02:19calico: yup with the kernel 6.12 (so at least 6.7) I'm finally getting Steam games working properly
02:21calico: just sad my 2017 laptop that have a GTX 1050 will never support reclocking in nouveau
02:26calico: and since it's not using the radeon gpu games like Death Stranding will not randomly crash
02:27airlied[d]: okay texture size works now
03:30orowith2os[d]: calico: it could, if you feel like tinkering with nvidia's binaries :v
03:31orowith2os[d]: Probably easier to rely on the Nvidia proprietary driver though. On the opposite side there, you lose dmabufs across vendors.
03:32calico: orowith2os: If I had some tutorial from the nouveau devs, and I knew what to search in those binaries I could probably do it yes.
03:39calico: and we won't be able to rely on proprietary drivers forever, since they might break completly with future kernel versions
03:55orowith2os[d]: calico: both true. Messing with the binaries is also illegal if you redistribute them.
03:56orowith2os[d]: 900 and 10 series are just SOL, unless someone finds a viable workaround, like a bug somewhere that lets nouveau get unrestricted control of the GPU.
04:23redsheep[d]: Yeah we talked about this just the other day, all that bypassing firmware signing wins you is the wonderful task of writing gpu firmware from scratch. I still think a tool to extract the firmwares from their .run with support for their firmware added to nouveau would be the best way if you were really determined to get 8+ year old gpus working well
04:24redsheep[d]: I would have to get really really bored before trying that myself, given I am not overly interested in old hardware
04:28redsheep[d]: That approach seems especially preferable if some time in the distant future whatever is blocking redistribution of their firmware ends then that support would be ready to go. I have no idea what that actually is though, and likely nobody in the public like myself will ever know. If it's just plain copyright that could be a very long time indeed.
04:43magic_rb[d]: If its copyright it could be forever, copyright never expires
04:44redsheep[d]: At least in the US it's supposed to
04:45magic_rb[d]: Realistically it wont since it can be passed down and the laws seemed to always been extended when a major work was just about to expire
04:47redsheep[d]: Yes disney is trying very hard to keep it alive but steamboat willie's copyright did expire, it does happen eventually. We might just be dead before redistributing pascal firmware is legal, but if that's the issue that should eventually lift
04:48redsheep[d]: Should have official support in kernel 25.7 or something lol
05:29orowith2os[d]: I doubt the hardware would even be alive today *to* use the firmware
05:29orowith2os[d]: Pascal would be the least of anyone's worries
05:44redsheep[d]: If you replaced every single capacitor I bet it would still work
06:13x512[m]: orowith2os[d]: Why dmabuf across vendors is important? Why display server can't manually copy data from one GPU buffer to another GPU buffer using standard Vulkan API?
06:26airlied[d]: Manually copying data is slow
06:27airlied[d]: Esp using the CPU
07:06airlied[d]: soo many texture ops, whack-a-mole fun
07:20airlied[d]: mhenning[d]: any idea what the difference between tex_urc and tex_b_urc is in the encoding?
07:20airlied[d]: is the bindless, it seems a bit wierd if everything is bindless
07:22airlied[d]: oh just got tld4 to start working
09:14trevortrident: My dad has supported me throughout years, when i have crashed and felt worthless, he is always my hero. So as he get's older but shows still super condition, things can be turned another way around (me supporting him), though he does not ask for support of mine, but i am already staring to do well. Computers have so smart design and intel driver on 2011 macbook has never failed here , i
09:14trevortrident: suspect entirely miscoordinated people are not found here hence, even the blue people and ben skeggs have understood , that yes compression and compressed execution is possible probably, thanks to super design of hardware in general. It derives from a fact of such 288.00+328+368−328−328−328=0 328−40+328+368−328−1024=-40 so the function over this can be anything for index/value 144 aka
09:14trevortrident: 112, including twice the delta feed back and subtracted etc. can also use the spectrum of 80 or 40 etc. it's all a longer story, the condition itself is 256(can be different constant)-144=112 256-112-112=32 and 112-32=80, that is how the bound is gotten for index112 aka 72 or twice of that 144. That is pseudo, but it works on any format that respect logics of invariant in arbitrary sum
09:14trevortrident: power representing numbers. Ladies and Gentlemen the theory is correct it is fair to say likely, so redsheep[d] i am afraid what i said wasn't the funniest thing to say.
10:49snowycoder[d]: Oh no redsheep[d] the bot noticed you, run😂
12:25mangodev[d]: what did i miss 😟
13:38gfxstrand[d]: snowycoder[d]: At least they can't find their way into your Discord DMs. They were in my IRC DMs for years.
