New GPU Comparison - RTX4000 Ada SFF

Comments

• 3DSaga Posts: 712

  2:44AM

  I would say the 3080Ti acquits itself quite well. It's significantly faster, but that's just one specification. Would a 4060 or 4070 be a closer comparison?
• Gordig Posts: 10,652

  3:28AM

  Not terribly surprising. Even with the newer CUDA architecture, that is quite a difference in the number of cores, and also the workstation cards are designed more for reliability than speed.
• TimberWolf Posts: 313

  3:50AM

  Gordig has this correctly. Workstation cards have a lower TDP, lower thermals, the components are parts-binned (which means tested individually to fall within certain performance parameters) and will almost certainly outlast the nearest equivalent gaming GPU by a significant factor. The trade-off is lower performance on a roughly equivalent architecture. Don't be disappointed - lower electricity bills, longer time between replacement. Small commercial studios like mine will always choose lower bills and reliability over faster rendering times. Pixar we are not.

  I did notice you said you bought a SFF card, but is there room for both GPUs in your system? Use your 3080 to drive your monitors and save at least 1Gb of VRAM on your 4000 and, if your render fits into both cards, excellent performance. If it only fits on the 4000, nothing lost.
• Matt_Castle Posts: 3,061

  4:04AM edited 4:08AM

  This information might be aggravating and unhelpful at this stage, but a key thing to understand is that workstation cards are not designed to prioritise speed, they're designed to prioritise reliability.

  Their Error Correcting RAM is only one part of that; they also run their cores less hard than the equivalent GPU dies on gaming cards, in order to minimise any potential for errors. If you've ever looked at overclocking a GPU, then it's not actually particularly dangerous from the perspective of frying your card - the card becomes unstable and unreliable long before you're at risk of doing permanent damage to the hardware. But it's not like there's a precise point that is "too far"; the error rate is a matter of probabilties. The slower the core is run, the more the error probability dwindles. (Lower heat output can also be a consideration, as these cards are often in server farms).

  As such, their performance can be a significant fraction below the gaming cards with the closest core configurations. The closest gaming equivalent die to the RTX4000 Ada SFF is the RTX 4070, and the 4070 runs at a Boost Clock of 2475 MHz compared to the 4000 SFF's 1560 MHz (and in that respect, it's worth noting that as a small form factor card, the SFF is being run less hard again than the regular RTX4000 Ada, which has a boost clock of 2175 MHZ).

  So yes, the workstation cards aren't just the gaming cards with more VRAM, they're tuned differently as reflects their more critical role - we're talking about cases where a data error could compromise a multi-million dollar film project or mess up weeks of scientific simulation. It's quite a different case to gaming, where frame rates are more important and a crash is usually merely frustrating rather than disastrous.

  EDIT: I see that Gordig has said something similar, if less lengthily, while my fractured attention was taking too long to write this post.

   

  Post edited by Matt_Castle at 4:08AM
• kyoto kid Posts: 42,050

  10:06AM

  ...the choice is more VRAM "overhead" v. more raw speed. Hiwver should the proces dump to the CPU if it exceeds VRAM all those extra cores and clock speed mean nothing. I often had scenes push close to 11 GB (I have a 12 GB Titan-X).

  While it is t rue that elements in a scene (partcularly textures) can be optimised to better fit within GPU memory however that still takes extra time.

  One advantage of the 4000 ADA SFF (Small Form Factor) is that ot runs much cooler as it consumes far less power than the 3080Ti - 70w v. 350w respectively (which is ⅕ the power consumption).  This low power rating has a price though, and that is render performance.

  A skightly better comparison would have been against a full length 4000 ADA which is 25 - 30% better in rendering performance than the SFF version while it still uses significantly less power than the 3080 TI (130w).  The lower performance factor of the SFF version is that it draws its power solely from the PCIe slot whereas the standard 4000 ADA uses a 16 pin connector to the PSU.
• Gordig Posts: 10,652

  10:37AM

  kyoto kid said:

  The lower performance factor of the SFF version is that it draws its power solely from the PCIe slot whereas the standard 4000 ADA uses a 16 pin connector to the PSU.

  Apparently I'd missed that point in all my looking into the various options, so thanks for bringing that to my attention.