AMD 8000G APUs launch Jan 31, $329 8700G competes with desktop GTX1650 in gaming | Hacker News Reader

AMD 8000G APUs launch Jan 31, $329 8700G competes with desktop GTX1650 in gaming

videocardz.com

63 pointsnateb202267 comments

Am I the only one disappointed in AMD for still not releasing a proper PC gaming APU and instead they keep trickling down these lukewarm parts to us like it's the best thing since slice bread when they already have the monster X86 Ryzen-RDNA APU from the PS5 and Xbox released 3 years ago, which slap these new PC APUs into oblivion at gaming.

So why can't they just strip the proprietary Sony/Microsoft IP from the silicone and release a PC version of those APUs?

They're still just bog standard Ryzen-RDNA parts with more graphic CUs and higher TDP than what they sell for the PC, and the PC motherboards can take higher TDPs of the PS5/Xbox APUs anyway. So what gives AMD? Why you playing us like that?

It's more complicated it isn't that they can't. Desktop memory is still behind GPU memory in bandwidth. So making a GPU greater than a certain footprint looses efficiency. However you could probably achieve low-midrange dGPU performance.

Apple seems to be doing a pretty good job, AMD can figure it out.

Apple's SoCs aren't running against consumer DDR5 slapped into slots on a motherboard

That's the point, though. Apple are doing a pretty good job because they're using unified memory on the same chip. When AMD gets the same opportunity -- e.g.: PS4/PS5 and XBO/XSX APUs, Steam Deck APU, Z1 Extreme, etc. -- they get good results because they control the memory layout, capacity, and speeds. (And end up deploying VRAM as chip RAM, essentially)

What exactly is “good results”?

The fact that the PS5 and Xbox Series X have a great GPU, despite being an APU, merely because they can package the RAM along with the chip and not rely on slot-in DDR5? Look up the benchmarks and compare to dGPUs.

Dedicated desktop GPUs are significantly more powerful than PS5 and XSX chips - even GPUs released in the same year.

Nvidia’s Ampere series released in 2020.

You might not see the huge difference in performance because games are optimized around consoles. But in terms of raw specs/tflops, dedicated GPUs are much more powerful.

That said, PS5 and XSX take the cake in cost/perf. But that’s only because they are subsidized.

Because PS5/XSX SoCs are way too power hungry for laptops and too impractical for desktops. They’re designed to be cheap to produce for a console which means they use fewer transistors and run in high clock speed leading to higher power and cooling requirements than laptops can handle. Discrete Nvidia GPUs blows it out of the water on desktop. So do standalone desktop CPUs.

Furthermore, they use unified GDDR RAM which has high bandwidth but also high latency. The CPU would not be competitive. This is unlike Apple Silicon which takes the more expensive route by using LPDDR5 with many channels.

I think what you want is an efficient SoC with a very large GPU. Basically an M3 Max. But Apple is way ahead of AMD at making mobile SoCs. It would also cost $3000+ starting - which brings us back to the territorial of Nvidia GPUs having more value if you’re a gamer.

With that said, I think chip makers were woken up by Apple or caught with their pants down. I do think we will see bigger GPUs in SoCs on laptops and desktops. Nvidia is rumored to be making an Apple Silicon Competitor. Qualcomm is with their Nuvia based chips. AMD and Intel are putting bigger and bigger GPUs into their SoCs.

>So why can't they just strip the proprietary Sony/Microsoft IP from the silicone and release a PC version of those APUs?

They could, but it would run like shit. DDR4 is way too low bandwidth for video memory. Zen2 was a pretty slow architecture to begin with; the ancient 14nm Intel CPUs are faster.

For those keeping score: consoles are getting behind AMD's current tech -- a generation behind on the memory (DDR4 vs 5), 10 nm behind on the CPU fab (14nm vs 4nm), and two generations behind on the zen architecture (Zen 2 vs Zen 4). I guess that means we are due for a new console generation in the next year or two? Which will probably be based on this faster, better 8000G series (or the next one).

PS5 and Xbox Series X use GDDR6. PS5 has 512MB of DDR4 that it uses as an SSD cache, but everything else is done using GDDR6. Both are fabbed on 7nm processes, and later revision PS5s use a 6nm process.

They are just entering their 4th year. We won't see a new console gen for probably 3+ years. They've barely started getting non-cross gen games in any decent number. We might see a PS5 Pro, but it won't be a huge leap.

PS5 and XSX are significantly faster than this chip in GPU performance. Not even close. This 8000 series is using 780M which is a rebrand of 680M released a while ago.

Pretty much everything you said is wrong.

I believe they never release good APUs because they would cut into the market for video cards... So the best APU has to be inferior to a mediocre GPU combo.

