AMD’s Phoenix SoC

https://chipsandcheese.com/2023/09/16/hot-chips-2023-amds-phoenix-soc/

206 155 3 years ago

pbj1968• 3 years ago

Can it run Gauntlet Legends 100% in MAME?

xnx• 3 years ago

Hot Chips related, I found this the keynote from Google on the state of the art (that they're willing to talk about) on ML hardware fascinating: https://www.youtube.com/watch?v=EFe7-WZMMhc

I had no idea about so many of the topics discussed like power management and detecting and mitigating calculation errors.

jauntywundrkind• 3 years ago

Documentation suggests a 1 GHz clock, but Phoenix’s XDNA might be running at 1.25 GHz as AMD says BF16 is supported with 5 TFLOPS of throughput.

It's amazing seeing the ML FPGA come out on a CPU. There's such an incredible opportunity for AMD here, if they can get the world adopting & using this kind of hardware that most folks in dev are not used to using. Great seeing some internal details here.

I haven't heard a peep thought about devs actually having access to this hardware. One of the big criticisms at launch time was that there was essentially no material for using this sizable part of the chip.

It's not just important for Phoenix & latter chips, it's potential bridge for AMD to make their FPGA in general used & adopted. But like the GPUs, it's questionable whether AMD can get broadscale enough adoption for advantage to mean anything for them. Ideally AMD would be working with yosys or openxla or someone to have an easy to adopt synthesis pipeline folks could play with. Right now I still haven't heard that they have anything, even proprietary or self-made.

Edit: oh well, AMD just dropped something! https://www.tomshardware.com/news/amd-posts-ryzen-ai-softwar...

dhruvdh• 3 years ago

Direct link to developer docs - https://ryzenai.docs.amd.com/en/latest/index.html

cherryteastain• 3 years ago

Such a shame this is not a part of ROCm. It'd give developers a big incentive to actually use it, unlike the dumpster fire state it is in now.

arein3• 3 years ago

Windows only software. I am/was a big fan of AMD for finally giving the market cheap powerful laptops. My laptop is amd, my PC and a few other PCs I build were all AMD, but I might have to go with Intel on the next GPU because they have open source GPU drivers.

hamandcheese• 3 years ago

I thought AMD has open source drivers for their GPUs?

wmf3• 3 years ago

XDNA is not an FPGA just because it comes from Xilinx.

tyingq2• 3 years ago

Searching around some suggests it is not just an FPGA, but includes one.

mathisfun123• 3 years ago

It's amazing seeing the ML FPGA come out on a CPU.

I don't know what "on a CPU" means here but the AIEngine tiles aren't FPGA-like or reconfigurable logic - I don't know where you got that idea. As the diagrams show they are VLIW with vector instructions. They (the fabric, not AIE tiles) do have fully programmable DMA engines but I don't know how much is exposed to third-party devs.

it's potential bridge for AMD to make their FPGA in general used & adopted.

This isn't happening and won't ever happen because absolutely no one wants to "program" reconfigurable logic.

yosys or openxla

These two things are so far apart that I have no clue what you're suggesting here regarding synthesis.

brucethemoose2• 3 years ago

AMD is seemingly pitching this as a "power efficient" AI engine, where the integrated GPU is still better for maximum throughput.

The iGPU has a much better chance of broad adoption for GenAI type of stuff because its basically mapped to CUDA through ROCm, and probably works with triton.

imiric• 3 years ago

Absolutely love this chipset. It powers my favorite tech of 2023.

