AMD has developed an 8-core desktop APU

[mongeese]

A month ago, AMD released a smattering of mobile pieces to immediate success. Now they’re prototyping the same hardware in desktop form: Renoir, fully unlocked and power hungry.

Renoir is AMD’s newest product. It is a monolithic APU; it pairs eight Zen 2 cores with eight Vega compute units (equivalent to 512 shaders) on one die. Subsequently, it’s a little pricier than other recent Ryzen releases, but it’s also very efficient which has made it an excellent choice for laptops.

Renoir-based products stretch from 4-core, 4-thread pieces boosting up to 3.7 GHz to 8-core, 16-thread behemoths that spike up to 4.2 GHz. But limited by the cooling potential and power delivery of a laptop, we’re yet to see exactly how far Renoir can be pushed.

Soon that could change. A Renoir-based processor, codenamed “100-000000149-40_40/30_Y” has been found in Userbenchmark’s database. According to its codename, it should operate with a 3.0 GHz base clock and a 4.0 GHz boost clock, and during the Userbenchmark trial, it did. Its scores were about 5% above the average 4800H result which is the highest performing mobile Renoir piece. As an engineering sample, however, it is likely to be an underperforming member of its species.

Renoir Desktop

Two chips are being tested

> 3.5GHz CPU, 1750MHz iGPU
> 3GHz CPU, 1750MHz iGPU (likely the 8C Eng Sample: 100-000000149-40_40/30_Y from Userbenchmark)

I believe both are 8C

— _rogame (@_rogame) May 8, 2020

While Userbenchmark is a very poor metric of performance, it did outpace the 3400G, the current desktop Ryzen APU, by about 20% core-for-core. But the 3400G has only four cores while Renoir has eight so a fully unlocked Renoir piece should easily trounce its predecessor.

"We've actually done a tremendous amount of optimization around these graphics cores... they have 59% more performance than the previous generation."

- AMD CEO Dr. Lisa Su, at the announcement of Renoir for mobile

Indeed, where Renoir will sit within AMD’s desktop lineup poses an interesting question. APUs are generally affordable alternatives to CPU + discrete GPU combos but at eight cores, a flagship Renoir piece would offer CPU performance that’s about what a typical power user would require. At the same time, its GPU component – while very capable for an APU – won’t stand up to AAA gaming or heavy rendering workloads.

Permalink to story.

https://www.techspot.com/news/85166-amd-has-developed-8-core-desktop-apu.html

 

[HardReset]

That is targeted into businesses. Lots of CPU power and powerful GPU for computing purposes. Many businesses do not want PC's with external graphics and so far AMD only offers 4 cores with GPU. Now AMD offers 8, same as Intel while GPU is much stronger.

 

[Lew Zealand]

But does an APU need eight cores?

The Core i9-9900K is an 8-core APU. Last I looked it comes with a UHD 630 iGPU.

 

[candle_86]

It's a feature check mark, makes it easier for oems to sell

 

[neeyik]

The Core i9-9900K is an 8-core APU. Last I looked it comes with a UHD 630 iGPU.

Doesn't that make almost all of Intel's products APUs?

It's really only an AMD term, as these particular products will shift certain processing tasks automatically onto the GPU where required (known as heterogeneous computing). Intel's, on the other hand, doesn't do this as the iGPU only does dedicated graphics and video tasks.

 

[Gomos]

Actually I wish we had APUs with more than 8 cores, as they are perfect for my workloads.

 

[Dsirius]

So you're basically write an article which was resumed by neeyik:
"Doesn't that make almost all of Intel's products APUs?"
Neeyik wrote a better and shorter article than you

 

[DukeJukem]

I'm sure there are many companies out there that don't demand a high end gpu that would appreciate the other cores.

 

[Aqueltic]

The most popular GPU on amazon is the NVIDIA GT 710. The reason is people want a more powerful CPU than a budget APU and they need graphics. AMD, by launching this, is removing the problem of wanting more power but no demand for a discrete GPU. Also, the Intel parts with IGPU are not good value and their IGPU is weak compared to AMD.

