Hi all, it's been awhile since my last video.
I've been too busy with work to formulate my thoughts for a coherent presentation on
many interesting topics.
But the recent events in the past week are quite extraordinary and I felt it's important
to provide some clarity as well as contribute to discussion among the tech enthusiast community.
The topics I will cover today will be Raja Koduri's move to Intel along with my thoughts
on Vega and Navi.
Let's start with the the Raja saga.
The common question many of you have would be, why?
Was he fired by AMD or lured by Intel?
Did Raja help or hindered AMD and what will be his legacy at AMD?
Ofcourse I have to point out, only those at the table will know the truth, I am not an
insider and so this is only researched speculation from third party info.
To understand this whole situation better we do need to understand the past and what
led us to our current situation.
If you've seen my previous video, Vega and the Compute Conumdrum, you will know that
it wasn't long ago when ATI's GPU architecture was the leader in the graphics industry.
Note, I do not mean marketshare, raw performance or even performance per dollar.
As a tech enthusiast, to me, the best architecture is one that delivers high performance with
the least power and the fewest transistors because it signifies a true engineering lead.
But at around 2008, AMD decided to throw that lead away to turn towards a new direction,
one towards General Purpose GPU computing.
The reason, many believed the shrinking PC gaming market would not last and the already
low profits would not be sustainable given the vast R&D resources required moving forward.
The discrete GPU market was also volatile because of fierce competition, ATI and NVIDIA
traded blows, and when either wasn't on top, they had major drops in revenue and profits.
NV's CEO, Jensen Huang, realized they had to diversify away from PC graphics much earlier
and so CUDA and Tesla accelerators were born.
The idea was that these new markets would dampen the volatility of the PC graphics revenue,
and it worked.
As we saw, when there's strong competition from AMD GPUs, NVIDIA made a profit and a
large chunk was due to the professional market.
In essence, Quadro & Tesla revenue enabled further R&D investments even if the PC gaming
market was volatile.
Returning to Raja and AMD, as you may know, Raja has a long history at ATI as one of their
major graphics engineer.
When ATI merged with AMD, under the leadership of Hector Ruiz, AMD believed combining the
CPU & GPU would enable AMD to be the sole player in emergent graphics markets, where
Intel could not compete.
The drive for this Fusion approach was strong and the ATI crew were well funded, morale
was high and people were passionate.
During this time, Raja contributed with the TeraScale architecture, optimized for graphics.
The efficiency allowed AMD to have a strong presence in notebooks and recover close to
50% of the PC marketshare.
However, AMD was suffering financially during these periods, due to the overwhelming bribes
that Intel paid to major OEMs to not buy superior AMD CPUs.
This topic though, was thoroughly covered by Jim over at AdoredTV so I will not go into
it, but there's a link in the description for those interested.
Basically, to survive the tough times, AMD needed to restructure, a euphemism for firing
many good hard working staff.
The new CEO, Dirk Meyer, felt that his predecesor had gone off-focus with the ATI acquisition
along with others, and moved to refocus AMD back to it's core business, that is, CPUs
and Servers.
Also under Dirk Meyer, the decision was made to drop the focus on making the most efficient
graphics architecture, because many thought the future of gaming would be consoles and
mobiles.
What's left of ATI had to refocus to develop a brand new General Purpose architecture so
that AMD could compete in the emerging HPC market with their own accelerators.
Now, if you have studied up on Raja's career and him as a person, you will know that Raja
is first and foremost, a graphics guy.
It is indeed his passion to bring great computer graphics to as wide an audience possible.
Because of AMD's new GP-GPU focus and the shrinking ATI budget along with ever smaller
staff, Raja decided to take an offer at Apple to lead their graphics engineering teams so
he left AMD in 2009.
Coincidentally, Jim Keller, a renown CPU architect who was at AMD also left to join Apple around
this time.
It's no wonder then, with the talent & budget available, Apple quickly became the industry
leader in CPU & graphics on mobiles.
Fast forward a few years to 2011 and AMD was in dire position, with bankruptcy looming
on the horizon.
Rory Read was the new CEO and his mission was to do whatever it takes for AMD to survive
until the new CPU architecture being developed by Jim Keller, who also had returned to AMD
was finished and launched.
