Are Graphics Cards superior to CPUs or is it all just smoke and mirrors?

[TheDarkKnight]

I've recently purchased an AMD 2200G CPU and I think it's a pretty good chip overall (bang for buck). But I am having issues getting my DDR4-3000 memory to run at it's top-rated effective speed (2934MHz, according to ASRock and AMD) for my CPU and motherboard combo.

I went back and forth about whether I should just spend $150 on a discrete graphics card and be done with it or go with the iGPU since I already had purchased my fast DDR4 memory much earlier. So I went with the APU system. The system worked fine @2133MHz OOTB but It definitely hasn't been a plug-and-play experience when it comes to getting my top-rated effective speeds out of my new system. Which is of utmost importance for these APUs.

So, I was looking at buying a discrete RX 560 and I seen that the effective memory speed is 7000MHz and I thought to myself, "Holy Moses, that's fast". But the memory itself is running (in reality) at anywhere from 1200MHz to 1750MHz. So to get to 7000MHz there's a factor of 4 involved. And that sounds very similar to what have today on the CPU side. We have quad-core CPUs.

Along with using GDDR5 memory, how is this effective speed achieved. What's the X factor of 4? In other words, are GPUs really better at graphics than a CPU could be or are they just doing things in parallel, like 4 bozos might on an assembly lane in a factory. The same operations all day long but in a group of 4?

 

[whm1974]

I've recently purchased an AMD 2200G CPU and I think it's a pretty good chip overall (bang for buck). But I am having issues getting my DDR4-3000 memory to run at it's top-rated effective speed (2934MHz, according to ASRock and AMD) for my CPU and motherboard combo.

I went back and forth about whether I should just spend $150 on a discrete graphics card and be done with it or go with the iGPU since I already had purchased my fast DDR4 memory much earlier. So I went with the APU system. The system worked fine @2133MHz OOTB but It definitely hasn't been a plug-and-play experience when it comes to getting my top-rated effective speeds out of my new system. Which is of utmost importance for these APUs.

So, I was looking at buying a discrete RX 560 and I seen that the effective memory speed is 7000MHz and I thought to myself, "Holy Moses, that's fast". But the memory itself is running (in reality) at anywhere from 1200MHz to 1750MHz. So to get to 7000MHz there's a factor of 4 involved. And that sounds very similar to what have today on the CPU side. We have quad-core CPUs.

Along with using GDDR5 memory, how is this effective speed achieved. What's the X factor of 4? In other words, are GPUs really better at graphics than a CPU could be or are they just doing things in parallel, like 4 bozos might on an assembly lane in a factory. The same operations all day long but in a group of 4?

Yes GPUs are usually much better at graphics then CPUs are. However the 2200G comes close to the 1030 dGPU so you would want a 1050 or better yet the 1050Ti.

 

[TheDarkKnight]

Yes GPUs are usually much better at graphics then CPUs are. However the 2200G comes close to the 1030 dGPU so you would want a 1050 or better yet the 1050Ti.

Okay yes. They are faster. But by what mechanism? Is it all about the GDDR5 memory bandwidth? What's the other factor, parallelism? I'm just trying to understand why CPUs have such a hard time keeping up with discrete GPUs. And to be honest, it seems AMD is closing the gap here with the most recent APU releases. I know it's still a long way behind and I'm trying to understand how they overcome these gaps.

 

[whm1974]

Okay yes. They are faster. But by what mechanism? Is it all about the GDDR5 memory bandwidth? What's the other factor, parallelism? I'm just trying to understand why CPUs have such a hard time keeping up with discrete GPUs. And to be honest, it seems AMD is closing the gap here with the most recent APU releases. I know it's still a long way behind and I'm trying to understand how they overcome these gaps.

Well APUs have to share bandwidth with both the CPU and iGPU portions and of course die space is split between graphics and processors cores. While dGPUs have the whole to themselves.

 

[DXDiag]

Okay yes. They are faster. But by what mechanism?

By fixed hardware functions, GPUs fetch textures from memory with specific Texture Mapping units, they also do triangle setup through fixed Rasterizer and Tessellation units, they also do a lot of frame buffer and Anti Aliasing operations through specific Raster Out units and then they have A LOT of FP32 cores, several times more than any CPU, which perform shader and compute operations. CPUs don't have any of that, so they render graphics VERY slowly.

it seems AMD is closing the gap here with the most recent APU releases. I know it's still a long way behind and I'm trying to understand how they overcome these gaps.

