Overclock.net - An Overclocking Community - View Single Post - Sffpc build recommendation for rendering

View Single Post
post #9 of (permalink) Old 08-06-2019, 04:15 PM - Thread Starter
New to Overclock.net
kevindd992002's Avatar
Join Date: Apr 2009
Location: Philippines
Posts: 10,975
Rep: 85 (Unique: 67)
Quote: Originally Posted by AlphaC View Post
RTX 2070 has more CUDA cores and more Ray tracing and tensor cores, it's outright better (than RTX 2060 Super). Because you're in the Philippines the pricing is the main thing to consider.

As for pro GPU: Not for 3dsmax or Autocad. Sketchup is a consumer application as well since they don't mention anything on workstation GPUs.

In your case there's no way you're going to fit a proper $400 GPU in there if you go with Quadro as the cheaper (more affordable rather, <$1K) Quadros are a generation old and have much less performance over two to three times less based on FP32 TFLOPs) for the main applications used.

About Quadros:

Quadros are only worth it when you are thermal/power constrained , need professional drivers, and can get a Quadro of similar performance in your application for about the same money or less. (For example if a TU106 RTX 3000 came out costing $500 vs a RTX 2070 Super for $500-550 , or if you didn't need Ray tracing and an unlikely 75W TU116 T2000 GPU comes out with GTX 1660 Ti performance for $350 vs a $280 GTX 1660 Ti / RTX 2060 6GB for $350 . The other point would be if your application runs poorly on gaming cards such that you get 10 FPS on a RTX 2080 Ti / RTX 2080 such as in Siemens NX but a Quadro P1000 gets 40+FPS).

The roadmap appears stark on the Quadro lineup below the $1K RTX 4000 (TU104, 2560 CUDA). There would need to be a RTX 3000 (2304 CUDA on mobile right now , but 1920 as RTX 2060 or 2176 as RTX 2060 Super is possible) around $600-700ish but that would only come out if AMD launched a midrange GPU to replace their really old 150W WX 7100 (RX 480 basically) at <$600. The current $400 offering is 1280 CUDA P2200 5GB GDDR5X based on GP106 (Pascal) launched in June 2019.

Right now in the professional market it is a Quadro landslide:
RTX 6000 (TU102 , 4608 CUDA , 24 GB GDDR6 on 384-bit bus) = ~$4K (launch at ~$6K), about $0.80 per CUDA core
RTX 5000 (TU104 , 3072 CUDA , 16 GB GDDR6 on 256-bit bus ) = $2K , about $0.65 per CUDA core ---- demolishes the $1.5K Radeon Pro WX 9100 (Vega64) ; competes price-wise with their own TU102 based RTX TITAN at $2.5K
RTX 4000 (TU104 , 2560 CUDA , 8GB GDDR6 on 256-bit bus) = $900-1K , about $0.35 per CUDA core ---- competes against $1K Radeon Pro WX 8200 8GB based on Vega 56 ; Vega Frontier Edition 16GB HBM2
<--- RTX 3000 mobile part with 2304 CUDA cores based on TU106
<--- TU106 would go here based on RTX 2070/RTX 2060 Super/RTX 2060
<--- TU116 based on GTX 1660 Ti / GTX 1660 could possibly go here
P2200 (GP106, 1280 CUDA , 5GB GDDR5X on 160-bit bus) = $400ish , about $0.31 per CUDA core but actually more money since Pascal cores have around 20% less instructions per clock = ~$0.38/CUDA core ---- competes against Radeon Pro WX 7100 based on the RX480
<--- T2000 mobile part , TU117 1024 CUDA cores with GDDR5 on 128-bit memory bus is pretty lame
P2000 superseced by P2200 (GP106, 1024 CUDA , 5GB GDDR5 on 160-bit bus) = $400ish , about $0.40 a CUDA core but with Pascal IPC
P1000 (GP107, 640 CUDA , 4GB GDDR5 on 128-bit bus) = $300ish , about $0.46 a CUDA core but with Pascal IPC --- competes with the Radeon Pro WX 5100 (cut down RX470D) and Radeon Pro WX 4100 (RX 560) on pricepoint but is faster oftentimes in professional workloads
Don't worry about prices because I usually source my parts from the US. If I want to go with a 2060 instead, will I be losing significant amount of performance due to the slower and lesser VRAM and other things in consideration?

And what's your take on the 5700 and 5700XT for my use case?

kevindd992002 is offline