Read also leverage AMD's graphics technologies and IP to start the semi-custom & embedded
business, leading to major wins in the Xbox and PS4, providing revenue to keep them afloat.
In 2011, AMD also launched GCN to consumers with Radeon 7970.
Immediately, it was obvious that this architecture was not optimized for gaming graphics.
It was a relatively large & power hungry chip by AMD standard, on a new 28nm node, and it
was only ~30% faster than their previous gaming flagship, the TeraScale 6970.
30% may sound like a lot by today's standard, but back in the days, a true generation and
node leap yielded typically 80 to 100% performance gains.
NVIDIA was very happy with the situation because it meant that they no longer had to compete
with AMD in gaming using a much larger and power hungry General Purpose GPU like they
were forced to do for many years.
Instead of selling big Kepler to compete against a small mid-range AMD GPU, they could actually
sell their own mid-range Kepler at a high price.
A chip such as GK104, instead of being sold at $299, sold well at $499.
Meanwhile big Kepler instead of being $499 it sold at $699.
Ofcourse, no one should blame NVIDIA for taking the opportunity given to them.
It was infact Rory Read who let this happen, by pricing a mid-range chip such as Tahiti
at $549.
The justification being that it was faster than the previous NVIDIA Fermi flagship, the
$500 GTX 580.
When we think about this logically, it's perfectly normal for a next-gen mid-range to displace
the prior-gen high-end, at actual mid-range prices.
That's real progress and the result of innovation.
Because TeraScale was superior in graphics, it enabled fierce competition which drove
down GPU prices.
GCN was barely competitive versus Kepler in graphics, and it's pricing caused the floodgates
to open allowing NVIDIA to grow their gaming revenue and profits to unprecedented levels.
Meanwhile, GCN which was designed for General Purpose Compute and actually very good at
it, never stood a chance in those lucrative markets.
What AMD didn't plan for, because their leadership came from CPU backgrounds, is how important
software is to succeed in the GPU accelerator markets.
A CPU doesn't need an insanely complex driver layer or a major software stack to enable
it to perform.
A GPU does.
NVIDIA and Jensen Huang are especially aware of this fact, and so CUDA was designed first
and foremost to secure their gains in these markets.
The entrechment of CUDA is real and something not even Intel with vastly more industry connections
and money can easily displace.
Basically, AMD already lost the battle the moment they half-hearted and with a small
budget, decided to enter into NVIDIA's established GP-GPU domain.
At around this time, Lisa Su became the senior VP at AMD.
She very much like Rory Read, understood the need to diversify, but she also knows how
important the gaming market is.
As such, a strong graphics architecture was required to compete, but AMD cannot have a
clean slate restart because GCN backwards compatibility is a critical factor for their
lifeline, the semi-custom business.
Also, the promise of the growing HPC market and emerging AI & MI markets was too great
to not pursue.
Lisa wanted a single GPU architecture that can do it all, and do it well, no easy task
as you can imagine.
Raja was asked to return to AMD, to lead their graphics division and tasked with creating
future GCN ISA-compatible architecture that would enable AMD to compete in all their markets.
Now is where I have to point out something very important that a lot in this community
don't realize or want to discuss.
Guys like Jim Keller and Raja Koduri are highly respected, the cream of the crop in their
fields.
They would have received offers from many companies, some of which undoutedly carry
a huge paycheck and a well funded division for them to lead.
Why did they return to AMD which was struggling financially and as far as they knew back then,
heading towards bankruptcy?
Their return to AMD was not about money, it was about saving the company & the jobs of
many good people.
Ultimately, it is because of loyalty and friendships and for that, we have to give them due respect.
Back to the GPU topic, a major architecture always takes a long time in this industry,
typically 3 to 4 years ahead is the time-frame when major engineering decisions reach the
market.
While Raja may have an influence in Polaris, it was mostly completed before his work at
AMD began.
Vega however, certainly would be the result of Raja's guidance.
While gaming performance of Vega leaves many wanting more, the biggest problem with Vega
is that it's actually late, had it launch in Q4 2016, even as a GTX 1070 and 1080 competitor,
it would have a better reception.
Why was it late?
Mostly due to HBM2.