14:05redsheep[d]: Discord also makes it much easier to protect yourself from unwanted DMs
14:08mohamexiety[d]: the fact that someone would be dedicated to consistent spam for like 15 years (from what I heard) and even going as far as DMs is scary tbh
14:14redsheep[d]: Yeah there's a few things that can cause extreme paranoia. The degree of obsession, and the obsession with nouveau in particular, is more odd to me
14:14redsheep[d]: Like they've never once been on topic that I've seen
14:16gfxstrand[d]: It's not nouveau. It's all of Linux graphics. He moves from channel to channel.
14:16tiredchiku[d]: I started compiling a list of their hosts to find a patters a few days ago
14:16gfxstrand[d]: Sometimes he's rambling about things that are graphics-adjacent and even vaguely comprehensible.
14:19orowith2os[d]: So did he just show up like this, or was it a downwards spiral, or what :ohcrap:
14:21redsheep[d]: I feel bad for their father if they're ever actually going to need to depend on them, of any of that was real
14:21gfxstrand[d]: He's gotten worse over the years
14:37rinlovesyou[d]: the messages i've seen over my time here from that person seem like genuine mental illness like
14:37rinlovesyou[d]: that person is not okay to be doing this for that long
16:04calico: maybe his father is Linus Torvalds?
16:05calico: kek
16:10calico: or maybe that's somekind of secret code like in the 7 days of the condor
16:15snowycoder[d]: It seems like a markov chain tbh
16:20mhenning[d]: airlied[d]: The b is bindless. That's what I was trying to ask you when I asked whether you were using bindless or not. The bindless forms exist before blackwell, and are largely the same other than the addition of the new ureg source. That's why I was saying I think the bindless forms will be similar to before.
16:21mhenning[d]: My guess is that the bindless forms will not require the handle from ldc.texunpack but the bound forms will require it.
16:24mhenning[d]: I don't know why you say "everything is bindless" - I don't think nvidia considers that to be true
17:46karolherbst[d]: tiredchiku[d]: I'm sure the pattern is "every VPN and every VPS"
17:46karolherbst[d]: rinlovesyou[d]: correct
17:47karolherbst[d]: it's not a bot, it's not AI, or markov chains or anything like that. It's just an unfortunate person, with an unfortunate past
17:49karolherbst[d]: and I also don't like that people keep repeating the term "bot" for him. How would you feel being called a bot if you are suffering from mental problems/brain damage/whatever enough that you end up like him?
18:12calico: tbh he remembers me of a guy with a very similar behaviour on Discord and GitHub
18:13calico: jredfox
18:14calico: but if you're tired of unremovable spam on IRC, why not moving to something like XMPP or Matrix?
18:15snowycoder[d]: karolherbst[d]: I'm really sorry, I genuinly thought it was a bot because I cannot understand what they are saying and they always seem to use different usernames.
18:15snowycoder[d]: I don't know the full story so I just assumed they were.
18:15snowycoder[d]: Thanks for letting me know, and sorry again for any insult, that wasn't my intention
18:16gfxstrand[d]: mhenning[d]: That's my guess as well
18:16calico: orowith2os[d]: 'Messing with the binaries is also illegal if you redistribute them.' <-- who says I'd have to redistribute the said binaries?
18:31redsheep[d]: calico: That was my point with a conversation tool of Nvidia's driver download. Still without redistribution it would never be seamless. Users would have to be aware there's extra steps needed
18:33redsheep[d]: Way I see it basically the only upside of the monumental effort involved in breaking through the signing somehow and writing new firmware would be that in theory in the end it would be seamless. I just can't see anyone ever doing that.
18:49calico: monumental effort? so there's no way to grab the keys at runtime?
18:51calico: from what I can understand the Nouveau driver was mostly written according to analysis of the behaviour of the proprietary driver at runtime
18:52tiredchiku[d]: if there was, we'd have had something in the 8-odd years Pascal's been around
18:52tiredchiku[d]: or is it 9 now
18:52tiredchiku[d]: I forget
18:56snowycoder[d]: calico: If it's really a digital signature, not easily.
18:56snowycoder[d]: The hardware can only check signatures with the public key, and even if we dump that, we can only verify the signature, not generate new signatures.
19:00redsheep[d]: Nearing 9
19:01calico: and there's no way to either to grab the exact 'signed reclocking requests' sent to the gpu by the proprietary driver at time T and just replicate them in Nouveau?
