Imagination Technologies Share Discussion Threads

[caradog] 04/5/2017 08:21

Charismatic CEO, supine and stagnating board, extravagant and overambitious expansion, lovely new premises, neglect of core business, selling out piecemeal after missing opportunities to sell the business for much more, going out with a whimper. Textbook example of how not to run a publicly quoted company. Oh, and a gong for the Great Leader to complete a very British story.

[richardc77] 04/5/2017 08:04

In reality IMG should have concentrated on their world class core activity instead of wasting, time, focus and money but it's an oft made mistake. Back to basics

[swiftnick] 04/5/2017 08:00

It must be sad for SHY watching it all being dismantled. But difficult to think what other strategic options were left.

[richardc77] 04/5/2017 07:46

So MIPS & Ensigma being sold.

Good.

[nearlythere] 02/5/2017 11:48

look like the shorts are increasing

[rbalakrishnan] 02/5/2017 07:56

Not great news. Mediatek will be reprioritising their product portfolio away from the X series smartphone to focus on the P series. Looks like they have given up the fight competing with Qualcomm's 835 at the premium end of the market.

[dt1010] 29/4/2017 13:08

for more than 100p?? hmm

[amt] 29/4/2017 06:37

It would make sense for Samsung to buy IMG

[borromini1] 29/4/2017 04:31

Latest ARM Mali in big fail against IMG PowerVR from two years ago. No wonder Samsung are rumoured to be moving to IMG PowerVR to replace ARM Mali in next Exynos chip design.

Just released top ARM Mali 20 shader core GPU chip FAILS to surpass performance of over 18 months out of date IMG PowerVR 7XT 12 cluster GPU chip according to FUTURE MARK graphics and gaming benchmark tests!

Latest ARM Mali G71 MP20 20 shader core brand new Bifrost architecture on 10 nm process node in Samsung Galaxy S8 flagship FAILS to achieve higher average score on Future Mark's 3D Mark Sling Shot Extreme and Ice Storm Unlimited benchmark tests than over 18 month old Apple A9X GPU containing IMG PowerVR GT7800 12 cluster design older Rogue architecture on 16nm process node in Apple iPad Pro. Since then the IMG Series7XT Plus has been released by MediaTek in the Helio X30 and the new IMG Furian architecture is in the process of being adopted by licencees.

Samsung Galaxy S8 Exynos 9 Octa
Mali-G71 MP20 20 x Shader Cores on 10 nm process node



Apple iPad Pro A9X CPU
A9X GPU PowerVR GT7800 12 cluster on 16 nm process node




Sam Mobile Samsung next Exynos chip design report from two months ago ...

[orkney] 27/4/2017 10:33

hxxps://www.imgtec.com/news/press-release/imagination-shatters-price-barriers-to-soc-embedded-software-development/

[orkney] 27/4/2017 10:03

What, 98000 sold, 177000 bought thus far

[mallorca 9] 27/4/2017 09:55

Heavy selling here today.
Looks as though the decisive break down below £1 may be about to happen.

Gap down then to 60p.

[stunger] 24/4/2017 12:07

Without knowing otherwise, there is a chance that Apple may infringe upon IMG's IP and if so IMG would have some presumably lucrative leverage. But if it is assumed that Apple have dotted the i's and crossed the t's and that their technology doesn't make any such infringement on IMG or anyone else's IP. If it were also assumed that their resulting graphics technology is far better than IMG's and the rest of the competition then wouldn't IMG have greater opportunity to get into the high volume lower/middle end of the market which is now going to need a much higher level graphics engine so as to not look prehistoric against Apple?

[mallorca 9] 24/4/2017 06:58

Rossi,

I've just returned from holiday and have read your post ... can you please help me to understand how this can be fairly valued at nearly £300m ?

[orkney] 23/4/2017 11:04

Borromini, Apple only gpu. Based upon Apple statement 15-24 months, I would have thought it is highly unlikely.

[willie99] 23/4/2017 09:27

Sunday times says img is considering selling a stake in mips to plug funding gap.

[borromini1] 23/4/2017 09:01

TSMC reported as fabricating and delivering 50 million A11 chips between April and July 2017 and 100 million by end of 2017. Does anyone think these will contain an Apple only GPU with no IMG IP?

