NVidia Losses half of its value

[maartendc]

18 hours ago, yian88 said:

Market crash? bear market?  Seems like everything thats overvalued is at risk of crashing in 2019 hopefully  not the banks again.

Same thing is happening with EA for example... .losing a lot of value currently. Investors are getting a bit worried, and so everything that has a bit of negative news now is being sold off.

 

Granted, EA hasn't had much positive news even since Battlefront 2.

[williamcll]

Big oofs

[Adramalech]

It was just an opinion

22 minutes ago, ThePD said:

Nobody in this thread is praising ray-tracing performance.

that was a general statement however you proceeded to praise ray tracing as 

 

23 minutes ago, ThePD said:

the future

 

24 minutes ago, ThePD said:

This thought process is silly, you would of said the same about dual core processors in the day because all the software was developed for single cores.

You mean how multicore processors didn't become mainstream until they became affordable? Yes i would

43 minutes ago, ThePD said:

Ray Tracing IS the future. It is used in almost every high budget hollywood film today. The fact that we can do rendering in real time using RTX is an insane technological leap.

Only reflections are being ray traced in games at this point with the current hardware and pricing, think ahead while keeping in mind the current market share, developer compliance and public opinion from an investment point of view and the current prospect of Ray Tracing starts to look a lot like a liability.

 

1 hour ago, ThePD said:

Just because you find it to be expensive for what it is, does not mean it isn't relevant. People in this post keep talking about it, so it is clearly a relevant technology.

please keep in mind this post is titled  NVidia Losses half of its value so it is understandable why ray tracing would be a recurring topic 

[ThePD]

1 minute ago, Adramalech said:

please keep in mind this post is titled  NVidia Losses half of its value so it is understandable why ray tracing would be a recurring topic 

There are many companies on the stock market that have lost half their value.... AMD was trading at 34.14 in October and dropped to 16.15 within the same month. There are many companies in the market right now that lost all of the gains they had from 2018. This is not a Nvidia issue, but a issue of investors not trusting the market as a whole.

[Syntaxvgm]

Wallstreetbets summarized the current market conditions quite well earlier this week 

[LAwLz]

23 hours ago, ThePD said:

The whole market is down. This isn't just Nvidia, the same applies to Intel and to AMD. The same trends are seen in all tech stocks. 

A thousand times this.

AMD went from a peak of 33.37 dollars per share, down to their current 19.91 dollars per share. They lost 40% of their value during roughly the same time Nvidia went from 292.02 to 148.74 (49%).

 

Intel has been surprisingly stable. Peaked at 57.6 and is currently at 48.07 dollars per share (16.5% loss).

 

Maybe I should make a thread about AMD losing 40% of their value as well. Would be fun to see people speculate on why. Hilarious to see some people actually believe that it's because of RTX.

Also kind of funny how clearly biased a lot of people are against Nvidia. When something negative affects both Nvidia and AMD, for some reason people brush the AMD news under the rug and only focus on how poorly Nvidia is doing.

[Rohith_Kumar_Sp]

aren't all stocks down ?
 

 

[DemureGamer]

20 minutes ago, LAwLz said:

A thousand times this.

AMD went from a peak of 33.37 dollars per share, down to their current 19.91 dollars per share. They lost 40% of their value during roughly the same time Nvidia went from 292.02 to 148.74 (49%).

 

Intel has been surprisingly stable. Peaked at 57.6 and is currently at 48.07 dollars per share (16.5% loss).

 

Maybe I should make a thread about AMD losing 40% of their value as well. Would be fun to see people speculate on why. Hilarious to see some people actually believe that it's because of RTX.

Also kind of funny how clearly biased a lot of people are against Nvidia. When something negative affects both Nvidia and AMD, for some reason people brush the AMD news under the rug and only focus on how poorly Nvidia is doing.

It's an interesting thing for people to have a bias on. I don't have a loyalty to a certain company. As long as they turn out some good products and they don't charge us our first born, I'm fine. I want to buy a new RTX 2080 Ti, but I'm probably going to go with a Vega 64, for the time being.  I would love to see more done with RTX, but I'm not about to be price gouged to be an early adopter.

