Looking at this: https://dcemulation.org/1-newsdump/QRandom/DC%20stuff/info/150-000105-1e_SegaDreamcast_title.pdfThe dreamcast cost at least $220 each in the year 2000. It was sold at a loss. So here's how Dreamcast hardware can possibly be sold at a profit, while being better equipped to stand alongside other Gen 6 consoles.>GPURendition Verite 2200 instead of PowerVR2. Since it was manufactured in 350nm node, it was cheaper than PowerVR2 (250nm node) and could be fabbed in the same plant that produced the SH4 CPU. But the biggest benefit is the programmable RISC core running the GPU. Verite was the first programmable GPU and games could partially or entirely offload geometry calculations from the CPU to the GPU. This would improve performance in games like GTA 3 and other PS2 generation games. The most fatal downside is the 66MHz clock speed. It was a relatively slow GPU, giving it only slightly better fillrate than the original 3dFx Voodoo. But it's still well above PS1, and nothing smaller textures, lower color depth, lower Z-buffer precision, and lower resolution couldn't fix.>memoryBy the end of 1998, 64Mbit RAM modules actually cost half of 16Mbit RAM (Source: https://phe.rockefeller.edu/LogletLab/DRAM/dram.htm). That means Sega wasted money giving the Dreamcast 8MB VRAM. Coupled with chip packaging costs, Dreamcast should've had 16MB VRAM with the same RAM modules as the CPU, giving it 32MB total. GPUs back then were only set up to use 2-bank SDRAMs, but V2200's specs were completed by early 1997, so they had plenty of time to modify the memory map.
>>12903864cont.>soundTo minize costs, abandon Yamaha and instead use an old SH-2 CPU coupled with a 32KHz-44kHz internal sound DAC (saving chip and board costs). RAM runs at double the CPU clock speed so cache isn't needed, instead replaced by ROM containing ADPCM decompression and sound mixing routines. Devs can write their own decompression and mixing software too. This chip is hooked to a 1MB PC66 SDRAM to shave off costs even further. >controllerPotentiometer analog stick on the left, trackball on the right. Hall effect is pricey. >other thingsNo modem. No VMU.
Finally xisters, the dreamcast is... le savedBut piracy would have killed it anyways, so eh.
>>12903869Piracy saved playstation. Lack of piracy killed N64.
>>12903864Dreamcast was the only 3d console ever made. You want to take that from us and make it yet another obsolete 2d triangle accelerator.
>>12903881PowerVR2 was a 2D triangle accelerator with oversized 32KB SRAM tile buffer to maximize fillrate. Not only it didn't have 3D accelerator, tiled rendering actually causes a ton of geometry overdraw and wastes 3D performance.
>>12903864>>12903865>>12903869>schizo lolcow spam threadgo away, lolcow
>>12903889>tiled rendering actually causes a ton of geometry overdraw>wastes 3D performance.https://blog.imaginationtech.com/a-look-at-the-powervr-graphics-architecture-tile-based-rendering/according to this blog you're wrong. tile based rendering is pretty fuckin based.
>>12903889Unlike every other gpu released before or after, the PowerVR2 is a true 3d gpu, it actually knows what depth is, and draws geometry in the correct order. Your modern gpu has no concept of what depth is, and blindly draws polygons in the exact order you submit them.
>>12903903It minimizes the time spent on texture lookups and fetching, and eliminates the external memory needed for Z-buffer, but this method isn't great at handling triangles that overlap several tiles. Also in case of PowerVR2, there's no on-GPU geometry processing, the CPU had to render the primitives, and that means much higher CPU overhead and poorer performance in CPU intensive games like GTA. That's how Kyro lost to Nvidia and Radeon post 1999. >>12903905>Unlike every other gpu released before or afterYeah nah, it's just another rendering method. Xbox One, Intel Arc Alchemist, and Apple have it. Nowadays GPU cache is so big and VRAM bandwidth is so high that IMR method actually yields more performance than TBDR, or at least >it actually knows what depth is, and draws geometry in the correct order.All GPUs have z-buffer. The only difference is TBDR method breaks each frame into tiles, so it's small enough to store the z-buffer in on-chip RAM. With 1998 technology though, chip density was so low and you needed expensive advanced process to manufacture a GPU chip with on-chip RAM that big.
