Why not create something far more advanced than the competition?

Discussion created by phatec on Mar 11, 2018
Latest reply on Mar 12, 2018 by ajlueke

Like • Show 0 Likes0
Comment • 2

When AMD releases a GPU, it seems like the competition has something far more advanced either already being sold, or waiting to be sold right after AMD releases their product.

How can AMD beat Nvidia's GPU's?... Build a GPU that is more powerful than 2x 1080TI's, and is several generations ahead.

Below is the dream GPU which will destroy the competition. I understand that engineering such a GPU takes time, but why not invest in something like this? Obviously don't make a dual GPU single card such as in the past, those were a mistake. Make a card with extremely profound specs, something Nvidia would never expect. Because the next gen game people will be playing is games such as Star Citizen.

Nvidia only does 10% performance boost increments each year, when a new card is released, so release one that is 100% better.

Chipset	AMD Radeon™ R-XI Eclipse (because it's going to Eclipse and block out the competition) GPU: GP102-350-A1
CUDA Cores	7,168 Stream Processors
Engine Clock	Boost Clock / Base Clock: 2,200 / 2,000 MHz
Performance	FP32 Compute: 16 TFLOPS or higher
Memory	24 GB GDDR5X HBM2 RAM (High Bandwidth Memory 2)
Memory Interface	512-bit
Memory Bandwidth	1000 GB/sec
Bus Interface	PCI Express x16 3.0
AMD CrossFire™	Supports 2, 3 or 4-Way CrossFire™ without Bridges
Resolution	8K (7680 × 4320) @ 60 Hz 5K (5120 × 2880) @ 120 Hz 4K (3840 × 2160) @ 200 Hz
I/O Ports	1x HDMI v2.1 port 3x DisplayPort 1.4 or 1.5 ports (if available)
Supports	API Support DirectX ®12 OpenCL™ 2.0 OpenGL® 4.5 Vulkan™ API

Outcomes

wimpzilla Mar 12, 2018 4:25 AM

Hello,

Simply because it cost a lot of money, time, resources, etc to be able to design, then build and sell a powerful gpu, taking also into account the present and future technological landscape.

Not all the gpu can be designed, often you have limited design capability, there are patent to take into account, then how many people, money you are throwing into the design and R&D phase.

The same apply to the manufacturing, you have limits on how big a gpu die could be, you have efficiency limits, how many good gpu dies come out from a whole piece of silicon. The manufacturing capabilities need also to be took into account, the technology to produce these gpu dies cost a ridiculous amount of money.

Finally you need to take into account the software and developer landscape, the games and programs that will use your gpu. Again this is not an easy task to match the requirement of the user and at the same time being really more powerful than everything else.

Hope it help you a bit to understand how hard it is to build a gpu, even harder a successful and powerful one.
I like tho your name proposition, RX-Eclipse sounds pretty awesome for a gpu name!!

Actions

ajlueke Mar 12, 2018 8:47 AM

What it really comes down to, is that NVidia is a large operation that has the ability to produce multiple cores from a single architecture for different purposes. Back in the Fermi/Kepler days, NVidia added quite a few FP64 cores to their GPUs which gave a great deal of compute performance. Unfortunately FP64 cores a larger and more power intensive than FP32, so when playing games (which use very little FP64), there you can fit less cores on a die and the run much hotter than FP32 in gaming workloads. So the Fermi/Kepler GPU dies were massive and inefficient when compared to the Radeon 5800 series, and worse yet, workstation users started buying gaming cards instead of the much higher margin professional line.

So with Maxwell, and ultimately Pascal as well, NVidia stripped out a lot of the FP64 functionality in gaming cards, saving that for the Tesla line. That allowed them to have drastically better power efficiency, more cores, and higher clocks.

AMD on the other hand, doesn't have the resources to develop individual workstation and gaming cores. So they are stuck developing one core that can be used for both. RX Vega has twice the FP64 performance over the GTX 1080Ti, and it shows in the compute performance.

This allows AMD to use the GPU for machine learning/professional rendering/AI applications as well as gaming. But naturally, those extra FP64 cores make Vega less efficient at strongly FP32 tasks like gaming. To help combat that, AMD created the Mantle API, which strongly leverages compute functionality to deliver better gaming performance. A lot of that functionality was then built into the DX12 and especially the Vulkan spec which is why those specs tend to perform better on AMD hardware.

NVidia, reticent to add strong FP64 back into consumer cards to keep their professional margins intact, created Gameworks. A means of doing a lot of the compute style graphics enhancements within a standard FP32 workload. Naturally then those enhancements perform better on NVidia hardware.