AMD is closing nothing, no company is, no APU that exists today can approach the capabilities of dGPUs. dGPUs have separate ultra fast memories, they have thousands of cores and hundreds of fixed function units, no APU is capable of holding that because they have to share die space with a general purpose CPU while dGPUs dedicate their space only to graphics. The fastest GPU available now is the Titan V, and this one is at least dozens of times more powerful than a Ryzen 2400G.

 

[Carfax83]

AMD is closing nothing, no company is, no APU that exists today can approach the capabilities of dGPUs. dGPUs have separate ultra fast memories, they have thousands of cores and hundreds of fixed function units, no APU is capable of holding that because they have to share die space with a general purpose CPU while dGPUs dedicate their space only to graphics. The fastest GPU available now is the Titan V, and this one is at least dozens of times more powerful than a Ryzen 2400G.

Yeah, I remember when the first APU was released years ago, and there was similar talk on these forums and many others about how the APU was eventually going to replace the dGPU by virtue of everything being on the same die. I scoffed at it back then, and I still scoff at it right now.

The only thing APUs will be better at is performance per watt. For raw performance though, nothing beats discrete parts.

 

[TheDarkKnight]

By fixed hardware functions, GPUs fetch textures from memory with specific Texture Mapping units, they also do triangle setup through fixed Rasterizer and Tessellation units, they also do a lot of frame buffer and Anti Aliasing operations through specific Raster Out units and then they have A LOT of FP32 cores, several times more than any CPU, which perform shader and compute operations. CPUs don't have any of that, so they render graphics VERY slowly.


AMD is closing nothing, no company is, no APU that exists today can approach the capabilities of dGPUs. dGPUs have separate ultra fast memories, they have thousands of cores and hundreds of fixed function units, no APU is capable of holding that because they have to share die space with a general purpose CPU while dGPUs dedicate their space only to graphics. The fastest GPU available now is the Titan V, and this one is at least dozens of times more powerful than a Ryzen 2400G.

Well, that's all good stuff. But let's be honest here. Discrete graphics have a lot of what you are talking about also. The TFLOPs capabilities on an AMD 2400G are just slightly above the original PS4's capabilities. I'm not saying that integrated graphics are gonna crush discrete GPUs anytime soon. I know they are a special device for a special purpose. But I still don't think they're necessarily using any black magic to beat down CPUs. When the memory speed is reported as an effective rate of 7000MHz and then you look at base clock you can see it's a factor of the base clock. So, it's not black magic, it's as you said, putting a whole lot more of the same things on a separate device. It's basically allocating more of the same resources to the same operations a CPU would/could be doing. That's basically it.

I'm not arguing the performance gaps aren't there. But I think it's a matter of choice and convenience, as most things are, about whatt people want to buy. But there is nothing special per se about discrete graphics capabilities other than GDDR5. Would you at least admit this is the one of the most powerful tools in the discrete graphics cards arsenal? At least you did admit to the performance per watt being better on iGPUs. That's nice.

 

[whm1974]

Well, that's all good stuff. But let's be honest here. Discrete graphics have a lot of what you are talking about also. The TFLOPs capabilities on an AMD 2400G are just slightly above the original PS4's capabilities. I'm not saying that integrated graphics are gonna crush discrete GPUs anytime soon. I know they are a special device for a special purpose. But I still don't think they're necessarily using any black magic to beat down CPUs. When the memory speed is reported as an effective rate of 7000MHz and then you look at base clock you can see it's a factor of the base clock. So, it's not black magic, it's as you said, putting a whole lot more of the same things on a separate device.

I'm not arguing the performance gaps aren't there. But I think it's a matter of choice and convenience, as most things are, about whatt people want to buy. But there is nothing special per se about discrete graphics capabilities other than GDDR5. Would you at least admit this is the one of the most powerful tools in the discrete graphics cards arsenal? At least you did admit to the performance per watt being better on iGPUs. That's nice.

Not do dGPUs have GDDR5 for more bandwidth they are also larger, faster, and have more features then any iGPU will be or have. I think that at best iGPUs might reach 1050Ti performance levels, but that might be pushing it.

 

[Flapdrol1337]

Along with using GDDR5 memory, how is this effective speed achieved. What's the X factor of 4? In other words, are GPUs really better at graphics than a CPU could be or are they just doing things in parallel, like 4 bozos might on an assembly lane in a factory. The same operations all day long but in a group of 4?