I don't know that there's much of a market for a desktop APU. People who don't care about graphics can use the CPUs with built-in graphics; most people who do care about graphics are likely going to prefer a dedicated card, possibly from Nvidia anyway. There just aren't a lot of people looking for a non-upgradable desktop

The 8700G should be roughly as powerful as an Xbox Series S. These modern APUs have come a long way. I suspect they don't bother making them much more powerful because the cost/performance ratio can't compete with the standard CPU+GPU combos.

These are pretty beastly APUs, imo.

To go further, iGPUs need much more ram throughput than they have now. 128-bit ram is not enough!

I'd love love love to see an m1 style on-package ram, but I don't think it's likely (in the next couple years). What is happening though - if things go ok - is Strix Halo.

This announcement today is a desktop Hawk Point (same specs but higher power than existing mobile Hawk Point: 8c Z4, 16MB L3, 12 CU 780m GPU). Next APU due end of year or early next year is Strix Point, on Zen5+??. The real excitement though: there's supposed to be a "big" Strix Point called Strix Halo coming out a bit after that is alleged to boast a double wide 256b DDR5 memory interface. That doubled-up memory bandwidth should provide the headroom to go from good to great!

I really hope we get some ridiculous gaming decks with Strix Halo. Just turn off half the chip when on battery, but plug it on, put it on an actively cooled dock, and bamo, monster beast system. Hey, a nerd can dream, right?

Most interesting to me is the built-in NPU / "AI Accelerator" / Ryzen AI / XDNA. While descriptions of the chip are relatively easy to find, I haven't yet found out how the thing is programmed. Hetergeneous programming across APU, GPU, and NPU is going to be interesting.

Update: https://ryzenai.docs.amd.com/en/latest/

It’s basically an NPU, which has been shipping on phones for 5 years already.

300euros for the 8700g, perfect for a mini-PC to play dota2 on my custom linux distro.

I would push the envelop for a dirt cheap motherboard, without network or audio chips, only many USB-C high-speed ports (and dongles for those). I guess the power supply or battery pack will be the most expensive component after the APU and before the RAM.

We are talking <500 euros serious gaming mini-PC.

I'm more interested in the custom linux distro part, was it hard to setup?

I used to dabble a lot with Linux and its how I got my career in tech actually.

Closest I got to something like that was a gentoo distro, although I think I messed something up because I wasn't able to boot into it after all that setup.

I have 2 decades of linux system dev programming.

Most of open source SDKs are abominations and code is being literaly brain fucked by ultra-complex syntax languages like c++, rust and similar. I am disgusted, to a point for me, open source software is now hardly less worse than msft or apple trash.

I have open source software only because I can still customize it enough to let me bear with it and it is free as in free beer.

That to say, the amount of insane and broken knowledge you need to properly maintain a fully custom linux distro (very custom) is obscene.

I am trying to find ways out of some of the critical issues for myself, basically trying to move a mountain with a spoon.

Audio and ethernet are probably cheaper onboard than with whatever you need to enable usb-c and an external dongle.

The real way to get a cheaper board is going chipsetless; something knoll3 like this [1], but without the fancy bits. Single 1g ethernet, no ipmi, no oculink, and the price probably comes way down. Although, I do see there's a MSI PRO A620M-E Micro ATX out there for $75, so maybe we can have a chipset after all.

[1] https://www.asrockrack.com/general/productdetail.asp?Model=A...

We don't need a chipset at all: we have the APU USB and NVMe disk. RAM slots, PCIE slots, M.2 slots and a bunch of usb-c connectors, power connectors, an UEFI bios, and we are good to go.

But for the normal users, yeah, an audio chip and a Gbits network chip would be appropriate (probably routed directly to some APU PCIE pins).

You can easily achieve a 500 dollar Dota PC if you're willing to buy used. I bought a 5600x / 32GB RAM / RTX 3070 PC for 550 USD a bit ago.

I'm still a little sad they aren't at least allowing 90W parts. More power breathing room and maybe another 4-8 GPU CUs would make for a compelling entry tier gaming rig.

These parts could also be interesting for NAS and other home server setups if the monolithic die leads to better idle power usage and they haven't lost ECC support. Unfortunately getting verified ECC support in AM5 seems to be quite difficult. In AM4 it was easy but ECC on APUs was only enabled on the PRO parts which were only sold to OEMs.

It's worth noting AMD already included minimal GPUs in the desktop line for the Zen 4 generation, including video encode/decode. These, with the exception of the 8300G model, simply have larger GPUs but smaller cache and this might not be a good trade off for a NAS. The exception is the 8300G which is a good budget option having only 4 cores and no parallel in the desktop line.

AM5 has iGPU in all parts which makes it easier to use as server indeed. But normal AM5 parts are chiplet designs so they idle at relatively high power. These are monolithic dies from the laptop line so they have that going for them. According to the AMD website ECC support is enabled in these too:

https://www.amd.com/en/product/14066

It seems AM5 motherboards are back to having ECC confirmed in the specs after AMD AGESA updates:

https://www.asrock.com/MB/AMD/B650M%20Pro%20RS/index.asp#Spe...

