|
FuturePastNow posted:I suppose AMD's approach to making smaller cores by simply giving them less cache will work better for PCs, especially since AMD can glue more cache on top of them for gamer versions If you're referring to Zen 4c and its successors, one of the ways they shrank the cores was to remove the space used for 3D v-cache TSVs.
|
# ? Jul 30, 2023 20:53 |
|
|
# ? Jun 8, 2024 04:16 |
|
Klyith posted:It's not just less cache. I saw a neat article about it, a whole lot of what makes them compact is just removal of "extra" space. Some of that means that the compact cores have to run at lower frequencies, due to signal interference between components that are now closer together, and also power density. re: signal interference etc, kind of but not really? Especially the signal interference part. You need to flip your mental model over: AMD chose to make a lower frequency version of the core because (a) lots of servers want core count more than high clocks and (b) clock speed costs area and power, so reducing it enables them to create a variant better optimized for such servers. I'll try and explain why B is true... That article claims AMD uses the same RTL for 4 and 4c and I'll take it at face value. RTL is a cycle-accurate model written in a high level language (usually Verilog) in a style which is amenable to translation (usually automated, sometimes augmented by human effort) to a netlist - a big list of logic gates and the wires which connect them together. This netlist is the input to physical design - placement and routing. So how can the same RTL be translated into less or more area, depending on frequency target? Well, for any given logic gate, physical design tools can choose between several different versions in the cell library. These variants all do the same logical function (e.g. 5-input NAND gate, or whatever), the difference lies in the tradeoff between area and output drive strength. High-current output drivers cost more area. Why have multiple drive strengths? Each gate's output has to drive a wire which goes somewhere. Wire capacitance scales with wire length and loads, so the longer and/or more heavily loaded the wire is, the harder the gate output has to yank on it to quickly move it between logic 0 and logic 1. With a low speed target, most gates can be the minimum size versions. Doesn't matter if the wire takes its sweet time moving around, there's plenty of slack in the timing budget. As target clock speed goes up, so does the number of gates which have to be bigger versions of themselves. This even has a (usually mild) cascading effect: the extra area used by larger gates increases average wire length, which upsizes even more gates. In some cases, extra long wires may need repeaters inserted in the middle. This is how speed literally costs area. (And power - longer wires switched at higher speed burns lots more power than shorter / slower.) The modularity thing is not a first-run thing to improve speed of prototyping, it's a side effect of chasing clock speed. Trying to push Fmax to the limit creates lots more physical design optimization work. Since engineers work best in small teams, most of the time the design gets broken down into smaller chunks, each of which is assigned its own team and some floorplan to play inside. Nobody's perfect at predicting the ideal area and layout for each block, so there's usually some give-and-take iterations during the design process. The end result is never ideal on area efficiency, but it's faster. However, when you choose to reduce speed targets (as in Zen 4c), you can greatly reduce the amount of labor put into physical design optimization, live with fewer partitions, and as a side effect gain some area back. Re: normal and compact at the same time, it depends on the size of AMD's labor pool. The compact variant doesn't actually need the big variant to go first, and should require fewer engineer hours, but if they both have to be done in parallel that still needs more engineers on staff.
|
# ? Jul 31, 2023 01:23 |
|
https://twitter.com/Olrak29_/status/1685343071213604864 With the possible existence of a 4 + 8 part, I wonder if this means that they're going back to four-core CCXes (or eight for 5c cores?) or if there's going to be multiple CCX sizes now. It seems like one of the issues Zen 3 and 4 have run into has been power efficiency for devices at 15W and under, and my suspicion is that the eight-core CCX that has to be in every single CPU is a part of the problem (and perhaps that's why Mendocino uses Zen 2 still). Being able to do proper four-core parts with no wasted silicon would be good for handhelds and other low-power devices. Strix Halo seems kind of insane, based on everything that's been leaked about it so far. 16 full Zen 5 cores with up to 40 RDNA 3.5 CUs is a shitload of performance for an APU.
|
# ? Jul 31, 2023 05:19 |
|
Dr. Video Games 0031 posted:https://twitter.com/Olrak29_/status/1685343071213604864 This would also explain why AMD has finally sunsetted the G-series ultra-low-power SOCs. If they can spit out 4-core Zen 5/5c CCXes that can operate under 10 watts then there's no point in continuing to produce Jaguar dies in 2024 and beyond.