Graphics is a somewhat specific task so you can make specialized hardware for it. This hardware will perform better in terms of performance/chip size and performance/watt.
Graphics is also easily multithreaded, since chips are more efficient at low speeds (by allowing lower voltage) you can expoit this by making a slow chip with a massive number of "cores".
Since the chip doesn't have to be fast you can tweak the manufacturing process to make the transistors of the chip more efficient at the price of them maybe not working at all at high speed.

These all multiply to make a chip that is massively more efficient at graphics, yet the a 1080Ti still burns over 200W. That means there's a lot of extra performance to be had.

 

[SirDinadan]

dGPUs have have more features then any iGPU will be or have

This is simply not true. The new 2200G and 2400G is on par with the Vega dGPUs feature wise. It is an entirely different matter that hardly any game or compute app can take advantage of the new tech - an-age old problem for AMD.

 

[sontin]

The only thing APUs will be better at is performance per watt.

That isnt true either. Biggest problem is memory transfer. It takes much longer to get information from the system memory than from dedicated GDDR which sits much closer to the GPU.
Thats one of the reason why Raven Ridge iGPU cant compete with MX150 when it comes to power consumption.

 

[Carfax83]

That isnt true either. Biggest problem is memory transfer. It takes much longer to get information from the system memory than from dedicated GDDR which sits much closer to the GPU.
Thats one of the reason why Raven Ridge iGPU cant compete with MX150 when it comes to power consumption.

I guess it depends on whether we're talking about GPGPU or gaming.

Memory transfers isn't really a problem for gaming, because discrete GPUs are designed to hold as much data close to them as possible to minimize and mask latency. That combined with a well designed 3D engine that has a good streaming algorithm, and memory transfers aren't problematic at all; even in large open world games.

In fact, PC games typically have larger and more detailed textures than their console counterparts, and longer draw distances.

But for GPGPU and HPC stuff, then yeah I agree it is a problem because the datasets are much, much larger and there is more communication between the CPU and the GPU. But there are solutions for that as well in the form of high speed and high bandwidth interconnects like NVLink.

 

[Blitzvogel]

Along with using GDDR5 memory, how is this effective speed achieved. What's the X factor of 4? In other words, are GPUs really better at graphics than a CPU could be or are they just doing things in parallel, like 4 bozos might on an assembly lane in a factory. The same operations all day long but in a group of 4?

Graphics processors are all about simple easily parallelized operations. These days we're talking thousands of bozos, not just 4. You may know them as "stream processors" (AMD) or "CUDA cores" (Nvidia), each of them being a 32 bit FMAD (I think) ALU grouped into vector/SIMD engines called "Compute Units" for AMD or "Streaming Multiprocessor" for Nvidia.

That's pretty simplistic, but you have the right idea about extreme parallelism. Floating point units since the integration of them on CPUs have continued to grow larger, able to crunch numbers 64 bits in precision, but store 128, 256, and even recently 512 bits worth of numbers (16 x 32 bit or 8 x 64 bit) and operate on them in certain ways in a single clock. Having such large and capable capabilities requires alot of transistors while being integrated into a very large complete CPU core. That total floating point/SIMD capability goes unused in many of today's games since they have to be written to use it. Often enough that specific type of CPU performance is not the bottleneck, it's the GPU, when using a modern recent CPU.

CPUs do have the benefit of much higher clock speeds than graphics processors, but the extreme number of 32 bit ALUs more simplistic available methods of calculation in GPUs by far wins out. Less number of transistors are required to get the same GFLOPS. The ironic trend is that GPUs have been integrating more CPU like features that past 11 years, but usually it's in lieu of supplementing the high floating point performance for specific uses that need a bit of branch prediction or more programmability.

 

[whm1974]

I'm wondering if DDR5 and 7nm manufacturing process could make for more powerful iGPUs, so maybe up 1050/1050Ti levels? Or would they need some memory on the APU/CPU package dedicated to the iGPU?

 

[2is]

Okay yes. They are faster. But by what mechanism? Is it all about the GDDR5 memory bandwidth? What's the other factor, parallelism? I'm just trying to understand why CPUs have such a hard time keeping up with discrete GPUs. And to be honest, it seems AMD is closing the gap here with the most recent APU releases. I know it's still a long way behind and I'm trying to understand how they overcome these gaps.

The gap is certainly not closing unless you're looking at improvements in iGPU and comparing them with dedicated cards from 5 years ago. The gap is huge. That said, iGPU's have improved a great deal over the years and for the vast majority of consumers, are plenty capable on their own. Gamers still need a GPU. As far as what makes it faster, it's a bit of everything really. Access to much faster and dedicated pool of ram, higher clock speeds, more processing units. Need to remember that GPU's have 250watts to play with all on their own (or more if you want to exceed the standard)

It's been a while since I've checked the specifics on CPU's but last time I did I remember most being at around 100 watts with the upper end being 140, and that's split between the CPU and iGPU.