So pending some benchmarks 8700G+B650+ECC sounds like a good home server that should idle at low power.

Oh wow I thought they abandoned the monolithic APUs in newer generations as well. I guess that does help then.

What's the easiest way to cheaply get ECC on AM5 working w/ separate GPUs nowadays? I currently have an ASRock B450 AM4 motherboard + ECC memory + Ryzen 5 3600 and while I think ECC works, I'm not certain.

I with there were a "cheap ECC on <AMD newest socket>" guide where the components and settings were validated by a reputable party.

As someone with four Ryzen 7 PRO 5750GEs, I’ll say I’m looking forward to whenever the Ryzen 7 PRO 8750GEs exist. :)

In raw CPU performance, I found the 5750GE was 90-92% of the CPU performance of a 5700X, but capped out at a total package power of just 39W. 5750GEs made for fantastic home server CPUs, so here’s hoping there’ll be an 8750GE later this year (plus some ASRock Rack mATX motherboards with ECC support and BMCs).

I agree the OEM only was a pain. I had to get my 5750GEs through mostly disassemblers/recyclers, though some boutique shops like QuietPC buy a half tray of them at a time to sell directly to end users, so keep on the lookout!

Seems like there's no need to wait for the PRO versions as the 8700G has ECC support:

https://www.amd.com/en/product/14066

Power can probably be limited in the BIOS to whatever level you prefer.

There's also already AM5 ASRock Rack with ECC support in the specs:

https://www.asrockrack.com/general/productdetail.asp?Model=B...

There are reasons for folks to wait for the PRO.

* First are the EPYC memory protection features. No other Ryzen SKU has them. I actually care about these protection and isolation features.

* Second are the GE variants. Think similar to Intel’s T and TE variants. These are super binned for extra low power consumption. An 8750GE will be binned to actually be at max performance at a 35W TDP. Yes, you could set a G to Eco Mode or whatever, but I’ve done those experiments with 5700Gs and 5700Xs, and the performance nor power consumption is not quite there compared to the GE versions.

* Last is AMD DASH support. Think Intel vPro IPMI without needing a dedicated BMC (just an onboard NIC that supports DASH). Non-PRO Ryzen SKUs do not support DASH.

I won’t be waiting regardless as my 5750GEs are good for many more years still, but the PRO variants genuinely have real differentiating features.

Nothing to praise AMD about: 1. Same greedy strategy to sell 8 cores with top igpu while put heavily cut down igpu to 4 and 6 cores. Nobody need 8 cores for this little gpu performance. In fact you have to disable half of cores to get better fps and most users of RDNA handhelds already did it. Steam deck have 4 cores for the same reason. 2. 780M is rebranded 680M released 2 years ago with same performance. Before it, AMD had been rebranding vega for 3 years. 3. Intel igpu is already on pair in laptops. 15 gen will have same igpu performance and same cuts as AMD while having cheaper 2 and 4 cores for casual players. 4. AMD drivers still awful and most gamers switch to Nvidia just because of it and never look back. 5. “Insane gaming performance ” is equal to performance of 8 years old 1060 that can be bought for 30 euros. New parts like 6500, A380 or 1650 GDDR6 are available for 100-150 euro. Gaming PC for kids with performance of 300+ euro 8700g can be built for 100 euro (chinese LGA 2011-3 Xeon + RX 570 refab). 6. Non gamers have no reason to buy APU at all. It will stay as niche of SFF and corporate boxes.

I had an AMD APU setup that I used for a while as a light gaming rig. I'm impressed to see these chips continue to grow in capability and be offered at reasonable prices.

I'm curious how the FOSS video driver for these fares on Linux. The newest discrete cards have random full system lockups and driver crashes that make them unusable for anything that stresses the GPU.

I currently have a 3 year old 5800H APU from 2021, which replaced a 2400G (2018) I had before it. I've generally had no issues at all with the integrated graphics.

The AMD APU you can find in the steam deck running valve FOSS 3D drivers had lockups?

Green screen of death indeed. Had them on 4800h laptop regularly. Laptop worked fine on Windows. AMD already dropped support for vega igpu so I guess these will never work in Linux. May be RDNA drivers is better because Valve using them in Steam Deck. Still AMD can break it at any release as they did many times.

Can you point me to resources regarding these lockups? I regularly game on Linux using the FOSS amdgpu driver (kernel 6.5) with a 7900 XTX, which is relatively recent (i.e. RDNA3), and I haven't seen any yet.

I have had a Ryzen 7 Pro 3700U based Thinkpad T495 since 2019, running Fedora.