The 7840U is a beast in handhelds, and the 7940HS likewise in laptops and mini PCs. They're great for gaming, media, productivity, and anything you throw at it. Intel cannot compete with the performance and TDP of these. AMD is completely dominating the portable x86 segment, so much so, that I'm not envious of Apple's dominance of ARM. Can't wait to see the next generation of this chipset, as well as whatever Intel can produce in response.

jwells892• 3 years ago

Just wish I could find a 7840U in a flagship ~13” ultrabook with a big battery. That’s surprisingly hard to find, for some reason AMD tends to get relegated to second-tier models and paired with middling batteries.

bronson• 3 years ago

Not far off: https://frame.work/products/laptop-diy-13-gen-amd?q=processo... Impressive, especially for the price.

bas3l1ne• 3 years ago

The Thinkpad z13 (gen 2) hasn't released yet, but is probably the closest thing to what you're looking for

suprjami• 3 years ago

Current ThinkPad T14 has 7840U. This is the same width/length as my "13 inch" Skylake ultrabook (Asus UX305UA) and only 3mm thicker.

ThinkPad T14s Gen 3 has a 6850U and is ~1mm thinner than T14.

You're right though, none of them are superthin (<15mm) like ThinkPad X1 Carbon or Dell XPS 13. Maybe it's not possible to cool these faster AMD CPUs in such a small volume? Nobody wants to type on a space heater.

Edit: found these pages with complete lists, hell yeah

https://www.ultrabookreview.com/36030-amd-ryzen-7-u-laptops/

https://www.ultrabookreview.com/35985-amd-ryzen-9-laptops/

grumpyprole• 3 years ago

I have a work provided X1 Carbon with an Intel that often overheats and throttles. Would an AMD solution be better?

nik736• 3 years ago

In Germany we have the Lenovo Yoga with a 7840U.

tempest_• 3 years ago

You may or may not know that the term ultrabook is an intel trademark, so you are not going to find an ultrabook branded machine with AMD in it.

I'm sure intel is pumped that the term is being genericized.

ksec• 3 years ago

Its funny the word "Chipset" being used here. ( As it used to mean something different ). But yes I Love these AMD SoC. And it is a truly general purpose PC SoC with wide availability. Sometimes I just wish to stick one in a Smartphone ( Or should I say a Pocket Computer ) just for the sake of it. Especially when Zen 5 is suppose to be much closer to A16/A17 in terms of IPC.

I do wish AMD GPU department could work better though. Not just on ROCm but the overall GPU market shares. They have the advantage of being in the console but on PC they are still a minority.

sneak• 3 years ago

Is it just me, or does the inclusion of a Microsoft Pluton HSM (and of course the AMD PSP) inside the SoC make this a no-go for security critical stuff?

I'm sure systems vendors will love putting these in Windows laptops (a growing market with a bright future, I'm sure) but I can't imagine hackers having much use for these things.

wmf3• 3 years ago

It's just you. It's unlikely that Pluton is actually evil but MS keeps creating bad PR by refusing to document it.

rhaway84773• 3 years ago

If not creating documentation is leading to bad PR why doesn’t MS just create the documentation instead?

Occam’s razor would suggest it’s because accurate documentation would create worse PR.

That’s not reassuring at all.

pjmlp8• 3 years ago

It is documented for OEMs.

ilyt4• 3 years ago

Historically anything MS wanted to have in hardware was to make it more PITA to install windows alternatives.

It's extremely likely it is.

monocasa2• 3 years ago

Not any more than the existing PSP I'd imagine.

jakobson14• 3 years ago

https://gabrielsieben.tech/2022/07/25/the-power-of-microsoft...

the documented applications of pluton are far more horrifying than the PSP

hhh• 3 years ago

Why would it?

roenxi• 3 years ago

Article says:

Another cool use of low power, always-on DSPs is using ultrasound sonar to detect humans. AMD’s ultrasound runs at above 20 KHz but below 35 KHz, letting it get through the microphone and speaker’s band pass filters. Then, it can use Doppler shift to distinguish human movement from static objects in the same way that a look-down radar filters out ground clutter.

