innovation that are causing both Nvidia/RTG to keep nerfing consumer cards from having these features in order to differentiate their pro-lines. Instead they could innovate, create new features while retiring the old to work across all of their GPUs. It's the same shit intel has pulled for years. Thing is, these clowns (Nvidia/RTG) are new to compute. X86 (general purpose) compute dominated for the reason of being open and general and supporting loads of features.

GPU companies aren't going to continue to dominate large scale compute if they keep down this gimping path... Most of the chips and IP that they use to scale their GPUs is all 3rd party tech. Infinity fabric = 3rd party Phy.

So, if they keep pulling this shit, additional parties are going to enter and cut them at the knees. It's happened before and will happen again.

At this juncture, whichever of them matures first and realizes that general purpose compute is a mature market and starts selling prime time features wins. SR-IOV is disabled on consumer cards because these faggots charge insane money to enterprises that scale GPUs for VMs for clients. SR-IOV is not Nvidia/RTG tech. It's third party tech enabled by 3rd party asics which all are usually built by consortium. RDMA, same shit. It's been around since the early 90s .. the asics/ip costs pennies. Yet, they use it to differentiate between pro/consumer.

Literal faggots holding the future of computing hostage to keep their shekel margins in tact. Eventually you get BTFO for doing this. So, i'm not too worried. They can get on board the new wave or drown in the old.

>> No.62177869

>>62177833
Lol nobody is stupid enough to enter the home PC market.

Just let it stagnate and the mobile sphere will encompass it.

>> No.62177896
File: 298 KB, 515x438, 1491255558070.png [View same] [iqdb] [saucenao] [google] [report]
62177896

>>62177462
Ferrari develops their own motors right? That's what made them the company they are. They hand made the parts and components themselves. Ferrari isn't doing so hot as of late btw when a $40k chevy camaro can lap nurburgring better than than their hyped up $400k flagship.

Now, there's a lesson to be learned here.. You can only charge a premium for old tech so long before you get blown the fuck out. Reason being is because the tech eventually becomes ubiquitous and you no longer can differentiate yourself by features that everyone has.

> Get a grip on reality. NVidia and AMD need to earn back money spent on R&D. The only way of doing this is selling some of cards at much bigger margins to those willing to pay for pro features.

The pro features are decades old. It's no longer a pro feature when you're buying $0.50 asics from third parties and glueing it to your board.
> Muh 1000x fold markup for integrating 3rd party IP

Here's where things get spicy : Enterprise companies get tired of paying $10k for your relabeled gaymen cards with $10 worth of 3rd party IP slapped on. So, they spin their own shit and kick your nuts off : http://www.anandtech.com/show/11749/hot-chips-google-tpu-performance-analysis-live-blog-3pm-pt-10pm-utc


> Without product segmentation you'd paid 2-3x to just MUH GAYM on their cards.
Bullshit faggot.

> But hey, you'd get your FP16 at least! Totally worth it! Fucking retard.

You fuck people for long enough and they'll figure out a way to circumvent you or outright destroy your business model. GPU compute is set to be destroyed soon by far more entrants and players than Nvidia/RTG. Either they have true business dev individuals or they don't who see this coming. So, the least they can do after milking the market for all these years, is enable some old features on their GPUs and focus moreso on innovating new ones.

You say nah? I say, k... I spin my own shit.

>> No.62177903

>>62177869
> Just let it stagnate and the mobile sphere will encompass it.
Things aren't headed in that direction my friend...
That cycles over.

>> No.62178015

>>62177896
You dont understand semiconductor business neither economics.
ASICs fuck GPU's feautures ANY WAY, it don't matters if you enabled them on your cheapest GPU line or not. That's what asics are for. Tiny specialised chip, GPU with its shitload of different functions can't compete with them by default. 'Member bitcoin mining? Did CPU manufacturers restricted it by segmentation, turned off SHA1 in consumer chips? Did GPU manufactures? Nope. Did that helped them? Nnnope.
As long as there are no competitors, they will milk the market. That's what EVERY business about. Supply and demand and shiet. If they'll meet competition, they'll cut prices. If they can't compete anymore at all, they'll cut the feauture completely. Cutting prices for feautures right now will give them nothing.

>> No.62178050

>>62177896
>I spin my own shit.
'kay, call me when your 3-rd party company will invent FP16 asics.
There are a reasons GPU chips makers rule the market for such long time uncontested. Inventing, producing and, even more importantly sometimes (see NVidia and CUDA+AI libraries ecosystem), supporting chips are not all unicorns and rainbows.

>> No.62178062

>>62178050
>are a reasons
well, fug XD

>> No.62178077

Good let Nvidia and rtg stagnate and die just like intel

>> No.62178167
File: 133 KB, 333x333, 1496276450114.png [View same] [iqdb] [saucenao] [google] [report]
62178167

>>62178015
Welp, then its time for new competitors to cut them at the knees relentlessly. They luckily happened across compute. It's not their domain nor do they seem to be trying to push the envelope beyond what the market compels them.
So, using the same elements they use, some people will begin making custom boards for their applications in compute w/o any gimping and it will start to roll out to the general market (TPUs). Good luck beating 100TFLOP asics that run at 1/2 the power of your flagship product.