Vega engineering samples were running in 2015, it didn't launch until 2 years later which
is one of the longest delays as far as I can recall for a GPU.
HBM2 schedules were pushed back by SK Hynix as they had yield issues, and Micron's HBM2
production was completely bought out by NVIDIA for their P100 Tesla.
The question some of you may ask is why Vega needs HBM2 at all.
Contrary to what most people think, it is not power related, not memory footprint issues
and certainly not bandwidth given rising GDDR5 speeds and new developments.
As I've talked about before, an important unique property of HBM2 is within the memory
banks itself.
GDDR5 and such are very good for transfers of large chunks of data, but they are very
slow for rapid read/write cycles of small chunks of data to the same memory banks.
HBM2 is good for both workloads, and in the HPC market, fast fine-grain read/write cycles
of small pages is a very important feature to have.
Features like the High Bandwidth Cache & Unified Virtual Memory require this ability to really
shine.
It's a major reason why NVIDIA's P100 Tesla uses HBM2.
Vega 10 is a General Purpose architecture designed for Servers, AI & Machine Learning.
It's design also enables some novel solutions, such as the Pro SSG for big-data acceleration
& film studio workloads.
But is it also a gaming architecture?
Did Raja accomplish the mission given to him to make it good for gaming as well?
Clearly not, not yet.
GCN's weakness has always been the loopsided compute-focus design, with few transistors
dedicated to graphics, namely, geometry and rasterization.
To simplify this issue, imagine a typical frame of a game being rendered by the GPU,
it will comprise of several elements, such as vertices, geometry, textures, becoming
pixels which is then post-processed with shader effects.
If a GPU is limited by geometry & rasterization, it will scale poorly in performance even with
more cores, since all that extra raw Teraflops of computing power, cannot contribute to some
of the steps involved in creating a frame.
Another interesting aspect of graphics rendering, is that gamers often assume different resolutions
scale uniformly, as in all the elements scale but it's not true.
Vertices and geometry workloads do not change much at 720p vs 4K resolution, it's static,
while what changes are rasterization, pixel shaders and post-processing workloads.
A GPU that's geometry bound, but with ample shading power, will be slower at low resolution
while relative performance increases at higher resolution as the workload emphasis switch
to pixel related shaders.
This is the typical observed performance profile of GCN, especially in very loopsided designs
like Fiji.
With Vega, we do see better scaling, there is a consistent strong lead in performance
over Polaris & Hawaii even at low resolutions where relative geometry performance matters
more.
This is very much due to the new scheduler changes, which dynamically assign work to
the four Geometry Engines to lift efficiency of these units.
The Intelligent Workgroup Distributor, which is designed to assign work to idling Compute
Units first also helps in increasing shader usage at low resolution.
The scheduler also has a larger SRAM cache for storing more instructions, alleviating
CPU stalls, leading to higher performance under CPU bound scenarios.
But these hardware improvements are not enough.
Vega 56's raw performance should enable it to match the GTX 1080 in graphics, while Vega
64 should be closer to the 1080Ti.
The problem is still geometry & rasterization bottlenecks because these changes only lift
the bottleneck ceiling a little, while what GCN really needs is to break through that
ceiling entirely.
Primitive Shaders, if it's implemented by game developers or by AMD will help with the
vertice and geometry steps, so performance can improve, but a bottleneck still remain
with rasterization.
Importantly, NVIDIA has moved on with deferred rasterization and gained it's efficiency advantage,
while AMD is still on immediate mode rasterization.
On top of this, AMD has less resources on GCN dedicated to rasterization overall so
a GCN Teraflop will not equal a Pascal or Volta Teraflop in graphics until this bottleneck
is lifted.
What GCN really needs is the lifting of the 4 Shader Engine limitation.
It needs to be dynamic, scalable, just like NVIDIA's PolyMorph Engines & Graphics Processing
Clusters.
Ofcourse this requires more transistors dedicated to graphics function, which results in a bigger
die that uses more power.
This is the downside of making a single chip for many markets.
I believe the foundation for this change is already within Vega's scheduler & front-end,
but a bigger & more power hungry chip was off the table because of a combination of
limited resources, but also, it does not make economic sense at this time.
The solution should be obvious by now if you are a tech enthusiast.
Look at the accomplishment of Zen and it's scalability.