This chip sounds pretty creepy, although I doubt it'll be possible to escape from this sort of thing. The US spies will be loving these feature trends.

jimbob45• 3 years ago

Microphone on the chip or via an attached peripheral?

hypercube332• 3 years ago

I was just reading about the use of this and the company it looks like these are from (mobile so I don't have the name handy but it looks like Lenovo T14 g4 uses it) one of the "features" is to auto lock the machine and another is for auto monitor finding so it can lay out the desktop magically based on where it thinks monitors are to the laptop. Kinda neat. I'm not sure how it'll work for most offices based on the demos though. It looked like it could just decide if a monitor was right or left but most monitors are above AND right/left on mounts.

kibibu• 3 years ago

Now your Windows laptop can frustrate both humans and canines

silvestrov• 3 years ago

This will be torture for dogs. I is like a 5 kHz tone for humans.

All laptops with a 5 kHz tone would be returned as defective. In the 90's harddrives often had a 7200 Hz whine which was enourmously annoying. Todays harddrives are silent compared to the old ones.

acumenical• 3 years ago

Reminds me of BadBIOS, a little bit

pjmlp8• 3 years ago

Pluton is required for future Windows versions, exactly because of security in markets where full integrity hardware and software is required.

bboygravity• 3 years ago

Because security markets would use (future) Windows?!??!?

tedunangst• 3 years ago

Note that you can still buy a new 7000 series mobile ryzen with Vega GCN graphics if that's your thing.

icegreentea2• 3 years ago

Yes, the 7030 family will net you Zen3 + Vega cores. The 7030s are ultimately rebadges of Cezanne - so the old mobile 5000 series parts.

beebeepka4• 3 years ago

You sure about that? I think my 6800h has RDNA2. There are more rebrands in the 7000 family, though, and some do indeed have Vega chips. The person you replied to us correct but you are not.

unethical_ban2• 3 years ago

It's a massive disservice to customers to have 7030 series and 7040 series be different generations of chip.

speed_spread• 3 years ago

Weird quote of the week, from the article:

"As another note, dog hearing can cover the 20-35 KHz range, so the ultrasound engine may be able to detect static dogs by making them non-static, after which they will cause a Doppler shift."

So what's a static dog? Rust me wants a &'static dog, Java me wants a static final dog. And I also want a real dog named Doppler, or maybe just Shift?

nrp• 3 years ago

It’s a goofy joke. The human detection uses ultrasonic audio outside of human auditory range. However, dogs can hear those frequencies, so the computer chirping is going to make a resting (static) dog jump or run away, making it non-static.

hinkley3• 3 years ago

This is going to backfire.

treprinum4• 3 years ago

7940HS in my 4090 laptop is cool, though I am worried about the future starting with Phoenix 2 that will feature few good cores (Zen 5) and many crappy cores (Zen 5c). Not sure why AMD had to ape a failing Intel arch here that was introduced only to improve Intel's multicore benchmarks. Now with AMD we will get all the "heterogenous goodies" when programs keep running into weird states due to scheduler not being 100% in its core allocations. Even if 5c's only difference is in the smaller cache and slower AVX-512, this could bite in unpredictable ways. Moreover, if AMD really needed to go with this hybrid design, why not ape Apple M1 or M2 with only 2 or 4 crappy cores?

DeepYogurt• 3 years ago

"heterogenous goodies"

Like battery life

Klinky2• 3 years ago

You can do more power optimizations with efficiency cores than the performance cores. There are power tradeoffs to get max perf. You can significantly reduce vcore if your max frequency is much lower. Cache is also power hungry & takes up space.

treprinum4• 3 years ago

5c should be the same as 5 but with a smaller cache (unlike Core vs Atom distinction in Intel CPUs). I'd rather have a CPU full of 5c than a mix of 5 and 5c.

wtallis6• 3 years ago

Early indications are that Zen 4c actually requires significantly higher voltage for the same clock speed as plain Zen 4 cores (though still a lower voltage than Zen 4 at speeds far beyond the reach of Zen 4c). All the savings AMD gets by targeting ~3.5GHz peak rather than 5+ GHz has been put into shrinking the core area: https://zhuanlan.zhihu.com/p/653961282

wtallis6• 3 years ago

Phoenix 2 that will feature few good cores (Zen 5) and many crappy cores (Zen 5c).