Game developers have been slow to add a lot of the compute functionality to games, simply because of NVidia's much higher install base and market clout, in push Gameworks onto developers. AMD has had some luck working with developer Bethesda in using Vulkan in their games like Doom, and Wolfenstein II

You can see that when developers make use of those extra FP64 cores on the Vega 64, it's gaming performance rises relative to the competition. But bottom line, gaming will likely remain a strongly FP32 workload for some time, and NVidia will have a advantage here by their ability to fab out FP32 cores for gaming, while saving the designs with more FP64 cores for their Tesla line. The advantage Vega has, is that it can do both reasonabily well. It can game in the GTX 1080 ballpark, while climbing to 1808 Ti levels with Vulkan support. And while not as strong at compute as NVidia's workstation cards, it is worlds cheaper.

So in making the GPU you request, would that be all FP32 cores and just a gaming behemoth? Then what will AMD do in the professional space? Just cede that entire market to NVidia? Or if you include FP64 cores in your design, the GPU would be so large and hot, it would be difficult to keep the clocks up for decent gaming performance. Sure it would be nice if they could two separate cores the way NVidia does, but they simply don't have the resources for that. Such is life when you're the little guy.

Actions

2 Replies

wimpzilla Mar 12, 2018 4:25 AM
Hello,

Simply because it cost a lot of money, time, resources, etc to be able to design, then build and sell a powerful gpu, taking also into account the present and future technological landscape.

Not all the gpu can be designed, often you have limited design capability, there are patent to take into account, then how many people, money you are throwing into the design and R&D phase.

The same apply to the manufacturing, you have limits on how big a gpu die could be, you have efficiency limits, how many good gpu dies come out from a whole piece of silicon. The manufacturing capabilities need also to be took into account, the technology to produce these gpu dies cost a ridiculous amount of money.

Finally you need to take into account the software and developer landscape, the games and programs that will use your gpu. Again this is not an easy task to match the requirement of the user and at the same time being really more powerful than everything else.

Hope it help you a bit to understand how hard it is to build a gpu, even harder a successful and powerful one.
I like tho your name proposition, RX-Eclipse sounds pretty awesome for a gpu name!!
Like • Show 0 Likes0
Actions
ajlueke Mar 12, 2018 8:47 AM
What it really comes down to, is that NVidia is a large operation that has the ability to produce multiple cores from a single architecture for different purposes. Back in the Fermi/Kepler days, NVidia added quite a few FP64 cores to their GPUs which gave a great deal of compute performance. Unfortunately FP64 cores a larger and more power intensive than FP32, so when playing games (which use very little FP64), there you can fit less cores on a die and the run much hotter than FP32 in gaming workloads. So the Fermi/Kepler GPU dies were massive and inefficient when compared to the Radeon 5800 series, and worse yet, workstation users started buying gaming cards instead of the much higher margin professional line.

So with Maxwell, and ultimately Pascal as well, NVidia stripped out a lot of the FP64 functionality in gaming cards, saving that for the Tesla line. That allowed them to have drastically better power efficiency, more cores, and higher clocks.

AMD on the other hand, doesn't have the resources to develop individual workstation and gaming cores. So they are stuck developing one core that can be used for both. RX Vega has twice the FP64 performance over the GTX 1080Ti, and it shows in the compute performance.
This allows AMD to use the GPU for machine learning/professional rendering/AI applications as well as gaming. But naturally, those extra FP64 cores make Vega less efficient at strongly FP32 tasks like gaming. To help combat that, AMD created the Mantle API, which strongly leverages compute functionality to deliver better gaming performance. A lot of that functionality was then built into the DX12 and especially the Vulkan spec which is why those specs tend to perform better on AMD hardware.

NVidia, reticent to add strong FP64 back into consumer cards to keep their professional margins intact, created Gameworks. A means of doing a lot of the compute style graphics enhancements within a standard FP32 workload. Naturally then those enhancements perform better on NVidia hardware.

Game developers have been slow to add a lot of the compute functionality to games, simply because of NVidia's much higher install base and market clout, in push Gameworks onto developers. AMD has had some luck working with developer Bethesda in using Vulkan in their games like Doom, and Wolfenstein II

You can see that when developers make use of those extra FP64 cores on the Vega 64, it's gaming performance rises relative to the competition. But bottom line, gaming will likely remain a strongly FP32 workload for some time, and NVidia will have a advantage here by their ability to fab out FP32 cores for gaming, while saving the designs with more FP64 cores for their Tesla line. The advantage Vega has, is that it can do both reasonabily well. It can game in the GTX 1080 ballpark, while climbing to 1808 Ti levels with Vulkan support. And while not as strong at compute as NVidia's workstation cards, it is worlds cheaper.

So in making the GPU you request, would that be all FP32 cores and just a gaming behemoth? Then what will AMD do in the professional space? Just cede that entire market to NVidia? Or if you include FP64 cores in your design, the GPU would be so large and hot, it would be difficult to keep the clocks up for decent gaming performance. Sure it would be nice if they could two separate cores the way NVidia does, but they simply don't have the resources for that. Such is life when you're the little guy.
Like • Show 0 Likes0
Actions

Why not create something far more advanced than the competition?

Outcomes

Related Content

Recommended Content