[borromini1] 23/4/2017 08:38

Sheep_Herder - Having got to the position that a dedicated GPU/GPGPU is best for graphics and good for tasks such as machine learning and a dedicated parallel compute engine is best for compute and could be good for graphics...

Two questions, first can you provide lists or article links of the best applications of and examples of existing parallel compute engines at any size but also mobile size?

Second at what point would a combination of both chips be most effective for different power and area formats of mobile phone, tablet, laptop, and how similar is such a solution compared to other existing two chip complimentary combination examples such as in ARM's big/little approach or in laptops the integrated graphics combined with dedicated graphics card approach where entire pipelines are kind of replicated to address power and performance issues?

[sheep_herder] 21/4/2017 14:36

I'm not going to bother any more after this but the public release of an API is long after the companies involved have hashed it out for years. The reason IMG, ARM, AMD, Intel, NVidia etc pay lots of money to sit on these panels is so that they can get their features into each release and give their input on what they think should and should not go in. Vulkan was a very quick turn around as it was in response to Metal. The architectures and ISA for a GPU are defined in conjunction with the API development so that there are no surprises. It's no use building something that won't support the upcoming API's for 3 years after the GPU is released. That would be a very expensive mistake.

[pottsey] 21/4/2017 13:45

“You cannot build a GPU for a graphics API and then hope to some day run OpenCL on it.”
Then why have we seen GPU’s predate a release of an API by over a year and a half and still fully support the new API? Depending on the features the API adds all or parts of new API can be supported at a later date. For example the SGX543 predates the release of OpenCL 1.1 by well over a year and supports 1.1.

Take Vulkan it’s a new API that works on GPU's that got designed 2+ years before Vulkan was released.

I agree you can optimised a GPU for compute what I disagree with is how big a change this is. Relatively speaking compared to the rest of the GPU it’s not a massive change with loads of hardware sitting idling away doing nothing while you render graphics. A lot of what is useful for compute is also useful to render graphics.

Removing GPGPU support doesn’t mean you can remove all that hardware you use to do GPGPU as you still use a lot of that hardware for graphics. The way I see it is to make a 2nd chip for compute only that cannot render graphics you have to duplicate hardware you already have in the GPU. The way forward is integration not separating things out.

Thank you for coming back and explaining things in detail. I missed graphics discussions like this even if we don’t agree on how much hardware can be removed to lose GPGPU support.

I agree some hardware can be removed if you lose GPGPU support. But how much hardware would you estimate is used both in GPGPU and graphics and would have to be duplicated in a dedicated compute only, no graphics chip?

[pottsey] 20/4/2017 23:08

That not very helpful, if I made a technical error please point it out and I will correct myself. I know I sometimes make mistakes so I just reread up on GPGPU and everything I have says it’s fundamentally a software concept. You sometimes get a few small hardware changes like muti level caches to optimize for mainstream compute but at the core of it GPGPU is a type of algorithm, not a piece of equipment. After reading the documents on GPGU it appear as though its you who is wrong not me.

Take the SGX GPU or any other GPU which had a GPGPU API like OpenCL added at a later date without any hardware changes.

You can split a GPU into two core parts a fixed function area and a programmable hardware area. That programmable hardware area is used to render modern graphics like shaders and do GPGPU as they are the same thing. By improving the programmable hardware area you are both improving the rendering of graphics and GP Compute power. They are related as they are the same hardware for the most part. Like I said before there are a few small optimisations but it’s not a case of tons of extra hardware added to do GPGPU like you said. You don’t get tons of GPGPU hardware sitting idling away when rendering graphics just like you don’t get programmable graphics hardware idling away when doing GPGPU as mostly its all the same bit of hardware bar a few tiny changes.

The basic version is the programmable hardware for rendering graphics like shaders is the compute hardware that does GPGPU. You cannot remove this hardware from the GPU as then you cannot render graphics. So by making a 2nd compute only chip that cannot render graphics you are duplicating hardware you already have inside the GPU. All the 2nd chip is doing is wasting die space and wasting energy for no benefit. If you need more compute power you don't waste die space duplicating what you already have you use that space to improve the current GPU programmable hardware that does GPGPU.

Both GPGPU tech docs and the GPGPU wiki page confirm what I have said.

[sheep_herder] 20/4/2017 21:03

Bored now. You can lead a horse to water...

Edit: the sad thing is that you've got 3 likes which means not only are you misguided but there are others that are falling for it too. Shame really.