B70 or R9700? (help decide)

Hardware Hub GPU
122

Also also. Since I already have the 1 R9700, I can see myself falling further down this rabbit hole and itching for a second one. Again needing the x8/x8 board.

Not right away. Things to consider on my end is that I am still on AM4, and already have 128GB of DDR4 RAM. I would probably maxing out / pushing the limits of AM4 at that point. But I am not ready to jump to AM5 yet, and the same set of ram I bought over a year ago now costs 4x that. And the jump to 128GB of DDR5 is big oof…

Could be better to wait till the madness settles down and even wait for AM6. Skip a generation. I am patient. IDK, hard to tell.

That said, I wouldn’t be surprised if the 9700’s jump in price soon and maybe it’s not the worst idea…

123

If DDR5 is outside of budget and getting/setting up a x8/x8 is hassle, there’s another solution but it might be a bit of a rabbit hole in itself…

The PEX88024 can do x8/x8 PCIe 4 with a x8 to CPU, and the PEX88048 can do x16/x16 or even x8/x8/x8/x8 (assuming it has the DIP switches or the seller provides the firmware to change between modes). The 48 cards also tend to have MCIO connectors instead of M.2 and MCIO-PCIe adapters are less of a hassle imo.

There’s also a 96 variant that does 4x16 or 8x16 (10x16 should be possible but I haven’t seen it myself) and PCIe 5 switch chips with the same downstream capabilities, with a price to match of course.

124

Couldn’t add a link or edit, but searching for “A Neverending Story: PCIe 3.0/4.0/5.0/6.0/Sci-Fi+” should get you started

1 Like
125

btw i have the taichi x870 so i ll tell you how the r9700 + r9070 perform

1 Like
126

I am also curious what impact AM4 vs AM5 has on PCI-e lane speed. I guess that would be
PCI‑E 4.0 vs PCI‑E 5.0

If this table is correct: PCIe speed table (from gen 1 to gen 6) - NAS Compares

SPEC X1 X2 X4 X8 X16
PCIe 1.x (2.5 GT/s) 500 MB/s 1 GB/s 2GB/s 4GB/s 8GB/s
PCIe 2.x (5.0 GT/s) 1GB/s 2GB/s 4GB/s 8GB/s 16GB/s
PCIe 3.x (8.0 GT/s) 2GB/s 4GB/s 8GB/s 16GB/s 32GB/s
PCIe 4.x (16 GT/s) 4GB/s 8GB/s 16GB/s 32GB/s 64GB/s
PCIe 5.x (32 GT/s) 8GB/s 16GB/s 32GB/s 64GB/s 128GB/s
PCIe 6.x (64 GT/s) 16GB/s 32GB/s 64GB/s 128GB/s 256GB/s

Looks like its a factor of 2. So does that mean that PCI-e 5 on x8/x8 would be roughly equivalent to PCI-e 4 x16/x16?

I guess the more specific question is; how much impact would AM4 vs AM5 have on a 2 GPU x8/x8 split?

127

found it:

I guess I have more reading to do…

128

Well, here’s the update - it’s working flawlessly. The main downside with the B70 has been context loading, but because of the huge vram I have resorted to only running kv cache unquantized. I can fit 128k cache + image recognition with the gemma 4 26b full model very comfortably, ~30gb of vram, though loading a response with the full cache will take a few minutes.

Rimworld - I am running rimtalk, echotales, echocolony, rimmind, rimmind-advisor - all generating near instantly. I changed the model to Gemma 4 12b (not a MOE model) to load the context a bit faster, and am able to parallelize 4x with 16k context each. If you’ve ever just maxed out your tokens and needed something else - this seems to be the solution, and Gemma 4 has been by far the best LLM I’ve been able to run locally.

1 Like
129

Wtf is rimworld modding community doing. I need to install it again.

130

It seems like a natural evolution; the pawns complain directly to you. I’ve really cranked up the settings and am generating ~200k context/hr and 20k token output. Keep in mind it is based on in-game events - more pawns would increase usage. I will say my PC is belching out heat - even without a gpu intensive game.

131

Replying here, on a tangent, for the models you suggested: I am also running on a few 16GB cards, mainly A770 from intel and came accross this thread as I am contemplating B70 vs R9700. My use case is a local builder under a model that orchestrates. I am developing a skill for my smart tools to call a selection of local models on my A770, which willl generalise for other cards.
To the point, on the A700, the qwen3.5-9b is the fastest performing at max context, but fails some of the coding tests I have packaged to profile the models I use and I would like to know if you found the models you suggested as a marked improvement on coding tasks (well defined by a larger model, builder role ).

132

True, but other architecture stacks are improving and people are putting in the work to optimise for teh price benefit they get.
Your use case and your time make or break the value proposition.
I agree with you nvidia is very well positioned, but there is space for other options. I am watching general ecosystem and development push.
I have supported (bought) Intel’s forray into the GPU market. But AMD is really pushing on many fronts and on that metric alone, the R9700 looks to benefit the most, I believe.

133

For what it’s worth…I started off in the same place as you - a pair of A770s, on the grounds that they provided reasonable VRAM bandwidth and were the best-value proposition in terms of cash per GB.

In fact, I’ve done that experiment twice - and they just plain underperform to the point where the performance (in my opinion) didn’t even constitute a useful data point. The fact that Alchemist is basically abandonware as far as Intel are concerned doesn’t help.

Replacing them with an R9700 (and later a second one) put me in an entirely different world of capability.