>>12903967>handling triangles that overlap several tilesthat isn't much of a problem since programmers/3d graphics artists of games would reduce triangle counts to maximize frame rates. it looked ok. and most of the operations for tbdr are simple math operations anyway so it's mostly computationally free that isn't very taxing. perfect for one chip solutions in the 1990s but they had no choice back then with the shockingly low amounts of ram and caches they had to work with.
>>12903967> All GPUs have z-bufferThe z-buffer is not 3d, its a 2d framebuffer hack that fakes 3d in an incredibly limited way. Like go try render some geometry with an alpha texture and then you will see how 3d aware your "3d" gpu is. Modern games have resorted to using Sega Saturn style alpha because modern gpus are so useless at 3d.
Cant fix perfection.
>>12903865>Devs can write their own decompression and mixing software too. And end up with another Genesis situation, where sound quality varies wildly, depending on how good a programmer the composer is? Fuck no. >>12903864>66MHz clock speed You'll get extra inefficiency from the clock ratios not being integer. The PowerVR CLX2 runs at exactly half the SH-4 clock.
>>12903864>>12903865>Western CPU instead of JapaneseYAMEROOO SEGA'S HONERU IS ATU SUTAKUUU
but would this save the Saturn!?
>>12903986Didn't say it looks bad, PowerVR2 was pretty much the best choice for graphic fidelity back then. Just saying, without additional on-GPU triangle processing, Dreamcast would really struggle PS2 ports. >>12904074>And end up with another Genesis situation, where sound quality varies wildly, depending on how good a programmer the composer is?Genesis doesn't have enough processing power for ADPCM and software mixing. >You'll get extra inefficiency from the clock ratios not being integerSystems with 1:3 clock ratio do exist. Verite works differently from CLX2, being a programmable GPU composed of multiple internal processors working asynchronously. AFAIK it doesn't need to run at VRAM clock speed like the PowerVR Neon 250. Seems vendors could overclock or underclock it without changing the memory specs, like a modern GPU. I have no hardware documentation or deep dive for it though.
>>12903864The verite's lack of a z-buffer might make it a non-starter. I know the V1000 didn't have one at all and relied on software depth sorting or span buffering, but I can't find an authoritative source that says they properly fixed that in the V2200. The chips are supposedly identical, flaws and all, just faster and I think people are confused that a combination of higher clock speeds and faster host CPUs meant the z-buffer emulation was "fixed", (except let's just ignore all those mediocre benchmarks that are totally not z-buffer related.)If you're going to have a cheap CPU, you can't power out of a hardware flaw that massive. Software depth tests limit your fill rate to PS1 grade. It wouldn't be an upgrade.
>>12904018With modern GPUs it's the "2000 shader cores" problem. One transparent polygon can stall hundreds of cores because it has to be rendered at the correct depth. It's perfectly fine IF the game dev makes sure to minimise draw order stalls but most devs are so lazy they just use mesh rendering and let TAA/DLSS "fix" it. Tile based rendering wouldn't fix this problem because instead of stalling mid-frame you're just stalling at the top of frame. It was perfectly adequate in the DC days before GPUs changed from single core to thousands of slow cores but modern tile based solutions are more hybrid because you can't have a 10GHz GPU core to keep pace with 10-way 1GHz cores.
>>12905905>but I can't find an authoritative source that says they properly fixed that in the V2200Every source says V2200's pixel processor has a hardware z-buffer now. Many games running on V2200 suffered performance issues due to the rendering software being made for V1000 which doesn't utilize the newly added z-buffer feature, so that alone proves there are hardware changes, not just clock speed. >The chips are supposedly identical, flaws and all, just faster and I think people are confused that a combination of higher clock speeds and faster host CPUs meant the z-buffer emulation was "fixed", (except let's just ignore all those mediocre benchmarks that are totally not z-buffer related.)Nope. The pixel processors in both are clocked the same. Only in V2200 the RISC core is clocked twice as fast. Also moving from EDO RAM to SDRAM gave performance boost for sure. And again the driver issues provides some proof that that it's not just a higher clocked chip. It wouldn't be this bad if it was. >If you're going to have a cheap CPU, you can't power out of a hardware flaw that massive. Software depth tests limit your fill rate to PS1 grade. It wouldn't be an upgrade.Not being able to offload geometry processing from the CPU limits your performance to PS1 grade. Even N64 could do it, even though the unified 9-bit memory without any CPU/GPU L2 cache badly bottlenecked the performance.
>>12904094Saving saturd was easier but sega took at a look at snes' complicated architecture and chose the hard way.