|
# ? Jul 31, 2023 05:22 |
|
If it's a mobile part Strix Halo is going to need 4 (8) DDR5 channels to even have a hell of a chance of feeding that many CUs (and you really don't find all that many 15.6" laptops with 4 sodimm slots), or a GDDR6/7 controller, which means the memory has to be soldered and is unserviceable. Would be really interesting to see exactly what the hell is going on there.
|
# ? Jul 31, 2023 07:56 |
|
Anime Schoolgirl posted:If it's a mobile part Strix Halo is going to need 4 (8) DDR5 channels to even have a hell of a chance of feeding that many CUs (and you really don't find all that many 15.6" laptops with 4 sodimm slots), or a GDDR6/7 controller, which means the memory has to be soldered and is unserviceable. Would be really interesting to see exactly what the hell is going on there. It will supposedly have a 256-bit LPDDR5X interface (so, 8 x 32-bit channels). So yeah, it'll probably have over 200GB/s of memory bandwidth available (7200 MT/s would be 230GB/s), but lpddrx is also soldered. Dr. Video Games 0031 fucked around with this message at 08:28 on Jul 31, 2023 |
# ? Jul 31, 2023 08:04 |
|
Is this going into a new Steam machine?
|
# ? Jul 31, 2023 15:15 |
|
Probably. 20cu integrated graphics with 8 channels of soldered RAM is a game console chip. I wonder if it's low power enough for a handheld.
|
# ? Jul 31, 2023 15:34 |
|
If not a handheld then companies like Minisforum or Beelink are going to be able to make some real beefy mini PCs with a chip like that.
|
# ? Jul 31, 2023 15:42 |
|
just went from a 3700X to a 7700X, with a Deepcool AK620 on there. my understanding for the 7000 series is that it's pretty much intended for the chip to shoot right up to 95C under full load (especially with air cooling) and run there consistently? ran Cinebench R23 and it went from ~51C light load up to 95C and stuck there. looking at Adrenalin, it looks like it stayed consistent at 5.0GHz which sounds about right for multi-core based on what I've been reading since the advertised 5.4 is mostly for single-core boost, but I get neurotic about my thermal paste/cooler application with a new build so I wanted to make sure this seems right
|
# ? Jul 31, 2023 16:26 |
|
Rinkles posted:Is this going into a new Steam machine? It's going into laptops but I'm sure it'll make it's way into pre-builts down the line
|
# ? Jul 31, 2023 16:30 |
|
BobHoward posted:normal and compact at the same time Neato, thanks for the writeup. So a 5/5c hybrid sounds much more plausible that I thought. barnold posted:just went from a 3700X to a 7700X, with a Deepcool AK620 on there. my understanding for the 7000 series is that it's pretty much intended for the chip to shoot right up to 95C under full load (especially with air cooling) and run there consistently? ran Cinebench R23 and it went from ~51C light load up to 95C and stuck there. looking at Adrenalin, it looks like it stayed consistent at 5.0GHz which sounds about right for multi-core based on what I've been reading since the advertised 5.4 is mostly for single-core boost, but I get neurotic about my thermal paste/cooler application with a new build so I wanted to make sure this seems right Yup. Even liquid cooling will only keep it in the low 90s.
|
# ? Jul 31, 2023 16:51 |
|
FuturePastNow posted:Probably. 20cu integrated graphics with 8 channels of soldered RAM is a game console chip. I wonder if it's low power enough for a handheld. The smaller 4+8 Strix Point variant is supposedly 16 CUs, vs 12 CU in the 7840U, both on the same manufacturing process. The TDP range for the 7840U is 15-30W, and Strix Point is rumored to be 30-45W. I think it is more likely to be an "H" class part than the "U" class part that a handheld would require... but maybe with some big downclocks? The next "U" class part is scheduled to be Hawk Point, which appear to be a refresh of the 7840U with an updated GPU. Strix Point seems to be carving out a new middle ground in mobile for AMD with its own mask. Strix Halo is 8+8 and 40 CUs on TSMC 3 (supposedly, i could see it staying on 4nm). Even with a new process, I don't think there is any way that fits in a handheld power budget. Cygni fucked around with this message at 17:39 on Jul 31, 2023 |
# ? Jul 31, 2023 17:35 |
|
I just picked up the 7700X + 32gb - Pro B650-P Wifi bundle from Microcenter. You know, the $400 one. Except it cost me $230 plus tax. I asked them if that was the right price and they said yes. At this point if you have a vague interest in a new PC I suggest getting the gently caress over to Microcenter before they fix whatever is causing them to sell you that bundle at that price.