To give you an idea, the pair of A770s struggled to run anything useful above 40t/s even with the smallest models, whereas the R9700s run (in vLLM) Qwen 3.8 27B FP8 at a consistent 95-100t/s in code gen with spikes up to 140t/s, and over 3000t/s prefill.

Basically, it’s not just a step-change in terms of performance, but also capability and quality.

There’s also the new ggml-hrx backend for llama.cpp that was announced by AMD recently as a focused alternative to the ROCm + HIP approach. It’s nowhere near primetime yet (there’s a lot of work for them to do), but they’re dedicating resources to it in the same way that Nvidia have with the CUDA backend. The performance uplift seems significant, to the point where it may even compete with vLLM for dense models.

1 Like
134

That was my experience as well. I kept hoping, until it was clear battlemage is the way forward for intel. I even made a repo around the A770, a small harness, served as profiled builder roles to an agent. But performance is always problematic.

Moved to B580, then B60, with b70s on the way next week. vLLM on battlemage is good for the price of the cards. I am looking at the 9700 AI pro, but I am not biting a dual stack bullet yet…

Still had to do something with the A770s… They are running in an OpenClaw VM, PCIe passthrough on a proxmox box… One is the backend (qwen3.5 9b) one is used with the skill repo I built, to continue development.

135

I have a couple of these in a 7800 3d ryzen.I find them fairly energy efficient for the power and they run great on Ubuntu and they are well supported still and AMD only seems to be getting going here. I generally use them in preference to my DGX Spark. If you are having fun , just get the intel but be prepared for some software pain. TBH I get a lot of pain with the NVidia as well.

136

So my harness distributes sub-tasks to agents within 64k context, with many elements to minimize context use for fluff. The models don’t have a benchmark beyond that of “do they respond through 64K, does the response error?” If they fail at the harness subtask I don’t use them.

Long context tasks can always be divided, LLM’s are great at distributing and converging those segments. These models do not get long form multi hour tasks. The harness assembles the plan, task and sub-tasks with one-shot calls to various models, and the “swarm”(it’s sequenced on 16GB VRAM) handles the grunt work.

I developed a spec based workflow in march that I’ve used successfully with this workflow, but we built the harness greenfield, and I’ve not even kept the starting models we began with, as newer models became more capable.

Current trio:
fv78pd/gemma-4-E4B-Agentic-Opus-Reasoning-GeminiCLI-GGUF Q 4

empero-ai/Qwen3.8-9B-Distill-GGUF Q_5

yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF Q 4

137

Been a few months. Small status update.

Already has jumped by +$600 at my local hardware shop. Glad I pulled the trigger.

Re: AI

I still don’t use any IDE integration or agents. I copy paste code back and forth into Open WebUI to solve specific problems. Maybe I’m still too hesitant old man brain that doesn’t want too much deep integration yet. And I certainly won’t have agents touching my system settings or having the ability to install packages.

But man these local LLMS are objectively very useful and getting pretty decent. Dropping in classes or functions and asking it questions. Giving it stack traces. Pretty neat tool.

Often I don’t like the format of answers. I’ll reformat the code into my style and standards, remove unnecessary comments. And I hate the useless unnecessary emojis that a lot of models like to inject into answers (and sometimes even code) for absolutely no reason. Thankfully this is easily solved via a global system prompt Don't use emoji's in explanations or code.

I like to program without AI sometimes and only fall back to it for certain questions or problems. I have been programming for a long time before LLMs and I do not wish to deskill, so I try to keep sharp and exercise without overreliance on it. Even if the AI can be faster than me for certain things, I still want to use my brain.

It is also useful for many non-programming tasks too, such as general system maintenance. For example just recently I had an ssh session disconnect while doing an apt-get upgrade on another system, and it must have stopped at a prompt or something because when I reconnected I couldn’t update as the old process was lingering in limbo or something:

Waiting for cache lock: Could not get lock /var/lib/dpkg/lock-frontend. It is held by process 26785 (apt)

Told my GPU about my updating woes, provided the error and it told me the solution; first to kill that process then --configure:

sudo kill -9 26785
sudo dpkg --configure -a

And afterwords I can update again! All without having to touch the internet.

I have many other such trouble shooting examples solved locally. Life is good.

Re: Games

Don’t think I’ll bother with the hassle of water cooling. Again the noise doesn’t really bother me too much as I mostly game with headphones. It’s “good enough” IMO, but certainly far from silent.

Pretty much every game running flawlessly with the R9700.

Recently picked up Body Cam, and its a lot of fun. Bit of a twist on the classic FPS perspective.

The only game that has given issues so far is Battlefield 4 via Lutris. It’s an older title, but still holds up and modern BF titles wont run on Linux due to their garbage AC solutions…

Unfortunately with the R9700 there is some sort of rendering artifacting where parts of the map or certain layers are flickering in and out of existence. Tried tweaking all sorts of in-game graphics settings to no success. The flickering is bad enough to the point where it is unplayable. Wasn’t an issue on the 6750.

Nothing in the logs. Unsure how to diagnose that further. Perhaps I’ll have to have a session where I sit down and interrogate my GPU why it is unable to render that game…

Overall very happy with the upgrade.

The 6750 has sat on my desk unused since swapping. Might sell, might hold onto just in case, idk.

I’ve heard rumors of a 64GB RXXXX variation, which might be tempting if the prices aren’t too insane at that point. Probably better to have it all on the one GPU than deal with multi-GPU hoops and hassle.

Also very curios about something like the Huawei Atlas 300I Duo. Probably much worse for gaming but 96GB for AI would probably be pretty sweet.

138

Linking this thread because there is more good info on a similar platform (AM4) and similar R9700 consideration.