 

[Joray K Joseph]

I love this, Finally getting CPU if the graphics are good enough. Will make computers not only more efficient but will be a huge boon to alot of creatives and business. Photo Editors and light video work would be great on the APU as well as business that need graphics performance for light tasks and day traders needing multiple monitors I can see this being a huge boon for them while keeping a slim format. Not to mention the possibility for Slim AIO computers that are powerful and cheap enough for most people to afford. This chip would be great for Damn near most people who use computers as well as for gamers

 

[neeyik]

So you're basically write an article which was resumed by neeyik:
"Doesn't that make almost all of Intel's products APUs?"
Neeyik wrote a better and shorter article than you

I was joking. Intel’s CPUs cannot be classed as APUs as they’re not heterogeneous.

 

[Irata]

I was joking. Intel’s CPUs cannot be classed as APUs as they’re not heterogeneous.

How heteronormative of you...

 

[Dsirius]

Your'e right, just that you pointed what is important, the fact that almost all intel processors have integrated graphics (a very weak one).

 

[Lew Zealand]

Doesn't that make almost all of Intel's products APUs?

It's really only an AMD term, as these particular products will shift certain processing tasks automatically onto the GPU where required (known as heterogeneous computing). Intel's, on the other hand, doesn't do this as the iGPU only does dedicated graphics and video tasks.

Thanks! I had no idea that's what APU meant. I always assumed it just meant that the CPU had an iGPU but not that the iGPU contributed to more than just video display. I wonder if QuickSync is an APU-type function but I assume not as it's not done automatically. Off to so some online research.

 

[Tuxie]

I wonder if this one will be powerful enough for Dolphin, Cemu and PCSX2 at 1080p 60fps. Maybe even Yuzu and RPCS3?

 

[Theinsanegamer]

"does an APU need 8 cores"

The same question was asked of all desktop APUs a few years ago, and look where we are now. More importantly, if someone wants to start with an APU and build a more powerful system later, well now they can start with an 8 core APU, and when they get a dGPU they already have an 8 core ready to go. For progressive builds its great.

You know, as long as AMD doesnt randomly decide to only support it on the A520 chipset or something equally stupid to the "no support for 4000 on previous chipsets" thing.

 

[opckieran]

Doesn't that make almost all of Intel's products APUs?

It's really only an AMD term, as these particular products will shift certain processing tasks automatically onto the GPU where required (known as heterogeneous computing). Intel's, on the other hand, doesn't do this as the iGPU only does dedicated graphics and video tasks.

What tasks specifically does AMD’s heterogeneous compute APU accelerate which Intel’s iGPU currently does not?

 

[Evernessince]

"does an APU need 8 cores"

The same question was asked of all desktop APUs a few years ago, and look where we are now. More importantly, if someone wants to start with an APU and build a more powerful system later, well now they can start with an 8 core APU, and when they get a dGPU they already have an 8 core ready to go. For progressive builds its great.

You know, as long as AMD doesnt randomly decide to only support it on the A520 chipset or something equally stupid to the "no support for 4000 on previous chipsets" thing.

Progressive builds are a good chunk of PCs. The problem with AMD's current APUs on the market is the cut down cache size which hurts gaming performance. If they made an APU with good performance that scales well as you upgrade the system, they would certainly sell more than current products. They need to price them aggressively as well. A bit more expensive than CPUs without an integrated GPU and a hit to performance makes them a hard bargain to anyone who doesn't need a temporary GPU solution.

 

[Theinsanegamer]

Progressive builds are a good chunk of PCs. The problem with AMD's current APUs on the market is the cut down cache size which hurts gaming performance. If they made an APU with good performance that scales well as you upgrade the system, they would certainly sell more than current products. They need to price them aggressively as well. A bit more expensive than CPUs without an integrated GPU and a hit to performance makes them a hard bargain to anyone who doesn't need a temporary GPU solution.

True, but the type of budget that prioritizes building an APU system and later upgrading is typically not the type buying 2080tis. If you are using a 2060 super or 2070 level performance, the difference between a 8 core with 30MB of cache vs 45 MB or whatever wont be all that noticeable, 1-2% at most.