I can't remember many full system lockups, but there may have been a few. I ran Steam games on it for a year or two before switching those to a different system. But I ran a number of Linux native or Proton-based games and it was the best performing system I had for a while.

What I do recall is that OpenCL barely worked and could get into strange "stuck" states only resolved by rebooting. I could still use the windowing system for a bit, but some kernel calls seemed to begin hanging. It was very disappointing, as I knew this machine had a lot of FLOPS I could have put to use if the OpenCL drivers worked.

One quirk I never resolved in Linux is that the mouse pointer would seem to jump around sometimes while typing. I haven't seen this on any other Thinkpad and it seemed to persist across multiple different Fedora installations. No adjustments to the "mouse" controls menu seemed to make a difference.

I recently started using a Ryzen 7 Pro 7840U based Thinkpad P14s. It doesn't seem to have either of these two issues, but I haven't used it for very long. I find it to be an impressive machine so far.

I have a dual-GPU setup, with a 6900 XT passed through to a Windows VM for games. So I guess I can't say anything about new cards having issues. But the amdgpu drivers are fantastic on the RX 570 that Linux does touch.

Press F to doubt(c)

It says that highest SKU has Radeon 780M (12 CU) inside. That's exactly the same GPU as in 7940HS/7840HS and it scores slightly below 1060-3Gb cards in 3dmark.

The only difference between 7840HG and newer APUs is addition of the (useless?) NPU unit which functions are unclear to me. Train neural networks on a 35W APU?

I've entertained an idea of buying 7840HG SBC as a mini travel PC for 1080P. Still thinking :)

the NPU is expected for inference. in the near future, AMD is expecting small models to proliferate.

of course, this depends on a toolkit that abstracts the various competitors

It'd be interesting to benchmark this against say a 5600 with an actual 1650 (cost of both combined's about 300 bucks at current prices)

My first build was on an APU - if all you wanted was 'a working computer' it was fine, but I soon switched to a normal CPU and GPU

More interesting to benchmark the same CPU with an actual 1650.

True - I was just going for 'around the same price'

I went 3400G -> 1600 Super -> 5600X, and performance with the latter was much better (other than being 2 gens behind, the older APUs had half the cache of the non-G variants)

Depending on the state of iGPU transcoding this'd probably make a nice chip for a Plex box

What kind of LLM inference performance can we expect if we pair it up with 192GB of DDR5?

abysmal, as it ~100 times slower than A100

Still, it's possible to test/validate on your desktop and then move to the big machines.

This was, until a short while back, Apple's exclusive terrain.

Can you expand on this? What is the current co tender for LLM inference that is ~on par with Apples offerings?

It entirely depends on workload. This blog has a diligent mixed-model comparison of the M3 laptops lined up against the Titan RTX (2019) and Tesla V100 (2017): https://www.mrdbourke.com/apple-m3-machine-learning-test/

Would be interesting to see results for the beefier Macs with more RAM. Curious as to what one could do with larger AI models.

I'm aiming at 64GB or more for my next Mac, but not because of AI but to ensure longevity. Macs in this house tend to have very long and productive lives.

https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inferen... should give you consumer comparison

Expect the number of cases where it would give a real advantage would be pretty small. In all the cases presented, Apple gets completely destroyed even with their best chip at around 10x worse; that is before even talking about the cost, after which you can only conclude that Apple hardware makes absolutely no sense for this type of task.

Even if you found enough tasks where the larger memory available would make a difference to average mixed performance closer to a NVidia solution, they would seriously need to lower their RAM price for this argument to hold up...

With their current commercial offering, Apple isn't competitive for anything, even though the computers and chips themselves are not bad at all.

M3 Max 128GB RAM inference on LLaMA2 70B is much faster than on Zephyrus G14 64GB with 4090/16GB because one can fit the whole model into the RAM and share it with GPU. For smaller models that can fit into 16GB G14 smokes M3 Max.

Apple was, IIRC, the first company to make the fact they could have 192GB AI models fully accessible to the GPU a selling point. I remember an AMD chip that could have up to half the memory accessible to the GPU (but not shared with the CPU, IIRC), and read something similar about Intel ones, but no explicit references to sharing the full RAM pool with the GPU at full bus speed.

> the first company to make the fact [...] a selling point.

I don't think that's as impressive of an accomplishment as you think it is...

LLM? Probably ok. Image diffusion? Not so hot, but better than CPU.

You got that switched around. Text generation is harder than image generation.

It is also memory bandwidth bound for non-batched cases.

Very poorly as the GPU is weak and bandwidth is low.

What is the performance for number crunching (rather than polygon pushing) compared to similar APUs in the same price range?

There are no similar APUs in the same price range. Unless you consider regular i5 CPUs as APUs, which they technically are.

> Unless you consider regular i5 CPUs as APUs, which they technically are.

Yes. The integrated GPU is the defining feature.

Reply on news.ycombinator.com