|
# ? Jul 31, 2023 18:23 |
|
Cygni posted:The smaller 4+8 Strix Point variant is supposedly 16 CUs, vs 12 CU in the 7840U, both on the same manufacturing process. The TDP range for the 7840U is 15-30W, and Strix Point is rumored to be 30-45W. I think it is more likely to be an "H" class part than the "U" class part that a handheld would require... but maybe with some big downclocks? The next "U" class part is scheduled to be Hawk Point, which appear to be a refresh of the 7840U with an updated GPU. Strix Point seems to be carving out a new middle ground in mobile for AMD with its own mask. Neat. Yeah that's definitely a game console chip. By comparison, the XBox Series X is 56 CU and the Series S has 20, which are RDNA2 of course. Even for laptops that might be a bit much though gamer laptops can get pretty chonky still. I wonder if Gabe wants to take another go at the Steam Machine.
|
# ? Jul 31, 2023 18:31 |
|
K8.0 posted:I just picked up the 7700X + 32gb - Pro B650-P Wifi bundle from Microcenter. You know, the $400 one.
|
# ? Jul 31, 2023 18:56 |
|
Rinkles posted:Is this going into a new Steam machine? Strix Halo is too big to go in a handheld. Strix Point may be too big too, tbh. According to the MLID leak that first revealed these chips (several points have been confirmed already so it's probably legit), Strix Point (4+8 cpu and 16 CUs) will have a TDP of around 28 to 54 watts and Strix Halo (16 full cpu cores and 40 CUs) will be 55 - 120W. These seem to be primarily meant for laptops. AMD aims to replace low-end and midrange discrete solutions with APUs, and Strix Halo may be positioned as a chip for x86 Macbook alternatives. Strix Point will also have a more typical 128-bit memory bus. Honestly, I can see both still running into memory bandwidth issues unless AMD's mobile Zen 5 memory controller can handle really high speeds (over 8000 MT/s would be nice) The sub-28W range for Zen 5 APUs hasn't been leaked yet, but that's likely to be a different thing than these, with fewer CPU cores. edit: Here's the table of current known and rumored information provided by videocardz: Hawk Point is just a Phoenix refresh, and I guess Strix Point is the phoenix successor? I get that they probably aren't designing these specifically with handhelds in mind, but 4+8 sounds like a bad cpu config for handhelds to me. It's just too big for ultra low power, so maybe they'll release a different chip for ULP devices. I also doubt Strix Halo will be a console chip. I think the PS5 Pro will want something different. Probably 8 CPU cores again with more CUs, as is currently rumored. Dr. Video Games 0031 fucked around with this message at 19:41 on Jul 31, 2023 |
# ? Jul 31, 2023 19:00 |
|
K8.0 posted:I just picked up the 7700X + 32gb - Pro B650-P Wifi bundle from Microcenter. You know, the $400 one. gently caress, we really need to get a micro center in Phoenix. It’s a barren wasteland since Fry’s effectively ceased functioning*. * I’m not counting the several years with poor inventory and consignment status while technically still “open”
|
# ? Jul 31, 2023 19:12 |
|
K8.0 posted:I just picked up the 7700X + 32gb - Pro B650-P Wifi bundle from Microcenter. You know, the $400 one. What location?
|
# ? Jul 31, 2023 20:49 |
|
K8.0 posted:I just picked up the 7700X + 32gb - Pro B650-P Wifi bundle from Microcenter. You know, the $400 one.