 

[Guru]

Well, that's all good stuff. But let's be honest here. Discrete graphics have a lot of what you are talking about also. The TFLOPs capabilities on an AMD 2400G are just slightly above the original PS4's capabilities. I'm not saying that integrated graphics are gonna crush discrete GPUs anytime soon. I know they are a special device for a special purpose. But I still don't think they're necessarily using any black magic to beat down CPUs. When the memory speed is reported as an effective rate of 7000MHz and then you look at base clock you can see it's a factor of the base clock. So, it's not black magic, it's as you said, putting a whole lot more of the same things on a separate device. It's basically allocating more of the same resources to the same operations a CPU would/could be doing. That's basically it.

I'm not arguing the performance gaps aren't there. But I think it's a matter of choice and convenience, as most things are, about whatt people want to buy. But there is nothing special per se about discrete graphics capabilities other than GDDR5. Would you at least admit this is the one of the most powerful tools in the discrete graphics cards arsenal? At least you did admit to the performance per watt being better on iGPUs. That's nice.

Graphical DDR is different than DDR. In fact GDDR5 runs something like 15x times faster than your average DDR4. Its not simply about MHz, the way DDR and GDDR is built is different, the connections are different, the bus width is different, the latency is different, etc... its a completely different beast.

GPU's are obviously superior to CPU's in graphic workloads.

 

[cfenton]

APUs have a few important limitations. Tthey need to fit in a standard CPU slot and they need to stay in a reasonable power budget. This means that they have to be a certain size and be usable with a reasonable cooling solution. The most powerful APU at the moment is the custom unit in the Xbox One X. That thing competes with something like a 1060 or 580. However, you can't put that in a standard CPU socket and keep it cool. It has a custom cooling system designed specifically to cool that chip. You would need a massive air cooler, or more practically, a water cooler. Most people don't have those things and would rather just buy a dGPU and not have to worry about it. Having two discrete chips spread out on a board also means you spread the heat around. So, it's less that AMD couldn't release a massive and powerful APU, but that it would be a very niche product. The big advantage of an APU over a CPU + dGPU solution is size and power usage, but if you'd need a giant cooler to keep a high-power APU cool, then those advantages are mostly lost.

 

[2is]

Am I the only one scratching my head, trying to figure out exactly what the OP is trying to accomplish/prove here?

 

[Headfoot]

"Superior" floating out there with no specificity as to what it would be superior in = troll thread. As soon as you put a specific task to the question the answers are easy and straightforward. Are GPUs superior in memory bandwidth -> generally, yes. Are GPUs superior at gaming -> yes. Are GPUs superior at serial, branchy code -> no.

 

[whm1974]

Am I the only one scratching my head, trying to figure out exactly what the OP is trying to accomplish/prove here?

Trying justify not buying a dGPU? Not hard to do with today's prices.

 

[Grubbernaught]

Well, that's all good stuff. But let's be honest here. Discrete graphics have a lot of what you are talking about also. The TFLOPs capabilities on an AMD 2400G are just slightly above the original PS4's capabilities. I'm not saying that integrated graphics are gonna crush discrete GPUs anytime soon. I know they are a special device for a special purpose. But I still don't think they're necessarily using any black magic to beat down CPUs. When the memory speed is reported as an effective rate of 7000MHz and then you look at base clock you can see it's a factor of the base clock. So, it's not black magic, it's as you said, putting a whole lot more of the same things on a separate device. It's basically allocating more of the same resources to the same operations a CPU would/could be doing. That's basically it.

I'm not arguing the performance gaps aren't there. But I think it's a matter of choice and convenience, as most things are, about whatt people want to buy. But there is nothing special per se about discrete graphics capabilities other than GDDR5. Would you at least admit this is the one of the most powerful tools in the discrete graphics cards arsenal? At least you did admit to the performance per watt being better on iGPUs. That's nice.

By "discrete graphics" I assume you are referring to the GPU cores on an APU.

So you have to understand this IS a cut down GPU die that first appeared on dedicated cards.

Very roughly, these have been shrunk sufficiently and given a new memory interface to fit into the same package as the CPU.

So it's essentially exactly the same as running a bottom end dedicated card.

But you can scale up that die size substantially when you are not constrained for size (dedicated cards), so you end up with nearly an order of magnitude more "cores" (really bad way to put it with GPUs).