19:26pavlo_kozlenko[d]: Nobody wants to do this.
19:28tiredchiku[d]: pretty sure peeking into the binaries violates the clean-room rules
19:29tiredchiku[d]: or can at least get you into trouble
19:32pavlo_kozlenko[d]: tiredchiku[d]: You say that as if your PC has spyware installed.
19:33pavlo_kozlenko[d]: Make, hack, decrypt
19:33pavlo_kozlenko[d]: You just can't distribute your results.
19:33tiredchiku[d]: why not do it yourself
19:34pavlo_kozlenko[d]: Unfortunately, I have a different profession, at a factory. If I knew how to do it, I wouldn't give a damn about licenses until my results fly out of my space, and they don't fly out
19:35pavlo_kozlenko[d]: I tried before, but due to lack of experience my research stopped.
19:35tiredchiku[d]: and I'm just a student with some free time
19:36tiredchiku[d]: also I've seen people being banned from contributing to wine because they violated the clean-room rules, and I'm not keen on having that happen to me in any case
19:43airlied[d]: mhenning[d]: I think I was hoping the difference might just be in the LDC of the handle rather than in the Tex instr, so I'm not using the bindless forms, it seems like then we will need to distinguish between bound and bindless in some other way for blsckwell
20:51avhe[d]: tiredchiku[d]: wine is somewhat different because they are implementing windows apis verbatim, so they have a high risk of infringing on copyrighted code
20:52avhe[d]: fwiw they also prohibit writing programs that dump internal stuff which is essentially what envyhooks does
21:12tiredchiku[d]: yeah, I'd still not take the chance
21:13tiredchiku[d]: especially since Arthur warned me about not poking into the binaries since it could get me into trouble :P
21:22pavlo_kozlenko[d]: tiredchiku[d]: I can give you access to my computer, so to speak, in honor of science.
21:29gfxstrand[d]: airlied[d]: I'm not sure what you mean. I would expect bindless is just bindless and the new magic LDC thing is for bound.
21:29gfxstrand[d]: So if you're using a 32-bit bindless handle with non-bindless forms that seems like a recipe for trouble.
21:30airlied[d]: I'm not sure I know what I mean either which is why I'm confused 🙂 but bindless is just a handle, and now bound is just a handle, I was somehow hoping the handle magic was the only difference 🙂
21:31airlied[d]: but current code that works is all bound variants, I cleaned that up yesterday
21:31airlied[d]: I just need to work out how to put back bindless properly
21:32gfxstrand[d]: For the bound variants, you're doing the magic LDC and stuffing that into the ureg source?
21:33airlied[d]: yes but doing it as TexRef::Bindless
21:33airlied[d]: I think we need a new TexRef variant
21:33gfxstrand[d]: Ah. Well, we'll need a new TexRef type for Blackwell
21:34gfxstrand[d]: `TexRef::Blackwell` for all I care.
21:34gfxstrand[d]: I do care. I hate that name. But until we grok it, I don't have a better one. 😂
21:35airlied[d]: BoundHandle was where I was thinking
21:35airlied[d]: oh the qmd for hopper is out, not sure it covers blackwell, but we should import updated headers anyways
21:36gfxstrand[d]: airlied[d]: Sure. That's better than `TexRef::Blackwell` and not horribly confusing. I don't love it but it'll work for now.
21:36gfxstrand[d]: We should also figure out what, exactly, the magic LDC is doing
21:36airlied[d]: I'd settle for knowing if it's 64-bit or not 😛
21:37gfxstrand[d]: We should be able to make a unit test for it somehow. I think that will grant us a lot of insight into what's going on with textures.
21:37airlied[d]: since we clearly only create a 32-bit deref, I want to workout where to bump to 64-bit
21:41gfxstrand[d]: FYI: You can do `0.into()` instead of `SrcRef::Zero.into()`.