|
# ? Jul 31, 2023 20:57 |
|
bobfather posted:What location? Cambridge. Though, whatever caused it, I don't see why location would matter. I told a girl I wanted to buy the 7700X+ram+motherboard bundle, she went over to a computer and punched all the parts in, clearly having sold a bunch of them, printed out a list sheet, and handed it to me. I said are you sure this is right, I thought it was more? She insisted it was the right price and left to grab the CPU + ram. I was more concerned maybe I was buying the wrong stuff but it seemed to be the right things. Rung up the same price so I said huh, I guess it is right. Then after I left and worked a few other things off my to-do list I looked at the sheet + receipt and thought "Wait, what? $75.86 motherboard, $34.48 for 32gb of DDR5, and $119.65 for a 7700X?" The whole thing is handled through a web app and I don't think there's anything store-specific in that process, IDK if it was just a glitch that double-applied the bundle discount or something?
|
# ? Jul 31, 2023 21:18 |
|
that def seems to be a pricing error, but hey, winners win!
|
# ? Jul 31, 2023 21:27 |
|
Can someone explain why AMD chips are now ASUS branded, did they help design them in some way
|
# ? Jul 31, 2023 21:29 |
|
What, who doesn’t want IHS sponsorships?
|
# ? Jul 31, 2023 21:32 |
|
K8.0 posted:Cambridge. Though, whatever caused it, I don't see why location would matter. I told a girl I wanted to buy the 7700X+ram+motherboard bundle, she went over to a computer and punched all the parts in, clearly having sold a bunch of them, printed out a list sheet, and handed it to me. I said are you sure this is right, I thought it was more? She insisted it was the right price and left to grab the CPU + ram. I was more concerned maybe I was buying the wrong stuff but it seemed to be the right things. Rung up the same price so I said huh, I guess it is right. Then after I left and worked a few other things off my to-do list I looked at the sheet + receipt and thought "Wait, what? $75.86 motherboard, $34.48 for 32gb of DDR5, and $119.65 for a 7700X?" The whole thing is handled through a web app and I don't think there's anything store-specific in that process, IDK if it was just a glitch that double-applied the bundle discount or something? just bought this bundle from the same MC yesterday and I didn't get this discount so I would like half the money you should have spent, sent directly to me ty
|
# ? Jul 31, 2023 22:22 |
|
Klyith posted:Neato, thanks for the writeup. So a 5/5c hybrid sounds much more plausible that I thought. Yeah. Don't depend on the fine details of what I wrote, I'm reading a lot between the lines of that article and making tons of guesses/assumptions about AMD's engineering process, but I'm very confident there's no technical barrier which would block AMD from launching 5 and 5c at the same time, possibly on the same chip. It's just more expensive. (But note: this should still be far less expensive than designing a new small core with a different uarch. Starting from the same RTL saves them a lot of effort on both the design and validation sides.)
|
# ? Jul 31, 2023 22:33 |
|
Dr. Video Games 0031 posted:Strix Halo is too big to go in a handheld. Strix Point may be too big too, tbh. I should have capitalized. I was thinking of a Valve lead reboot of this: wikipedia.org/wiki/Steam_Machine_(computer)
|
# ? Jul 31, 2023 22:53 |
|
https://twitter.com/VideoCardz/status/1686069589245902871?s=20 The 8 inch screen would make it stand out from the crowd of AMD handhelds
|
# ? Aug 1, 2023 00:20 |
|
This is an area Intel can’t compete in, right?
|
# ? Aug 1, 2023 00:37 |
|
Rinkles posted:This is an area Intel can’t compete in, right? They can now that they have a better, semi-functional GPU architecture. There are also plausible rumors pointing to Intel releasing SoCs with big iGPUs based on alchemist and/or battlemage.
|
# ? Aug 1, 2023 00:52 |
|
I thought the problem was with power efficiency.
|
# ? Aug 1, 2023 00:54 |
|
Rinkles posted:I thought the problem was with power efficiency. Intel is about to ship a multi-fab chiplet wackiness part for mobile, including using TSMC N5 for graphics and intel's 4 for the CPU section. In contrast, Strix Point is monolithic on TSMC N4 (supposedly), and Zen 5 is the largest architectural change for AMD since Zen 1. So I think a lot of things are still in the air as to what the future looks like. This is the big show down people have been waiting years to see play out though. Frankly, I'm still not sold on the "big APU" concept to begin with, and the XPU train seems to have stalled before it left the station, although that might be more to technological limitations than market demand.