 

[Evernessince]

Nto sure where you are going with " A bit more expensive than CPUs without an integrated GPU and a hit to performance makes them a hard bargain to anyone who doesn't need a temporary GPU solution." The expensive ones have full cache, and the cheap ones have a iGPU. There is no "cut down CPU with worse performance with no iGPU" unless you are talking about the athlons, but those are CHEAPER then the APUs, not more expensive.

Reread the comment, the noun being substituted with pronouns in the last 2 sentences was APU, thus in this example I'm referring to APUs.

 

[Mr Majestyk]

Not much interest in Renoir, but next years APU's with Zen 3 and Navi are going to be monsters and will kill off discrete GPU's for laptops IMO, for all but gaming rigs or workstation laptops that need a Quadros or FirePro.

 

[krizby]

I guess the Renoir APU is a good choice for mini ITX PC, a worthy competitor to Intel NUC. AMD is really pushing Intel on all fronts now.

 

[neeyik]

What tasks specifically does AMD’s heterogeneous compute APU accelerate which Intel’s iGPU currently does not?

That's a good question and I made a hash of discussing it. Heterogeneous processors are ones that have multiple ASICs within them, for different roles or applications - mobile phone SoCs are a perfect example of this, but this also means that Intel's CPUs and AMD's APUs are both heterogeneous.

What I should have made clear was that when AMD uses the term APU, they are specifically talking about HSA - heterogeneous system architecture. This is different to the above, as it refers far more to management of memory resources. When you play a modern 3D game on a Windows PC, all of the working graphics data exists twice in the system: one copy is in the system memory, the other copy is in the graphics card's local memory.

Any changes required to either resource requires the data to be copied across, which is why when programming games, you make sure the GPU gets everything it needs to process a frame in the first copy, set it to work on that, and then copy the results across once finished.

In a system using HSA, there's no need to physically transfer data - instead the relevant hardware gets control of the memory pointers (code that contains the address location of data) and accesses it this way. Yes, the memory still needs to be accessed over some kind of a system bus, but it's not copied to another location first (generally known as 'zero-copy').

Now you might think that this doesn't apply to Intel CPUs, as the iGPU and the CPU itself all use the same system memory. However, the iGPU is allocated its own share of the RAM, and this is treated just as if it was local memory on a discrete graphics card. In other words, data still gets copied about, even though it's physically in the same hardware. The idea behind HSA is to completely remove the necessity for this.

HSA is best suited to devices that just the one pool of memory resources, such as a laptop, office PC without a discrete GPU, or a modern console. A top end graphics card has a memory system that is far better than the system's so it's better to copy the data it needs into the local memory: you get much better performance, but you've obviously got 2 sets of data floating about.

The true ultimate goal of HSA is for it to be completely transparent to the developer: one should be able to write a block of code in, say, Python and the computer should then automatically figure out what device is best for processing the work. This clearly requires software support, from the operating system, through to instruction sets and device drivers.

And this is where things become somewhat disappointing - at the moment, if you want to use your GPU to compute work in Windows, you're stuck with having to work directly with the likes of OpenCL, Vulkan, DirectML, CUDA, or using AMD's GPUOpen toolkits. If one used OpenCL, for example, then iGPU in Intel's Core processors will be utilised just as much as the Vega GPU will be in AMD's APUs, but as they're not HSA processors, there will be a degree of performance cost transferring data from the allocated iGPU memory back and forth to the rest of the CPU memory.

 

[bluetooth fairy]

That's a good question and I made a hash of discussing it. Heterogeneous processors are ones that have multiple ASICs within them, for different roles or applications - mobile phone SoCs are a perfect example of this, but this also means that Intel's CPUs and AMD's APUs are both heterogeneous.

What I should have made clear was that when AMD uses the term APU, they are specifically talking about HSA - heterogeneous system architecture. This is different to the above, as it refers far more to management of memory resources. When you play a modern 3D game on a Windows PC, all of the working graphics data exists twice in the system: one copy is in the system memory, the other copy is in the graphics card's local memory.