21:46mhenning[d]: airlied[d]: oh, yeah looks like we have a new "Queue Meta Data, Version 04_00"
21:51airlied[d]: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34440
21:53gfxstrand[d]: airlied[d]: I think this depends a lot on exactly how things work. If we can assume that magic LDC + tex is equivalent to the old cbuf tex ops, then I think we do something like
21:53gfxstrand[d]: ```c++
21:53gfxstrand[d]: uint32_t cbuf = 0;
21:53gfxstrand[d]: nir_def *bound_handle = NULL;
21:53gfxstrand[d]: if (tex_handl_as_cbuf(tex_h, cbuf)) {
21:53gfxstrand[d]: if (nak->sm >= 100) {
21:53gfxstrand[d]: bound_handle = nir_ldc_texunpack_nv(b, ...);
21:53gfxstrand[d]: ref_type = NAK_TEX_REF_TYPE_BOUND_HANDLE;
21:53gfxstrand[d]: tex_h = NULL;
21:53gfxstrand[d]: } else if (has_cbuf_tex(nak)) {
21:53gfxstrand[d]: tex->texture_index = cbuf;
21:53gfxstrand[d]: ref_type = NAK_TEX_REF_TYPE_CBUF;
21:53gfxstrand[d]: tex_h = NULL;
21:53gfxstrand[d]: }
21:53gfxstrand[d]: }
21:53gfxstrand[d]: and then
21:53gfxstrand[d]: ```c++
21:53gfxstrand[d]: if (ref_type == NAK_TEX_REF_TYPE_BOUND_HANDLE) {
21:53gfxstrand[d]: tex->src[num_backend_srcs].src_type = nir_tex_src_texture_handle;
21:53gfxstrand[d]: nir_src_rewrite(&tex->src[num_backend_srcs], bound_handle);
21:53gfxstrand[d]: num_backend_srcs++;
21:53gfxstrand[d]: }
21:55mhenning[d]: QMDv4 looks way different than v3
21:56mhenning[d]: in that almost none of the fields overlap
21:58airlied[d]: yes the redid it completely, my hope is v5 isn't a complete rewrite
21:59mhenning[d]: I wouldn't be too surprised if they left it the same on blackwell if they just redid it for hopper
21:59airlied[d]: also no idea if gb20x can launch a v4 or not
21:59mohamexiety[d]: we can check. I have the pushdump files already of some compute shaders, going to run it in a sec
21:59mohamexiety[d]: I am still not sure if envyhooks will work tbh but only one way to find out
22:02gfxstrand[d]: mhenning[d]: I'm typing QMD fill-out code now
22:06gfxstrand[d]: Ugh... So much stuff is renamed
22:14airlied[d]: I think v5 is different a lot again, as I know at least the version fields moved because I found those
22:21mohamexiety[d]: alright got a dump and ran it through envyhooks. how do I search for QMDs?
22:22gfxstrand[d]: I think it's called INLINE_QMD
22:22gfxstrand[d]: It's got QMD in the name
22:23mohamexiety[d]: found one in one of the text files at least. nothing too important though I think
22:23mohamexiety[d]: [0x00000004] HDR 200120a2 subch 1 NINC
22:23mohamexiety[d]: mthd 0288 NVA0C0_SET_QMD_VERSION
22:23mohamexiety[d]: .CURRENT = (0x50)
22:23mohamexiety[d]: .OLDEST_SUPPORTED = (0x1)
22:26mohamexiety[d]: mohamexiety[d]: hm could it be that the unknown methods following this have the qmd info? :thonk:
22:27mohamexiety[d]: mohamexiety[d]: yeah I can't actually find anything mentioning qmd except these
22:27airlied[d]: no it's uploaded different now
22:28mohamexiety[d]: oh
22:28mohamexiety[d]: where is it?
22:29mohamexiety[d]: if it's in the MME we do have mme dumps too, let me check that
22:30mohamexiety[d]: hm, nothing there actually
22:31airlied[d]: looks for a load inline data with 66 dwords
22:32airlied[d]: on subsch 1
22:32mohamexiety[d]: in the command buffer or the mme?
22:32mohamexiety[d]: oh
22:33airlied[d]: command buffer
22:33mohamexiety[d]: do you mean something like this?
22:33mohamexiety[d]: [0x00000006] HDR 6004206d subch 1 0INC
22:33mohamexiety[d]: mthd 01b4 NVA0C0_LOAD_INLINE_DATA
22:33mohamexiety[d]: .V = (0x40f0000)
22:33mohamexiety[d]: mthd 01b4 NVA0C0_LOAD_INLINE_DATA
22:33mohamexiety[d]: .V = (0x0)
22:33mohamexiety[d]: mthd 01b4 NVA0C0_LOAD_INLINE_DATA
22:33mohamexiety[d]: .V = (0x100)
22:33mohamexiety[d]: mthd 01b4 NVA0C0_LOAD_INLINE_DATA
22:33mohamexiety[d]: .V = (0x0)
22:35mohamexiety[d]: I guess the .v field is the data/qmd, but this doesn't look like 66 dwords
22:37airlied[d]: actually for me it was registers 3bc0/3bc4
22:38airlied[d]: there are 2 dwords of header then the rest is qmd
22:39airlied[d]: I expect maybe blackwell compute class has changed as well
22:41airlied[d]: so it might be a new INLINE_QMD location
22:42mohamexiety[d]: hm
22:42mohamexiety[d]: where do I go from here though
22:43airlied[d]: you work out the QMD v5 fields from those 64 dwords 🙂
22:44mohamexiety[d]: ohhh I seee
22:44mohamexiety[d]: one more thing though, what/where are the dwords? is it the `.v` section, or the numbers after `HDR`?