|
# ? Aug 1, 2023 01:17 |
|
Rinkles posted:I thought the problem was with power efficiency. The performance per watt of the desktop Arc GPUs wasn't terrible. They had problems with high idle power draw but that was all in the drivers being bad for quite a while. The same arch is used in Intel Xe mobile CPU integrated graphics already. Just with like 1/5th the execution units of the A750 or A770. Cygni posted:Frankly, I'm still not sold on the "big APU" concept to begin with, and the XPU train seems to have stalled before it left the station, although that might be more to technological limitations than market demand. I also don't know how big the market segment is really gonna be, and whether Intel needs to even care. Like the steam deck is cool and a great success, but IMO it very well could be a novelty more than a major product category. Klyith fucked around with this message at 01:30 on Aug 1, 2023 |
# ? Aug 1, 2023 01:25 |
|
Cygni posted:Intel is about to ship a multi-fab chiplet wackiness part for mobile, including using TSMC N5 for graphics and intel's 4 for the CPU section. That sounds expensive. Klyith posted:The performance per watt of the desktop Arc GPUs wasn't terrible. They had problems with high idle power draw but that was all in the drivers being bad for quite a while. The same arch is used in Intel Xe mobile CPU integrated graphics already. Just with like 1/5th the execution units of the A750 or A770. I was basing that on what I’ve heard about laptop battery life and thermals.
|
# ? Aug 1, 2023 01:28 |
|
I guess that 5600x3d deal was just too good and zen 4 had to be priced similarly barnold posted:just bought this bundle from the same MC yesterday and I didn't get this discount so I would like half the money you should have spent, sent directly to me ty Microcenter updates pricing sundays so I guess Monday morning must be your best bet to roll the dice on errors in your favor?
|
# ? Aug 1, 2023 01:30 |
|
PSA for people buying Starfield promo AMD CPUs from Microcenter. I didn’t get a redemption code at the register (not sure if they’re supposed to provide one there or not). I went onto Microcenter’s website and used their live chat tool to message a rep. They confirmed the phone number tied to my microcenter account, then sent me my starfield code to an email address I provided. So yeah, simple fix. Now I can get hyped for this game since I didn’t have to pay for it.
|
# ? Aug 1, 2023 01:47 |
|
Klyith posted:For gently caress's sake, just make an E core that can run 512 even if it takes ten cycles to complete an operation and makes the E core emit an audible shrieking noise. Whatever thread is doing that will peg the E core to 100% load and get shifted to a P-core ASAP. The scheduler knows how to handle that. The most pessimistic case would be a program that infrequently uses a burst of 512 ops on an otherwise low-budget thread, but that seems... unlikely. Again, every thread will be The core problem is that you have an operation that's conditionally faster and the condition is normally something you check at program start time: do I have 512bit operations available? Now you need some way to tell the program "yeah they're available but don't use them unless it's for media processing, even though not using them makes you slower." or worse, "ok but every time you are about to compare a string check this undocumented opcode result to see what core you're currently running on" or something ridiculous like that. The solution is just don't do asymmetric cores on a general purpose OS. It's only something you can do when you control the entire platform and design your different tasks around the available resources. Harik fucked around with this message at 04:52 on Aug 1, 2023 |
# ? Aug 1, 2023 04:44 |
|
AMD's 4/4c approach is so much cleaner in that regard. It's sad really how no sooner than AMD starts supporting it, Intel dropped support for AVX-512. Instead they're introducing a new instruction set just to handle the mess big/little made of AVX-512 adoption.
|
# ? Aug 1, 2023 08:56 |
|
I for one am very excited for powerful APUs
|
# ? Aug 1, 2023 09:02 |
|
|
# ? Jun 8, 2024 04:16 |
|
ConanTheLibrarian posted:AMD's 4/4c approach is so much cleaner in that regard. It's sad really how no sooner than AMD starts supporting it, Intel dropped support for AVX-512. Instead they're introducing a new instruction set just to handle the mess big/little made of AVX-512 adoption. i do like getting to tell people "to use intel's banner feature you have to buy an AMD chip" because it's all just so dumb.
|
# ? Aug 1, 2023 12:07 |