Phoenix 2 is already starting to ship, with Zen 4 and Zen 4c cores in a 2+4 arrangement. I don't think there's anything wrong with this approach: the main downside of Zen 4c cores is they cannot clock as high, but the kind of systems the SoC is going into are thermally limited to the point that using more than two CPU cores is already guaranteed to drop you down to the clock speed range that Zen 4c runs at. Ignoring the impact on chip floorplan and just looking at the implications for software, it's more like a more pronounced version of Intel's Turbo Boost Max 3.0 (preferred cores) rather than the heterogeneous P vs E core situation. All the cores have the same feature support and same performance per clock and the same sizes for their private caches.

treprinum4• 3 years ago

The smaller cache will make it a heterogeneous situation for some workloads and likely games as well (seeing how 5800x3D/7800x3D are demolishing the competition thanks to a larger cache).

Const-me3• 3 years ago

AMD has already released a product featuring Zen4c cores designed for servers. In contrast to Intel, AMD's small cores are not crappy; they are real Zen4 cores, with a reduction in size achieved by sacrificing some of the L3 cache and clock frequencies.

As for the future of Zen 5, we are uncertain about its development, but I am hopeful that AMD will follow a similar approach to Zen4c.

By the way, Intel employs different cores derived from their Atom line of low-power processors, specifically the Gracemont microarchitecture in their current products. Unfortunately, these cores are indeed crappy, particularly in terms of floating-point performance.

treprinum4• 3 years ago

I am fine with a CPU full of 4c/5c cores, I however don't want a mix of 4/4c or 5/5c.

Havoc2• 3 years ago

I wonder if this with lots of ram will make it similarly useful for large parameter llm like macs

fgblanch• 3 years ago

No, the reason macs are better on LLMs is memory bandwidth 800Gb/s on Ultra 2 . I couldn’t find a good source but it seems that Ally mem bandwidth is around 70GB/s

wtallis6• 3 years ago

A combination of high memory bandwidth and large memory capacity is necessary for good performance on LLMs. Plenty of consumer GPUs have great memory bandwidth but not enough capacity for the good LLMs. AMD's Phoenix has a memory bus too narrow to enable GPU-like bandwidth, and when paired with the faster memory it supports (LPDDR5 rather than DDR5) it won't offer much more memory capacity than consumer GPUs.

marci2• 3 years ago

Its theoretical 120GB/s (with soldered lpddr5, the fastest) is between a M2 (100GB/s) and a M2 Pro (200GB/s).

For a portable solution it can be a nice compromise (only considering LLMs here). The only way to get 64GB of Ram (or more) with apple is with a M1/M2 Max. It's way faster (400 GB/s) but the retail price is at least ~4000€. The M2 Pro doesn't get above 32 GB of RAM.

For a while I considered a macbook pro, but the price tag, mixed with the fact that the SSD is soldered (just with LLMs, I plan a lot of mileage on them, and the failure mode of these macs is to just never boot up ever again, even with an external drive), and then I heard about the GPD Win Max 2 with an AMD 7840U. It's a lesser known brand, and I'll have to wait for early october to get mine with 64GB LPDDR5, but seeing some people receiving small parts after breakage/malfunction for previous gen also tipped the scale. For LLMs, it should be about a third the speed than a portable M2 Max, but it's a quarter the price, and I like some stuff that it can do that no other laptop can, so I'm fine with the tradeoffs.

Havoc2• 3 years ago

Unconventionally choice of device!

I’ve had decent luck with zen minipcs- though the earlier gen 4700u. Take 64gb but too slow for LLMs realistically. Plus only 8gb max assignable to gpu.