22:46airlied[d]: all the .V =
22:46mohamexiety[d]: ah got you
22:47mohamexiety[d]: what number does the HDR stuff represent though?
22:47airlied[d]: that's the method, though I might also be wrong on the 66 dwords vs 64, it might be the QMD is 66 dwords and those header bytes matter
22:47airlied[d]: I didn't quite line things up well
22:47mohamexiety[d]: :thonk:: I see
22:48mohamexiety[d]: I do have a lot of slightly different shaders so will see
22:48mohamexiety[d]: thanks!
22:50airlied[d]: https://paste.centos.org/view/raw/122f8fee
22:50airlied[d]: is one I had here, with the small annotations I worked out in 5 mins 🙂
22:52airlied[d]: anything else will need proper RE I expect
22:53mhenning[d]: airlied[d]: I think I'm actually seeing the same thing for the QMDs on ampere: https://gitlab.freedesktop.org/mhenning/re/-/snippets/7835
22:53mhenning[d]: where 3bc0 is NVC7C0_CALL_MME_MACRO(120) and 3bc4 is NVC7C0_CALL_MME_DATA(120)
22:54airlied[d]: oh yeah I picked ADA instead of AMPERE
22:54airlied[d]: using AMPERE gives my proper names
22:54airlied[d]: so yes MME uplaods it is
22:54mohamexiety[d]: yeah I used Ada too
22:54airlied[d]: hence why I think the first 2 dwords are non QMD then
22:55mohamexiety[d]: airlied[d]: but wait now I am confused again, sorry. this is entirely different to what I have
22:55mohamexiety[d]: mohamexiety[d]: like this is the only instance of load_inline_data, and it's not cropped
22:57airlied[d]: ah push dump is missing ADA compute
22:57mhenning[d]: mohamexiety[d]: Right, it looks like it's using the MME instead of LOAD_INLINE_DATA so you should actually grep for 3bc4
22:57airlied[d]: they moved dispatch to mme at some point
22:58mhenning[d]: yeah, and I'm guessing that happened at a specific driver version rather than changing at a specific gen
22:59mohamexiety[d]: mhenning[d]: ahh ok. thanks and sorry
23:00mhenning[d]: mohamexiety[d]: no need to be sorry this stuff is confusing 😛
23:01airlied[d]: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34442
23:01airlied[d]: fix for ada compute dump
23:09airlied[d]: I wonder will I ever be motivated to get a remote hopper machine on loan and test it 😛
23:11mohamexiety[d]: I am just waiting for the $100 H100s on ebay after the AI bubble crashes :LeoKek:
23:15gfxstrand[d]: airlied[d]: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34443
23:16gfxstrand[d]: That should be at least closer for Blackwell (maybe?)
23:19gfxstrand[d]: gfxstrand[d]: IDK how long we want to leave that as a draft. I'm CTSing it now on Ada but none of us have a Hopper to test it on so... <a:shrug_anim:1096500513106841673>
23:21mhenning[d]: Do we know how the QMD version selection works? It looks like hopper has both v3 and v4, so maybe blackwell will have both v4 and v5?
23:22airlied[d]: I don't think we can know wihtout having the blackwell qmd headers
23:22gfxstrand[d]: If they moved the version number as airlied[d] claims, that won't work.
23:22airlied[d]: but I suspect blackwell to be v5 only
23:22gfxstrand[d]: There's a handful of annoying renames in the Hopper QMD. If Blackwell keeps the new names, I may switch it around to use the new names everywhere and make the old QMDs carry the rename. Still. Annoying.
23:23gfxstrand[d]: Also, there are a LOT of QMD fields that we're not setting. I suspect that's going to be a problem for us eventually.