[Enderman]

6 hours ago, schwellmo92 said:

AMD is the better buy IMO. Nvidia mcap is 90b, Intel 210b and AMD 20b.

AMD is rarely chosen for supercomputer or enterprise processors, and they have noting to do in the deep learning and autonomous driving AI market.

[mr moose]

7 hours ago, TVwazhere said:

snip

 

3 hours ago, LAwLz said:

snip

 

2 hours ago, Rohith_Kumar_Sp said:

snip

 

 

Sorry guys but I have it on good authority that Morgan Stanley said something bad about RTX and that caused the sole drop for Nvidia. Every other share going down at the same time must be caused by something else that coincidentally didn't effect nvidia. 

[Amazonsucks]

9 hours ago, Enderman said:

AMD is rarely chosen for supercomputer or enterprise processors, and they have noting to do in the deep learning and autonomous driving AI market.

Cray Shasta will be using Epyc. Theyll gain marketshare from Intel. Fujitsu still makes the best HPC CPU, the best supercomputers and theyll be doing a lot with the DLU and Digital Annealer as well despite their tiny market share.

[Enderman]

3 minutes ago, Amazonsucks said:

Cray Shasta will be using Epyc. Theyll gain marketshare from Intel. Fujitsu still makes the best HPC CPU, the best supercomputers and theyll be doing a lot with the DLU and Digital Annealer as well despite their tiny market share.

Take a look at what GPUs are being used by half of the top 10 supercomputers

[schwellmo92]

10 hours ago, Enderman said:

AMD is rarely chosen for supercomputer or enterprise processors, and they have noting to do in the deep learning and autonomous driving AI market.

Was rarely chosen*

 

New EPYC installations are being announced every week and with 7nm Vega they’ll be getting in to some of the deep learning installations as well, when you invest in a company you invest on what the potential is, rarely you invest in a company for them to continue doing what they already do. Another thing to keep in mind is AMD does both CPUs and GPUs.

[D13H4RD]

13 hours ago, Rohith_Kumar_Sp said:

aren't all stocks down ?

Yep, especially on the NASDAQ. Many of the declines can be attributed to similar reasons but for some, they have additional unique reasons as to why.

 

For companies like EA, one of the unique reasons is generally weaker game sales alongside the not-so-great reception to Battlefield V. For NVIDIA, it's generally because the crypto-boom has made them overvalued. The RTX thing may have played a role but not to this extent.

[Amazonsucks]

2 hours ago, Enderman said:

Take a look at what GPUs are being used by half of the top 10 supercomputers

Nvidia Volta is the best GPU for HPC and deep learning, but i was talking about CPUs. Shasta will be using Volta's follow up architecture or Volta. 

 

But the two fastest supercomputers currently use neither AMD or Intel CPUs. They use IBM Power9 which is better than anything AMD or Intel make.

[D13H4RD]

17 hours ago, Adramalech said:

In my opinion, people need to stop praising ray tracing performance. 

I don't think anyone has praised it. The best thing I've heard on current ray-tracing performance is that it's playable, albeit barely.

[Prox1ma]

On 12/13/2018 at 12:10 AM, DrMacintosh said:

Why would Nvidia be going down that much? The only negative new's I've heard from them is that they have had a bit of overstock. 

Because it was already overvalued. This was expected. But, more than the company, it's the investors who would have lost a lot of their money due to this.

[leadeater]

On 12/14/2018 at 6:18 PM, Amazonsucks said:

the best supercomputers

Pretty sure both Cray and IBM on technical levels, capabilities and deployments are higher than Fujitsu. Excellent company but not the best, I'd be very weary of giving that title to anyone tbh since it's such a fluid and subjective thing. What you are building the cluster for, what will be run on it, energy efficiency goals vs peak performance balancing all factor in more than what CPU or accelerator of choice you went with.

 

On 12/14/2018 at 9:13 PM, Amazonsucks said:

They use IBM Power9 which is better than anything AMD or Intel make

CPU performance wise Power 9 is slower than Intel's best, this however comes back to the more important factor of what are you doing. IBM Power 9, and previous, have much better memory bandwidth performance so if your workload is very memory bandwidth demanding the slower computational performance of Power 9 is greatly balanced out by the superior memory performance.