MassiveBonk51• 3 years ago

I want these to come to desktop at some point. I'd really love to have a compact x86 box powered by one.

bj-rn• 3 years ago

You could buy a framework mainboard and build one: https://frame.work/products/mainboard-kit-amd-ryzen-7040-ser...

ajmurmann• 3 years ago

Aren't these in mini desktops like the Minisforum (https://a.co/d/9BugA0r) already? I just bought my wife a older model for $250 to play older games together. It's ~1/4 the size of a mac Mini and the stuff we care about (CS:Source, Minecraft) runs like a rocket ship.

wsc981• 3 years ago

Is this PC quiet? Actually been thinking about a Minisforum Venus last week ...

bearjaws3• 3 years ago

Having recently bought a ROG ally, I have to say I am incredibly impressed with the SOC. It is pretty quiet and can run most of my games very well.

I am blown away to play OW2, Apex, Cities Skylines and even BOTW on a hand held device, incredible device for people who travel a lot.

ekianjo3• 3 years ago

The steam deck could do that a year before the Ally and with much longer battery life

rowanG0772• 3 years ago

And with a lot worse screen. The screen is the main thing keeping me away.

coder5432• 3 years ago

I am blown away to play OW2, Apex, Cities Skylines and even BOTW on a hand held device

All four of those games are available on Switch, which runs even more quietly and requires a lot less fiddling and maintenance.

I have both an ROG Ally and a Switch, I just find that selection of example games particularly amusing. Starfield works well enough on my Ally, and doesn’t exist for Switch.

hinkley3• 3 years ago

City skylines is on Switch? And it doesn’t melt?

izacus• 3 years ago

They all also run horribly on the Switch like most ports.

I was blown away just how much better experience those games on Steam Deck are in comparison to the Switch.

bearjaws3• 3 years ago

I have to buy all my games again to play them at lower quality and frame rate? Seems like a bad deal, I also dock my ROG Ally, and use it on my TV running 1080p60, something the Switch seldomly does.

As for fiddling, I really don't understand this commentary, the only 'fiddling' is toggling the controller mode and power target? I don't understand how basic settings that are built in to a easily accessible menu is 'fiddling'.

OW2, Apex, and all emulators I've tried have Just Worked (TM).

flakiness2• 3 years ago

Any recommended Linux laptop with this? Does Framework 16 have it? What else? I was thinking about Dell XPS but AMD's 16 threads CPU is so appealing.

gmokki• 3 years ago

We just put in orders for Lenovo P14s gen 4 amd laptops for developers. With 64G ram, 2.8k OLED display.

We have been very happy with P14s gen 2 amd with 4k display, but it has only 32G of ram. The 15W 8 core Ryzen apu has been silent and performant.

CoolCold2• 3 years ago

Thinking about this one for my sysadmins needs, would like to hear your experience with that machines, thanks in advance!

flakiness2• 3 years ago

Note taken, thanks! I hope they have larger screen variations though.

Edit: Noticed there is also P16 - This looks great!

iopq• 3 years ago

XDNA benchmarks when

pixelpoet• 3 years ago

No OpenCL drivers, no interest from me.

d3w4s95• 3 years ago

Have been using a 7735HS (6800H) based mini PC for the last few days, very impressed with it -- the entire machine idles at 10W, light load < 20W, medium load at < 40W and heavy load ~ 60W. The fan isn't on most of the time. The latest generation is likely better, and U series is even quieter. It has better performance and produces less heat than my Intel NUC and laptop.

andrea762• 3 years ago

Please tell us what mini pc is

d3w4s95• 3 years ago

Minisforum UM773 Lite

(I hesitated a long time before the purchase, because I wanted the combination of an Intel NUC warranty and reliability, AMD CPU, good design & cooling plus a good price -- which of course does not exist. However I haven't had an issue so far, and I hope it lasts long)

tyrfing2• 3 years ago

10 watts idle, really? That seems extremely high. Do you have C-states disabled?

getcrunk• 3 years ago

He quoted entire machine @ 10w not just the cpu.

Const-me3• 3 years ago

If you consider a laptop with Ryzen 9 7940HS, and you gonna pay with your own money, note that Ryzen 7 7840HS is pretty much the same chip. The only difference between the two seems to be a couple hundred MHz of various clock frequencies, which is barely relevant for laptops due to thermal throttling on sustained workloads.

