[HN Gopher] AMD's RX 7600: Small RDNA 3 Appears ___________________________________________________________________ AMD's RX 7600: Small RDNA 3 Appears Author : picture Score : 60 points Date : 2023-06-04 17:55 UTC (5 hours ago) (HTM) web link (chipsandcheese.com) (TXT) w3m dump (chipsandcheese.com) | gyudin wrote: | Yet their drivers still cause kernel panic and crashes even on | their own demos lol | mastax wrote: | > It uses TSMC's 6 nm process, which won't provide the | performance and power savings that the cutting edge 5 nm process | would. In prior generations, AMD fabbed x60 and x600 series cards | on the same cutting edge node as their higher end counterparts, | and used the same architecture too. However, they clearly felt a | sharp need to save costs with this generation, compromising the | RX 7600 a bit more than would be expected for a midrange GPU. | | It's also so they only need to design one memory controller for | 6nm. I believe I remember this corroborated from an AMD engineer | interview around the 7900XTX launch. Memory controllers aren't | just logic that can be "compiled" to whatever target node. They | have specific electrical requirements that take substantial | design work. For this generation AMD has a 6nm memory controller | that they use both in this 6nm monolithic design and in their 6nm | memory controller chiplets on the larger designs. | rowanG077 wrote: | It's analogue circuitry at the chip boundary far reaching into | RF. It isn't like the high-level digital code like Verilog/VHDL | which is relatively straightforward to port. | | The people who do that are the true gurus honestly. | speed_spread wrote: | The 8GB VRAM size limitation is the biggest downside for cheap | local AI. My feelings is that current LLMs start producing | interesting results from 12GB and above. | kramerger wrote: | It's "RDNA lite" for somewhat cheaper GPUs. It is not supposed | to have tons of VRAM. | | Besides, the _really_ interesting LLMs use far far more than | 12GB. But that sort of this changes from day to day here... | cypress66 wrote: | You're expecting too much from a cheap card. AMD offers many | 16GB cards, and used RTX 3090s (24GB) are very affordable and | are the go to for AI. | tigeroil wrote: | What do you consider to be affordable? In the UK used RTX | 3090s still go for well over 1k. | LaurensBER wrote: | For a company that's very affordable, for an individual | we're not there yet but I've seen some impressive demos | running on a M1/M2 laptop so no doubt time will bring down | the hardware requirements even more (and hopefully the | price of used 3090s as well). | epolanski wrote: | Cards in the same bracket offered the same amount of VRAM 8 | years ago to be fair. | smcleod wrote: | Used RTX cards (of any spec) are super expensive here in | Australia still. | ChuckNorris89 wrote: | Here's where the unified memory architecture of M1 Macs, gaming | consoles, APUs, and other custom SoCs can shine and shows the | inefficiencies of the traditional ageing desktop PC | architecture in the modern world where the graphics VRAM is | completely separated from the system RAM. | | I'd like to see PCs designed more like consoles or M1 Macs, | with GDDR shared as unified RAM between GPU and CPU. I have an | laptop with a last gen AMD APU that's no slouch, but the VRAM | slice of total RAM is still fixed as configured in BIOS between | 512MB and 4GB, instead of fully unified and dynamically shred | by the OS, which seems highly ineficient and wasteful in the | modern age. | | It's why PS5 and Xbox with their 16GB of fully unified VRAM can | compete with gaming PCS witch need 16GB of system RAM and over | 8GB of VRAM. Why have two separate memory zones where one sists | empty most of the time and when needed is actually too small | while the other is half empty, when you can unify them and make | use of the whole pie as needed? | coffeebeqn wrote: | I can't imagine the SOCs are too far away for PCs. It does go | against the upgradeability philosophy but I'm not sure how | many users really care about that | throw9away6 wrote: | Who really upgrades pcs anymore. By the time you need to | upgrade everything from socket to power standard is | outdated | beebeepka wrote: | Certainly not the case with AMD AM4. Stats show that | plenty of people have upgraded their AM4 builds. | | The Intel only crowd is such a sad bunch | heavyset_go wrote: | I'd disagree if AM5 follows the path of AM4. | bee_rider wrote: | A mid-lifetime upgrade where you pop in a new GPU, fill | up any unpopulated ram slots, or add in a new drive, can | be nice. CPUs last a really long time nowadays. | throw9away6 wrote: | So do gpus though. Are you really going to pop in a 800$ | gpu into 200$ worth of components? | coffeebeqn wrote: | Depends how good the original was if it's worthwhile. | GPUs don't need that much from the rest of the machine as | long as it's somewhat sensible for gaming. Especially at | 1440p or 4k the GPU is a huge portion of the performance | yjftsjthsd-h wrote: | > Are you really going to pop in a 800$ gpu into 200$ | worth of components? | | Why not? For the right workloads, the GPU is the | important thing anyways. It's not _that_ different from | rotating disks through a machine. | kramerger wrote: | Is this different from integerated GPUs from Intel and AMD? | ChuckNorris89 wrote: | As exemplified in my post above, AMD integrated GPUs | don't have unified memory, at least my 5000 series I | bought in 2022. Maybe the new 7000 series changes this. | fulafel wrote: | They have had it for about 10 years, since Kaveri. See eg | https://hexus.net/tech/news/cpu/54709-amds-heterogeneous- | uni... or https://www.guru3d.com/articles- | pages/amd-a10-7800-kaveri-ap... | ChuckNorris89 wrote: | Not exactly. While that may have been the case way back | in the past when AMD APUs were designed as a homonegnous | CPU-GPU unit from the start, I can tell you that's not | the case with the relatively modern 5000 series. | | I need to go in the Bios and specify explicitly how much | of the system RAM I want allocated exclusively of the | integrated GPU and the rest stays available for system | RAM. | | The reason is because Ryzen 5000 APUs seem to be a job | rushed out the door so they're just a Zen 3 CPU with a | separate Vega GPU glued together on the same die, but | they're not a homogenous design, designed to work as one | unit, like the APU of the PS5, so memory wise they're | unaware of each other, even though AMD calls them APUs, | they're not really, but more like separate CPU and GPU on | the same die. | | I wish I knew about this limitation at the time, as Intel | chips with integrated graphics have unified memory. | yjftsjthsd-h wrote: | I wonder if a clever enough driver+firmware could make | that hot-adjustable (or truly unified, rather) with | nothing but a software patch. It sounds like the pieces | are all there. | tsss wrote: | I imagine if that ever happens it will only be used to nickel | and dime consumers and you'll end up paying even more through | forced upselling than with the inefficient VRAM we currently | have. | bee_rider wrote: | The main argument I can think of for separating their memory | is that, of course, GDDR can be optimized for bandwidth and | regular ram can be optimized for latency or somewhere in | between. | | But, memory latency continues to make poor progress compared | to CPU speeds. So, since we're already going to need a | complicated system of caches on the CPU side, maybe it is not | such a big deal if CPU memory acts more like GDDR. | rektide wrote: | Almost twice the bandwidth of the rx580, from April 2017. About | the same msrp. | | It'll be interesting to see how this holds up versus Intel, whose | been doing pretty ok at this price point. Intel's initial launch | was pretty rocky but the drivers have gotten much much faster | already. ___________________________________________________________________ (page generated 2023-06-04 23:00 UTC)