AMD uses extra die-space with Zeppelin to enable scalability with the Infinity Fabric,
but they make up for it by not needing to produce one giant power hungry chip to reach
higher performance levels.
Raja has talked about the future of GPUs being multi small dies working together for awhile
and the foundation is done for Navi to succeed.
Let me explain why Navi is capable of being an all-in-one architecture.
To specialize an chip for HPC or AI and MI, like Vega, you require more transistors & you
increase the power usage.
Compared to a graphics-optimized chip, it will lose in performance per watt and performance
per mm2, both critical metrics to be succesful.
Here's the thing, what if performance per mm2 no longer matters because your production
yields are so high, since you do not need to make a huge die but multiple small ones?
Remember, yield is exponentially related to die size.
A smaller die not only yields a little better, but a lot, and so production cost is lower
while production rate is higher.
It should be clear immediately how Navi could be inefficient in performance per mm2 and
still be cheap to produce in mass quantities.
But what about power?
For GPUs, you simply go with lower clocks, lower voltage, and make up the performance
loss by having more cores.
Basically, performance per watt can increase at the expense of performance per mm2.
If die space is no longer a limiting factor, such as a multi-chip solution, the performance
per watt issue will solve itself in the process.
There's some concerns about the difficulties of making a multi-die single GPU, but it's
actually an easier task than with CPUs.
Unlike CPU workloads where thread dependencies are high, where out of order execution can
happen and latency is critical, GPU workloads are incredibly parallel and non-thread-dependent.
GPUs also hide latencies very well, unlike CPUs.
The next question is, how can GCN go wider, over multi-dies?
Well, by not being limited to 4 Shader Engines and with a scheduler clever enough to keep
many Shader Engines well fed.
You see, Vega has already laid that foundation for Navi.
But that's not really the end for Vega, as I've stated in the other video, if FineWine
is what you enjoy, Vega is the finest of wines.
All that untapped FP16 power, as new games begin to tap it, there's potential there for
some strong gains.
Due to a lack of gaming driver optimizations, the performance for RX Vega currently is nowhere
near it's potential, and so AMD had to sell them at a lower price than they hoped for
to compete.
In essence, as a gamer, you get competitive performance for your money today, with a lot
of untapped power under the hood.
Does this make Vega 10 a failed GPU design?
Far from it.
It's solid as MI25 & for professionals.
Vega 10's inclusion in arguably the best mass produced workstation, the new iMac Pros, is
also testament to it's one-chip for all markets design.
One could say, professional Vega is subsidizing RX Vega for gamers, an identical situation
NVIDIA was in when they had to sell big & power hungry chips for low prices to gamers.
With the typical volumes that Apple iMacs sell, and this upcoming model being the most
anticipated in years, it's obvious to me that revenue from Vega will ramp up in 2018, and
it will pay for it's R&D cost many folds.
This combined with a resurgent AMD CPU business, should put much needed money in AMD's coffers
to hopefully increase their odds of a smoother Navi launch.
Finally, regarding Raja's move to Intel, it's not something that happens quickly, a betrayal
as some call it, these moves along with the AMD & Intel collaboration would have been
discussed a long time ago.
Anyone with experience in large institutions knows these things take time, certainly back
porting Polaris to support HBM2 and Intel's EMIB technology is not a quick task.
And we know that engineering sample has been in the wild for a long time.
From what I've managed to dig up, Raja was personally affected by being asked to do too
much with too little, this forced him into demanding his RTG staff to work extra hard,
with constant overtime.
They had to get AMD's Pro drivers up to scratch, to utilize Vega features on PC workstations,
and to make it compatible with OSX & the Metal 2 API.
On top of this, they also had to create entire new HPC, AI & MI software stacks, and CUDA
to OpenCL compilers, which is a huge endeavor by itself.
All the while, they have to keep up with gaming optimizations for past and current GPUs.
For a software team a fraction the size of NVIDIA, and I cannot stress this enough, this
was simply unfair to all those workers.
I'm a software guy and I feel for those in AMD's driver team, so while gamers want their
Primitive Shaders enabled, it's just not at the top of their priority.
Lisa Su had to ensure Zen succeeds above all else because AMD's survival was at stake,
and so this limited the resources available for RTG.