The rest of the specs are the same, but Ryzen 7 is somewhat cheaper due to marketing. For example, at the time of writing the configurator for HP ZBook Firefly 14 G10 A laptops offered in US says the price difference is $285.

ilyt4• 3 years ago

Spikes of work is how most workflows work, max frequency is far from irrelevant.

Sure, if your main wait is waiting for compile it won't help that much but everything else it is helpful

Const-me3• 3 years ago

I don’t think in this case it matters even for spiky workloads. The relative difference is too small.

The difference in CPU base frequency is 3.8 versus 4.0 GHz, which is 5% slower. However, for spiky workloads boost frequency is IMO more relevant than base frequency. With boost, the clock speed difference diminishes to the laughable 2%. Specifically, they are 5.2GHz for the R9, and 5.1GHz for the R7.

AMD doesn’t publish boost clock for GPU, but relative difference in base GPU frequency is 3.6%.

I’m aware marketing generally works, making people willing to pay more just for the feeling they have the best hardware money can buy. But personally, I believe saving these $285 and getting 95-98% of the performance is a good deal.

P.S. Ryzen 5 from the same line up, the 7640HS, is substantially slower. It has fewer CPU cores, and fewer GPU compute units.

lakomen• 3 years ago

Lately, since I'm not in my home country right now, and like playing games, I've been using my home computer with a 4070 rtx to play and stream h264 encoded game play, using GeForce experience and moonlight. But since this uses quite some power and also I dualboot with Linux being the default (I turn on the machine with IPMI, Asus workstation board), then ssh into it, become root and efibootmgr temporarily to boot Windows on next boot and reboot. In Windows then both TeamViewer is installed, because doing anything in GeForce Experience locks up the Moonlight "server", and the Moonlight streaming stuff, aka firewall opener, so I don't have to wireguard every time. Although my desktop/workstation is 4k resolution, I stream at 1920×1080 resolution.

I can even stream to Twitch at the same time, but it's not ideal. I imagine, I need either a dedicated encoding card or a stronger Nvidia card than the 4070. I get some stuttering every now and then. 6Mbps goes to Twitch, 30 max to me and my 1920×1080 laptop.

Since this is all very complicated, I was thinking, have a dedicated gaming machine which doesn't consume that much power, which I can put to sleep or wake via Moonlight. This looks like a candidate, but it's AMD and idk if that will work. Ok, I looked it up and phoenix should work if it has VCE (which I assume it does, the article doesnt talk about it).

Ideally I could put it in some data center for a minimum fee, but that's wishful thinking, because that fee would not be minimum. With stadia gone and shadowPC having outdated hardware for expensive prices, it could be something. TBH I wouldn't mind renting a gaming computer with streaming, but let's be honest, all the offers assume you're not a pirate. And that's why they're all failing.

So yeah this little box (or soc rather) could be my next hardware purchase, a dedicated game streaming machine. :)

EDIT: I retract that statement, now that I've searched for 780M performance. It would require a dedicated GPU. Oh well. It's a little weaker than the Nvidia 980M of my laptop, but makes less noise. For energy efficiency it's ok, but if you like playing the latest and greatest, not an option. It's essential 30FPS 1920×1080 gaming.

sylware• 3 years ago

Hopefully it will be easy to find good and not that expensive laptops with such chips.

Now I wish for a 100% hardware raytracing BHV... if raytracing stays around. I hope RDNA4 will do that. Because those horrible mesa RADV glsl shaders to do "soft" BHV, erk... (I prefer to disable RT and compile that out, and remove the glslang SDK dependency...). They should have been coded directly in SPIR-V (with a plain and simple C coded SPIR-V translater, or better, direct AMD GPU assembly code (with a plain and simple C coded AMD GPU assembler).

fancyfredbot• 3 years ago

Does any software at all use XDNA?

AMD’s Phoenix SoC

Related Stories