 

 

IBM used to dominate the Top 500 list, now they don't because overall if we want to generalize Intel CPUs are faster and have been for a long time.

[JediFragger]

After utterly fleecing gamers with the RTX series with almost next-to-useless features I'm glad! Feels like karma for us!!!

[Amazonsucks]

On 12/15/2018 at 11:23 PM, leadeater said:

Pretty sure both Cray and IBM on technical levels, capabilities and deployments are higher than Fujitsu. Excellent company but not the best, I'd be very weary of giving that title to anyone tbh since it's such a fluid and subjective thing. What you are building the cluster for, what will be run on it, energy efficiency goals vs peak performance balancing all factor in more than what CPU or accelerator of choice you went with.

 

Cray is technically behind Fujitsu in every department but price and possibly some of the interconnect secret sauce. They don't even build the same class of system as Fujitsu anymore.

 

Cray doesn't have their own CPU architecture at all anymore since Threadstorm. Cray build much less expensive systems than Fujitsu for the FLOPS, however this is reflected in the rather dismal computational efficiency in heterogeneous systems. Dragonfly is a great interconnect topology, but its not nearly as efficient as the 6 dimensional torus interconnect that Fujitsu uses in their PrimeHPC line. I was specifically talking about PrimeHPC vs Cray or IBM's offerings btw.

 

I realize that Fujitsu, IBM and Cray probably make about the same level of commodity processor based systems using Xeons and Voltas. They're all about 65% computationally efficient in terms of max vs peak FLOPS in HPL.

 

IBM was definitely more advanced than Cray with their Blue Gene architecture, which also used a 5 dimensional torus interconnect similar to Fujitsu's, and had excellent computational efficiency and interconnect performance both in terms of bisection bandwidth and injection per node.

 

The Blue Gene systems were also extremely expensive, like Fujitsu's PrimeHPC lineup and used bespoke CPUs like the PrimeHPC series as well. That's why IBM abandoned them in favor of Power and Volta systems: price. No one wanted to pay for Blue Gene anymore since the "good enough" cheap FLOPS approach was so much cheaper.

 

A system like PrimeHPC or Blue Gene is a totally different approach than even the newer(and also designed to be much cheaper) IBM Power9 Volta systems or what Cray has to offer.

 

Shasta will be technically very interesting from a packaging perspective, and also from an interconnect perspective, but they're still using DIMMs in it ffs! DIMMs in 2018 when PrimeHPC has been using 3D stacked HMC memory since early 2015.

 

PrimeHPC started out with the FX1, which used SPARC chips that evolved into the K Computer. K Computer cost $1.3 billion dollars, but it was also the most powerful system on the Top500 at launch. However, that's not why it is still better than the vast majority of systems on the Top500. The fact that its Graph500 and HPCG performance stayed at #1 in the world until systems with literally 10x the FLOPS were released is the real reason that Fujitsu makes the best systems. As does the fact that even the new IBM Sierra and Summit systems don't come anywhere close to the 93% computational efficiency of K, or its successor the FX10 or the current PrimeHPC FX100.

 

They are systems designed without the kind of cost compromises that commodity processor systems typically have. I guess that should really be expected, since systems like PrimeHPC, SX-ACE and Blue Gene use CPUs specifically designed for HPC, not x86 architecture cores that aren't far removed from a normal desktop CPU. 

 

HPCG performance is really the more important benchmark than HPL these days, despite the Top500 still being what everyone looks at. Also, the % of the system's peak FLOPS compared to its actual HPCG measured flops is a huge indicator of real world performance. 

 

Most commodity processor systems are somewhere between 1 and 2.5%, with all CPU systems usually being on the higher side. K Computer is 5%, or double what most other sytems are. If you look at another architecture that was #1 on Top500 like Taihu Light, its performance is actually only 0.4% of peak, which is abysmal. It must be said that the "old" style of homogeneous all CPU systems tends to make for better real world performance, despite getting a lot of FLOPS per dollar with heterogeneous systems. The even older style of real vector supercomputing, using the last true vector CPU(NEC SX-ACE) has 11% HPCG efficiency compared to its peak, which is insanely high. 