23:24mhenning[d]: gfxstrand[d]: Yeah, even on ampere, I've been wondering if we have a perf impact from not setting some of them
23:26mhenning[d]: Looks like we have these methods pretty consistently:
23:26mhenning[d]: #define NVA1C0_CHECK_COMPUTE_CLASS_VERSION 0x0284
23:26mhenning[d]: #define NVA1C0_CHECK_COMPUTE_CLASS_VERSION_CURRENT 15:0
23:26mhenning[d]: #define NVA1C0_CHECK_COMPUTE_CLASS_VERSION_OLDEST_SUPPORTED 31:16
23:26mhenning[d]: #define NVA1C0_SET_QMD_VERSION 0x0288
23:26mhenning[d]: #define NVA1C0_SET_QMD_VERSION_CURRENT 15:0
23:26mhenning[d]: #define NVA1C0_SET_QMD_VERSION_OLDEST_SUPPORTED 31:16
23:28mhenning[d]: I don't see nvidia using them in the dump I have but it does do a version check for other things like:
23:28mhenning[d]: [0x00000005] HDR 200105a9 subch 0 NINC
23:28mhenning[d]: mthd 16a4 NVC697_SET_SPH_VERSION
23:28mhenning[d]: .CURRENT = (0x4)
23:28mhenning[d]: .OLDEST_SUPPORTED = (0x4)
23:28mhenning[d]: [0x00000007] HDR 200105aa subch 0 NINC
23:28mhenning[d]: mthd 16a8 NVC697_CHECK_SPH_VERSION
23:28mhenning[d]: .CURRENT = (0x4)
23:28mhenning[d]: .OLDEST_SUPPORTED = (0x4)
23:28gfxstrand[d]: It feels like we should probably be calling those....
23:28mhenning[d]: Yeah. We should at least try asking blackwell for v4 and see what happens
23:44redsheep[d]: gfxstrand[d]: Spinning up a hopper instance on some cloud service for an hour or two is maybe an idea. Dunno if any are still on hopper instead of blackwell
23:45redsheep[d]: Though... does running nouveau even work on a virtual machine that doesn't have gpu passthrough to own the whole gpu?
23:45gfxstrand[d]: Some of the clouds do passthrough
23:45gfxstrand[d]: But also, I suspect someone would need it for a lot longer than an hour
23:46redsheep[d]: Yeah last I checked instances with huge AI gpus are REALLY expensive and that's for ones that share it
23:46mhenning[d]: redsheep[d]: The cloud providers tend to keep old generations around for a few years'
23:46redsheep[d]: I can't imagine someone just has a hopper card lying around
23:47gfxstrand[d]: Maybe in 10 years, I can pick one up used on eBay for $1000
23:47HdkR: H100 and B100 also isn't super interesting because of their limited graphics support :D
23:48airlied[d]: I think RH should have a hopper somewhere, but no idea if I can steal it for a few days
23:48airlied[d]: okay first quick hack of that branch + SET_QMD didn't seem to work
23:48orowith2os[d]: airlied[d]: It's always easier to ask for forgiveness than for permission
23:48orowith2os[d]: :thumbsup:
23:49airlied[d]: well it would be automated access, so I have to ask 🙂
23:49mhenning[d]: airlied[d]: Yeah, I haven't gotten SET_QMD to work on ampere either
23:49mohamexiety[d]: HdkR: b100 may be kinda interesting. Spark (the little UMA machine) is derived from sm_100, and I think the newer driver socs/tegras are as well but dont quote me on that
23:49HdkR: Yea, Thor and GB10 are still interesting platforms.
23:50mohamexiety[d]: I'd assume for those they'd have more graphics support than what is exposed on B100 though but not sure :thonk:
23:50HdkR: Same
23:52redsheep[d]: I can't even find an EC2 hopper instance
23:52pavlo_kozlenko[d]: snowycoder[d]: Do you need kepler 2xx (GT 630 rev 2)?
23:53redsheep[d]: His kepler card is kepler B, so would already be 2xx
23:53mhenning[d]: redsheep[d]: https://aws.amazon.com/ec2/instance-types/p5/
23:53mohamexiety[d]: also yup, found global workgroup size! bit trivial but I think this does show we are going on the right track
23:55x512[m]: Another use-case of NVRM KMD backend.
23:56redsheep[d]: Wow yeah that is not the solution, p5 instances are about $100/hr
23:56x512[m]: No need to be obsessed with Nouveau KMD when testing NVK on rental servers.
23:58mohamexiety[d]: man I remember when GV100s were like $12/hr
23:58mohamexiety[d]: oh well