> But muh gaymen card has so many exciting features
Well, all i need are compute calculators. So, GTFO

Infinity fabric = standard Phy
HBCC = standard DMA/IOMMU/Cache
FP16 = Get blown the fuck out by TPU calculators

What happens when you milk this hard and gimp this long is that people look into your product and find out just how its made. They hire the same people you do and they make their own shit for cheaper, using less power, and without gimping.

DESU desu, you don't even need custom hardware anymore. You don't need to spin new shit. It's all right there right now. Which makes the potential for disruption all the more serious...

All that needs to be understood about the semiconductor business is that you operate on open standards for much of your IP. At the end of the day, there are drivers, firmware, and kernel hooks....

You don't want to push the envelope of your hardware's capabilities? Then I'll drag your ass from software. I hope they've enjoyed the margins thus far. All of this rage caused me to look more into specific IP blocks. I've found what I needed.

My work here is done.

>> No.62178256
File: 6 KB, 212x213, 1494976981523.jpg [View same] [iqdb] [saucenao] [google] [report]
62178256

>>62178050
> http://hexus.net/tech/news/industry/105838-google-cloud-tpu-takes-machine-learning-lead-tesla-v100/

No one's asking GPU manufacturers for their complicated ass pipelines for compute.
I don't think you realize how basic these calculation operations are...
The hardware is already being spun :
https://www.servethehome.com/google-tensorflow-asic-impact/

The competitor is already real... and its starting in enterprise.. GPUs bread and butter.

GPU manufacturers crossed into the CPU domain yet are continually trying to play on GPU margins. They played to hard with the wrong individuals. Certain other companies have done the same only to see deep pocket enterprise companies spin their own whitebox hardware and then make the standards and designs open to lower cost.

This is absolutely destroying high margin enterprise shekel seeking companies. HP sold its units for a reason.

The same is about to happen in GPUs targeted towards compute. General purpose computing is won by low margins, large volume, open standards and performance... Not by proprietary bullshit and artificial 'pro' segmentation.

This guys have by the end of the year before they get their shit pushed in IMO.
2018 and beyond is going to see the rise of whitebox hardware spinning out in insane flavors to serve different markets (enterprise/consumer). These fuckers can either get on board or get left behind.

I raged super hard until I saw it for myself. I have zero concerns at this point as I knw what's coming.

>> No.62178258

>>62178167
>then its time for new competitors to cut them at the knees relentlessly
Maybe. But by now, competitors are nowhere to be seen.

>> No.62178311
File: 2.30 MB, 1822x1224, 1499234569267.png [View same] [iqdb] [saucenao] [google] [report]
62178311

>>62178258
right...

> Nvidia will have to ungimp and so will RTG soon enough and with much lower margins if they want to survive. When it happens, remember this thread.

>> No.62178352

https://forum.beyond3d.com/posts/1998331/

>MinPrecisionSupport : D3D12_SHADER_MIN_PRECISION_SUPPORT_NONE (0)

>Apparently, AMD removed FP16-support from their newest (newer? didn't check every single driver) drivers for Fiji as well as Polaris (did not check Tonga yet).

> Yes, AMD disabled FP16 fon GCN3/GCN4 ISA GPUs since 17.2.x.

AYYMD, THE WAY IT'S MEANT TO BE GIMPED OLDER HARDWARE

>> No.62178353

>>62171548

confirmed 12yr old

>> No.62178391

>>62178311
Ungimp what? There are things that can theoretically be replaced by ASICS, and things that don't, or very hard to.
GPU+CPU+SSD Infinity Fabric ecosystem, for one, with its ROCm, MIopen and various other software shit support are hard to beat. Only by companies of size of AMD, Intel, Nvidia or Google and such. EPYCs+Radeon Instincts already heading to most cloud compute services, so... No competitors in near future, not till ~2020, or even 2025.

>> No.62178398

>>62170558
we literally have discussed this for 1.5 weeks now
we already knew vega is a compute card why you act suprised?

>> No.62178399

>>62173526
not all flops are equal

>> No.62178400

>>62178352
>FP16-support from their newest (newer? didn't check every single driver) drivers for Fiji as well as Polaris
Fiji and Polaris did FP16 at what, 1/8x? 1/16x? 1/32x? speed anyway.

>> No.62178408

>>62178391
IBM holds enough patents for releasing a compute chip, so maybe they'll enter the market sometime.

>> No.62178412

>wait for vega
>wait for them to fix vega
STOP ALREADY
STOP WAITING
Either give up or buy nvidia

>> No.62178449

>>62175648
>Were talking about FP16 which is becoming a standard for gaymen
is it?

>> No.62178460

>>62173282
mein got 56 doesnt have the juice to surpass 1080 moron
it will be close when the discard and primitives gets enabled but thats it people that expect wonders from it are the same that hyped the cards once more to infinity and beyond

>> No.62178477

>>62173526
vega 56 is 200watt more than 1070?
how are you even are allowed to use the internet?