 

The one common factor with the highest efficiency systems(especially for their size and peak FLOPS) is that they're all homogeneous CPU systems that use bespoke CPUs.

 

But yeah I was only really talking about PrimeHPC compared to commodity systems with regard to Fujitsu. As for the Power9 vs Intel and AMD, neither Intel or AMD support NVLink or have the kind of capabilities that Power9 has overall. Intel doesn't have a single CPU that supports anything close to 4TB per socket, nor can they scale gluelessly to 16 sockets and 64TB of RAM. I realize those are the SU versions of Power9 and most petascale systems would be using the SO Power9s but they're all Power9 like Xeons come in different flavours.

 

To say that they are "faster" needs to be quantified. In the one benchmark you showed they are, however there is more than just one aspect to performance when it comes to HPC systems. The top 2 computers on the Top500 currently use Power9, so they're the "fastest" systems until Frontier and Post K come out.

 

We already know the specs for the prototype of Post K, which is 48+2 core 7nm ARM CPUs with at least 32GB of HBM on package with a revised 6 dimensional torus interconnect and it will be around 1 ExaFLOPS. Given the computational efficiency of the SPARC based PrimeHPC architectures that use the Tofu interconnect, you can expect the ARM version, which now has 512bit SVE as well as a bunch of much more advanced features than FX100 had to be even better. The upgrade to Post K is going to cost another billion dollars. That's a huge contrast to the couple hundred million at best that most Top500 systems end up costing in other countries. Japan spends a lot more per system.

 

As for the Top500, its less relevant than it used to be, because now workloads have shifted away from needing more FLOPS to needing more memory and interconnect bandwidth. FLOPS are basically free now, as I'm sure you've heard the saying.

 

And since we have working silicon of the Post K ARM A64FX CPUs, they are now the fastest CPU with 2.7 FP64 TFLOPS and 1TB/s memory bandwidth. The PrimeHPC chips may not get talked about much, but they really are the pinnacle of supercomputing and have been for a long time.

[Canada EH]

11 hours ago, JediFragger said:

After utterly fleecing gamers with the RTX series with almost next-to-useless features I'm glad! Feels like karma for us!!!

Gotta milk it while the milking is good. Gotta take notes from Linus campaign right.

 

 

 

 

Too much to read, just give us the down low.

 

@Amazonsucks

[Amazonsucks]

1 hour ago, Canada EH said:

Gotta milk it while the milking is good. Gotta take notes from Linus campaign right.

 

 

 

 

Too much to read, just give us the down low.

 

@Amazonsucks

Fujitsu makes supercomputers that use bespoke CPUs designed specifically for supercomputing.

 

As a result, they are better at doing that task than commodity x86 CPUs. They also have their supercomputing specific interconnect router integrated to the CPU itself. Since the interconnect largely determines the performance of modern supercomputers they perform even better.

 

Brief specs of their last chip released in 2015:

 

20nm SPARC XIfx 32+2 cores. 1.1  DP TFLOPS. 480GB/s memory bandwidth. Used 3D HMC memory before anything else.

 

Upcoming CPU preliminary specs:

 

7nm ARM A64FX 48+2 cores. 2.7 DP TFLOPS, 1TB/s memory bandwidth using 4 stacks of HBM2.

 

Their systems have unmatched overall performance but cost a fortune compared to normal x86 based systems. 

[leadeater]

3 hours ago, Amazonsucks said:

Dragonfly is a great interconnect topology, but its not nearly as efficient as the 6 dimensional torus interconnect that Fujitsu uses in their PrimeHPC line. I was specifically talking about PrimeHPC vs Cray or IBM's offerings btw.

The interconnects from each of them are all actually very similar in design, meshes of meshes. Aries went with high compatibility so used PCIe 3.0 between the Aries SoC and the nodes, that's really where the limitations in bandwidth come from. Not that it's too limited, each Aries SoC can move 188GB/s of data around. Cray is switching to Omni-Path based last I heard so that means they are leaning back towards more specialized than generalized, the talk about other CPUs and accelerators never really came to actual fruition.