They can run hotter because there's no CPU heat to dissipate either.

Don't be too concerned about the memory type or speed, that is part of the equation which gives you what you actually care about when it comes to memory. ....bandwidth

Men bandwidth is import for several reasons, but a simple one is that it is going to limit your performance at high resolutions.

4k60fps gaming is not possible without plenty of it.

So dedicated cards are always going to I be faster than APU's because they are bigger, hotter, faster and have more memory bandwidth.

You are technically right in calling out the APUs GPU components to being essentially the same as a bottom tier card.

The top end is scaled up massively.

Confuse none of this with CPU software rendering, then you are heading into apples and oranges territory.

 

[lakedude]

Yeah, I remember when the first APU was released years ago, and there was similar talk on these forums and many others about how the APU was eventually going to replace the dGPU by virtue of everything being on the same die. I scoffed at it back then, and I still scoff at it right now.

Yeah, back in the 3dfx days I scoffed at the idea of having a 2D video card and 3D card combined into one.

Remember the Celeron 300A? Instead of a larger external cache they added a smaller but much faster cache on die and the result was fabulous.

I'm not saying an APU is better than a dGPU but for a typical surfer or casual gamer the APU has come a long way. The APU has made the $100 video card redundant.

If you look at the trend you can see the APU taking a larger and larger chunk out of the low end dGPU market.

 

[lakedude]

Am I the only one scratching my head, trying to figure out exactly what the OP is trying to accomplish/prove here?

Yes, but I think he is asking a legitimate question, not being a troll...

"Superior" floating out there with no specificity as to what it would be superior in = troll thread.

I think the OP has the notion that the APU is "superior" but at the same time wants a lesson on what makes things fast inside a computer. This is a difficult concept to grasp or even to teach. I know from years of experience and tinkering but putting 35 years worth of knowledge into a post is a bit tricky.

 

[NTMBK]

If you want to make an APU as powerful as a a discrete GPU, you need to give it a power budget as big as a discrete GPU. There's a reason why the PS4, PS4 Pro and XBox One X have heatsinks that look a lot like a GPU cooler, with a big blower fan pushing air through the fins:

Sadly, the standard PC ATX design is complete garbage, and not designed for this style of high performance coolers. It's a 20+ year old design built around 35W CPUs from the 90s, with their dinky little heatsinks and low airflow requirements. It's completely atrocious at getting cold air to the CPU socket, and then exhausting hot air out the back of the case. Instead everything sits in a warm soup, with hot air swirling around inside the case. That's why people have resorted to insane hacks like watercooling, with radiators bolted onto the side of the case so they can actually access some fresh cold air to cool their CPU.

If you really want high performance APUs, you either need a fully custom case design (like the consoles), or you need the industry to ditch ATX/uATX/mITX and move to a replacement that specifies a much better socket layout and airflow. Otherwise you'll be limited to ~95W for the entire CPU & GPU combined.

 

[DXDiag]

When the memory speed is reported as an effective rate of 7000MHz and then you look at base clock you can see it's a factor of the base clock. So, it's not black magic

That's the whole concept of DDR, it literally HAS to work in a factor of the base clock, the effective clock rate is actually deceptive, as manufacturers market the effective bandwidth as clock rate, It's NOT clock rate. It's just bandwidth capacity, it should be written like this: 7000MT/s, which means mega transfer per second.

here is how it goes in a very simplistic manner, take an example of a marketed memory speed of 400MHz:

SDR pumps one bit per clock, so base clock is 400MHz, so bus clock is also 400MHz, effective rate is 400MT/s.
DDR pumps two bits per clock, base clock is 200MHz, bus clock should also be 200MHz, effective rate is 400MT/s.
DDR2 pumps four bits per clock, base clock is 100MHz, bus clock is 200MHz, effective rate is also 400MT/s.
DDR3 pumps eight bits per clock, base clock is actually a stunning 50MHz, bus clock is still 200MHz, effective rate is 400MT/s.
DDR5X pumps sixteen bits per clock, base clock is actually an incredible 25MHz, bus clock is still 200MHz, effective rate is 400MT/s.


Do you see what's happening here? SDR had equal base and bus clocks, with equal effective rate (all 400s), however, DDR changed things in a big way, so that for a given speed (400), each generation of DDR effectively halves the base clock of the previous generation, maintains a fixed bus clock of half the effective rate (half the 400 is 200). Thus each generation saves power, voltage and durability for exactly the same effective rate of 400MT/s in my example.