Tough decisions were made and it's clear now that the partnership with Apple, Baidu and
others in the professional market was more important.
When you think about it, Intel's bribes over a decade ago still carry deep consequences
for AMD to this day, meanwhile it barely cost Intel anything in terms of fines.
I realize this is why some called Raja's move to Intel a betrayal, because he's joined the
enemy.
But times have changed, people have to move on from the past and find a solution for the
future.
Raja joined Intel for the same reason that Jim Keller had when he left AMD to work for
Tesla, two years before Zen was released to consumers.
The mission was done and it's time for a new challenge.
Raja had good working relationships with important leaders at Intel during their collaboration,
they recognized his talent and they want to grow markets where graphics & GPU accelerators
are priority.
Ofcourse a bigger paycheck and a huge R&D budget for his own division to lead is a big
draw card.
While there's so many interesting topics to discuss and I could go on and on, this video
is long enough.
I hope this one was insightful and as always, feel free to discuss anything related, whether
you agree or disagree, in the comments.
See you all next time!
For more infomation >> [TUTO] Comment avoir plein de vues sur sa chaine youtube ? [2017] - Astuces audience - Duration: 5:58. 
For more infomation >> [INTERVIEW] Chris Darril - Créateur de Remothered #PGW2017 - Duration: 3:42. 
For more infomation >> BERKATILAH - Live Koor Di Gereja Santo Yosep Purwokerto - Duration: 3:40. 
For more infomation >> Tebello Sukwene (We Nhliziyo Yami) - Duration: 4:51. 
For more infomation >> Tebello Sukwene (Jesus My All) - Duration: 5:53. 
For more infomation >> FutureNet prezentacija sa SRPSKIM titlom ! - Duration: 12:10. 
For more infomation >> В чем секрет хештега? Для чего хештеги нужны? #Хештег — навигатор социальных сетей. - Duration: 1:30. 
For more infomation >> THẦN DƯỢC GIÚP QUÝ ÔNG TRỊ DỨT ĐIỂM CHỨNG "TRÊN BẢO DƯỚI KHÔNG NGHE" ( D.ƯƠNG V.Ậ.T ) - Duration: 2:36. 
For more infomation >> Una musa que vale más de 200 pesos - Duration: 7:13. 
For more infomation >> Q अक्षर , Q name, Q नाम, Q naam, कैसे होते है Q नाम वाले, जानिए कैसे होते हैं 'Q' नाम वाले व्यक्ति - Duration: 3:29. 
For more infomation >> "Hustle Game" | 2 Chainz Typ...
For more infomation >> Não diga "NON LASCIARE DI VISITARE IL COLOSSEO" em italiano! I Vou Aprender Italiano - Duration: 3:25. 
For more infomation >> Genf 2012: Hyundai I-Qniq Elektro-Coupé mit Reichweite - Duration: 2:31. 
For more infomation >> JUST IN TIME Michael Bublè & Tony Bennett karaoke Playback instrumental wav edit by © Graziana '17 - Duration: 2:38. 
For more infomation >> Rob Kardashian : Dream, il fête son anniversaire - Duration: 2:27. 
For more infomation >> Types of Instagram DM's - Duration: 5:02. 
For more infomation >> Pulling my tooth out | Kailtlyn's Cool Club Weekly Videos - Duration: 4:19. 
For more infomation >> This Week in Unnecessary Censorship - Duration: 1:28. 
For more infomation >> Block B_BASTARZ - Make It Rain MV (Greek subs) - Duration: 3:05. 
For more infomation >> Puigdemont o el síndrome del presidente exiliado - Duration: 4:39. 
For more infomation >> 'El País' fulmina a Pablo Iglesias, Irene Montero y Echenique e incendia Podemos - Duration: 4:03. 
For more infomation >> KELLE PAÇA Geldi Çekim Patladı! | Tıka Basa Challenge - Duration: 7:23. 
For more infomation >> Co chcecie na kanale ??? - Duration: 0:09. 
For more infomation >> Monique fala sobre relação com Marcos: "Eu tô carente, trancada há 60 dias" - Duration: 4:10. 
For more infomation >> Critiche degli altri: effetto Napalm51 - Duration: 1:24. 
No comments:
Post a Comment