 

There's becoming little point in making your own high end interconnect transport now though, IBM has already moved to using Infiniband. With Infiniband, Omni-Path and Ethernet offering blistering speeds augmenting these with your own overlay logic is a lot more cost effective and doesn't really limit you at all, not when look at the bandwidth on offer now and coming soon.

 

3 hours ago, Amazonsucks said:

HPCG performance is really the more important benchmark than HPL these days, despite the Top500 still being what everyone looks at. Also, the % of the system's peak FLOPS compared to its actual HPCG measured flops is a huge indicator of real world performance. 

Both are important while also both not being, HPCG has a heavy bias towards memory bandwidth which makes it better indicator for those workload types but there are still many workloads that don't live and die on memory bandwidth. HPCG isn't actually run in a way like the system is actually going to be used either, tasks that would be run on dedicated hardware is run on everything and effects the scoring.

 

Basically as they put it:

Quote

HPL represents an attainable, but optimistic performance target for applications. While there are applications, for example in material sciences, that have performance signatures similar to HPL, very many applications do not.

 

HPCG represents a large class of applications, but demands a lot from a computer system and can be considered a lower bound on the achievable performance for many applications.

 

Together HPCG and HPL can be interpreted as bookends on the range of performance a collection of applications can experience. Using the HPCG and HPL results for a given system as an interval, we have a basic metric for the balance of that system. The smaller the interval between HPCG and HPL, the better balanced the system is for a range of problems.

HPL is best case and HPCG is worst case and neither of them represent the performance that is actually going to be obtained when running the workloads the cluster was designed for, performance that is above HPCG and below HPL basically. Cluster efficiency is much higher than HPCG indicates. 

 

3 hours ago, Amazonsucks said:

As for the Power9 vs Intel and AMD, neither Intel or AMD support NVLink or have the kind of capabilities that Power9 has overall. Intel doesn't have a single CPU that supports anything close to 4TB per socket, nor can they scale gluelessly to 16 sockets and 64TB of RAM.

Not sure what you mean by gluelessly, Intel supports hot add and remove of CPUs and up to 8 sockets per system. Far as ram per socket Intel regressed in that regard, I don't know why. Broadwell-EX supported 3TB per socket, why they limited Skylake-SP to 1.5TB and only on M variants I have no idea, maybe caution due to the new Mesh architecture.

 

NVLink only helps with Nvidia GPUs so I'm not sure how much that really needs to be talked up and you can still do NVLink between GPUs just fine, NVLink to the CPU would only help if you actually need that extra bandwidth. PCIe 3.0 isn't that slow and using NVLink on the CPU limits the number of GPUs you can put in the system, if you want NVLink between everything.

 

What other capabilities does Power 9 have that Intel doesn't? Intel has got on package FPGAs if you're in to that thing, on chip Omni-Path, faster vector math by a lot (GPUS are even faster so moot). Intel's performance superiority allows them to command a much large price tag to, IBM CPUs aren't actually that expensive but omg the ram is.

 

3 hours ago, Amazonsucks said:

To say that they are "faster" needs to be quantified. In the one benchmark you showed they are, however there is more than just one aspect to performance when it comes to HPC systems. The top 2 computers on the Top500 currently use Power9, so they're the "fastest" systems until Frontier and Post K come out.

They are the fastest solely due to the GPUs, the CPUs are still lower performance than Intel. I showed one because showing multiple wouldn't change what is being indicated and would be a waste of screen space, you're more than capable of independently fact checking when/if you so wish. Here are some starter links: https://www.phoronix.com/scan.php?page=article&item=power9-epyc-xeon&num=2, https://sthbrx.github.io/blog/2018/08/15/improving-performance-of-phoronix-benchmarks-on-power9/

 

As a reminder Xeon 6138 is not Intel's best/highest performance CPU, it's actually very cheap.

[asus killer]

nvidia fall kind of mirrors Nasdaq's fall. The problem is much greater than the company, it's the market as a all. MAGA

[Citadelen]

Nvidia's stock had been riding a multi-year high based on their consistent successive earning beats, so when they finally fell short, cut guidance and had racked up a massive Inventory, the market reacted quite violently.