Any changes required to either resource requires the data to be copied across, which is why when programming games, you make sure the GPU gets everything it needs to process a frame in the first copy, set it to work on that, and then copy the results across once finished.

In a system using HSA, there's no need to physically transfer data - instead the relevant hardware gets control of the memory pointers (code that contains the address location of data) and accesses it this way. Yes, the memory still needs to be accessed over some kind of a system bus, but it's not copied to another location first (generally known as 'zero-copy').

Now you might think that this doesn't apply to Intel CPUs, as the iGPU and the CPU itself all use the same system memory. However, the iGPU is allocated its own share of the RAM, and this is treated just as if it was local memory on a discrete graphics card. In other words, data still gets copied about, even though it's physically in the same hardware. The idea behind HSA is to completely remove the necessity for this.

HSA is best suited to devices that just the one pool of memory resources, such as a laptop, office PC without a discrete GPU, or a modern console. A top end graphics card has a memory system that is far better than the system's so it's better to copy the data it needs into the local memory: you get much better performance, but you've obviously got 2 sets of data floating about.

The true ultimate goal of HSA is for it to be completely transparent to the developer: one should be able to write a block of code in, say, Python and the computer should then automatically figure out what device is best for processing the work. This clearly requires software support, from the operating system, through to instruction sets and device drivers.

And this is where things become somewhat disappointing - at the moment, if you want to use your GPU to compute work in Windows, you're stuck with having to work directly with the likes of OpenCL, Vulkan, DirectML, CUDA, or using AMD's GPUOpen toolkits. If one used OpenCL, for example, then iGPU in Intel's Core processors will be utilised just as much as the Vega GPU will be in AMD's APUs, but as they're not HSA processors, there will be a degree of performance cost transferring data from the allocated iGPU memory back and forth to the rest of the CPU memory.

If it's mostly theory or marketing term, the original question raised is staying the same. I'd add another one: is there any widely used software benifitting from AMD's APU's HSA implementation? In other words, is there real difference between iGPU and APU usage now?

But probably HSA is not the main reason for discussion, started by Mongeese, because AMD's APUs were mid-tier products at best, while 8 core chip is still considered as top-end one. AMD has created our (consumer) view on their APU products, and now it feels strange when (if) AMD tries to change their product placement without thinking about shifts in our views on their APUs.

 

[neeyik]

If it's mostly theory or marketing term, the original question raised is staying the same. I'd add another one: is there any widely used software benifitting from AMD's APU's HSA implementation? In other words, is there real difference between iGPU and APU usage now?

Probably games: a lot of rendering process is very heavy on memory bandwidth, (which is why GDDR5 memory modules are always 32 bits wide, compared to DD4 used in system memory which are only 8 bit). Any measures available to reduce the load on the memory bus will always be beneficial. However, unless the developers have actively stated they're using AMD's GPUopen libraries and SDKs, then it would be difficult to tell for sure.

The biggest hurdle to identifying how beneficial AMD's approach is lies in the fact that it involves APUs and mobile processors - you just can't grab one by itself and test it in isolation; there's always the rest of the laptop that gets in the way and AMD and Intel take different approaches to the graphics part of their processors in a given price sector.

If we take the review of the Ryzen 3 2400G with its RX Vega 8 iGPU, you can see it's substantially better than Intel's UHD Graphics 630:

But the Vega 8 has 512 shader units, 32 TMUs, and 8 ROPs. The 630 has 192 shader units, 16 TMUs, 8 ROPs - so it's compute power is way down on the Vega. Any memory advantage the Ryzen has isn't going to be obvious. However, the GT 1030 the has 384 shader units, 24 TMUs, and 16 ROPs so it's not quite the same either (it's significant ROP count advantage is likely to be main reason why it's better here).

Unless Intel ever makes a CPU with iGPU that has the roughly the same internal configuration of an AMD APU, we're never going to be able to tell just how much (if any) benefit the HSA is giving the APU.