>> No.62178522
File: 87 KB, 637x622, 1496444863111.jpg [View same] [iqdb] [saucenao] [google] [report]
62178522

>>62178391
Ungimp the old as dirt open standard feature IP blocks that they continue to use as differentiators for their PRO cards vs consumers. Literally disabled by firmware/drivers/appl level software. That gimping.

> There are things that can theoretically be replaced by ASICS, and things that don't, or very hard to.
Yes, as it can, all GPU manufacturers did was include the IP blocks in their GPUs and called them server compute cards. Given that this can be done, Google spun their own .. named it a TPU, can do 100TFLOPS compute at half the power of their high end cards. Given that all they need to do is matrix math, and spit it out on a NIC, they're essentially going nobid on GPUs for this task. What they do next is open source the designs so that whitebox hardware can be spun and their margins cut as well as IP generation from the rest of the industry. Big data corporations have been doing this to every market segment and have literally cut billions from high margin hardware companies. Nvidia/RTG are going to get a taste of this in the upcoming years

> GPU+CPU+SSD Infinity Fabric ecosystem, for one, with its ROCm, MIopen and various other software shit support are hard to beat.
Nope it's not actually. That whole ecosystem is built on an old one that's been around in HPC for a decade+. Guess what? They're opening all of the IP blocks in the coming years.

>Only by companies of size of AMD, Intel, Nvidia or Google and such
Google, Amazon, facebook and others have been leading the charge for open source hardware which leads to whitebox boards which eat AMD/intel/Nvidia for lunch.

> EPYCs+Radeon Instincts already heading to most cloud compute services, so... No competitors in near future, not till ~2020, or even 2025.
EPYCs yes ... General purpose compute + low margins + no gimping
Radeon : TBD. They're launching at a time that IP blocks are going open and at a time that Whitebox enterprise solutions are igniting.

GPU divisions are in for a ride.

>> No.62178534
File: 246 KB, 1280x1233, 1497283296055.jpg [View same] [iqdb] [saucenao] [google] [report]
62178534

>>62178408
Lots of old are new entrants will be coming into the arena. It will be like the good ol days in hardware again. A flame will be ignited soon from software and it will bring all the boys to the yard

>> No.62178562

>>62178449
Yes, GPU divisions are trying everything they can to differentiate their products and appeal to general purpose compute. FP16 is part of that.Anything to squeeze performance out of that crusty GPU pipeline as they try to figure out how to make a traditional general purpose CPU pipeline. Doing so has awakened dragons. So, they'll be getting competition from all sides soon.

They stumbled into general purpose compute... Never forget that..
And no they wont survive for long trying to maintain GPU margins in general purpose computing. They'll learn that lesson soon.

>> No.62178577

>>62177460
Ah, less drivers = moar bloat on die.
And Radeon = consoles. That was a necessary sacrifice.

>> No.62178586

>>62178562
fp16 isnt really so much popular as fp64 are beggining to be

>> No.62178671
File: 181 KB, 888x1120, 1494387422977.jpg [View same] [iqdb] [saucenao] [google] [report]
62178671

>>62178586
FP64 is a meme being pushed by a misdirected simulation industry, overly complex and incorrect models of the world, and mathematicians turned applied engineering statisticians who, instead of trying to find accurate formulas for things, are resorted to brute force computational models + large width statistics.

They're about to get BTFO as well....

>> No.62179071

>>62178671
FP64 is essential for calculations in nature science and the benefits are real. It allows 15digits after the decimal point compared to FP32 which only allows 7. Physicists and chemists couldn't calculate anything without it.

It was invented in 1985 when things were actually designed out of need and not because they sounded fancy for marketing.

>> No.62179092

>>62178167
>FP16 = Get blown the fuck out by TPU calculators
At the cost of even more die area. GTFO frog man.

>> No.62179333

>>62178412
No, the jacketman would never cuck me again.
t. proud 770 owner

>> No.62179702
File: 21 KB, 417x405, 1494546940624.jpg [View same] [iqdb] [saucenao] [google] [report]
62179702

>>62179071
Because their understanding and models of the universe are shit.

>1985 understanding of the world
>1985 math
>1985 statistical models because no firm equations

>>62179092
> http://images.anandtech.com/doci/11749/hc29.22.730-tensorpu-young-google-page-037.jpg

Getting BTFO this easily.

>> No.62181595

>>62170558
>is a server card
>200.000million electricity costs/year
wat

>> No.62181958
File: 218 KB, 315x520, 1393944003201.jpg [View same] [iqdb] [saucenao] [google] [report]
62181958

>>62170558
So the upcoming Xbox is a Ryzen and Vega MCM contraption?

>> No.62181978

>>62181958
it's still jaguar and custom poolaris

>>
Name (leave empty)
Comment (leave empty)
Name
E-mail
Subject
Comment
Password [?]Password used for file deletion.
Captcha
Action