Nvidia RTX5000 up to +30% better than RTX4000 (only!)

S

stanlawj

Supremacy Member
Joined
Jul 11, 2021
Messages
7,972
Reaction score
4,280
Source: https://wccftech.com/nvidia-blackwe...-significant-improvement-cache-sm-throughput/

GPU NAMEGB202 (Blackwell, RTX5000)AD102 (Lovelace, RTX4000)
GPC12 (Per GPU)?12 (Per GPU)
TPC8 (Per GPC)?6 (Per GPC)
SM2 (Per TPC)?2 (Per TPC)
Total SMs192? (+33%)144
Sub-CoreTBD4 (Per SM)
FP32128 (Per SM)?128 (Per SM)
FP32+INT32TBD192 (Per SM)
CUDA Cores24,567? (+33%)18,432
WarpsTBD64 (Per SM)
ThreadsTBD2048 (Per SM)


GENERATIONBLACKWELLADA LOVELACEAMPERETURINGPASCAL
Process NodeTSMC 3nm? 5nm+ (4NP)?TSMC 5nm (4N)Samsung 8nmTSMC 12nmTSMC 16nm
Launch Year20242022202020182016
Ultra-Enthusiast SKUGB202AD102 (RTX4090)GA102TU102GP102
Enthusiast SKUGB203AD103 (RTX4080, 4070Ti Super)GA102TU104GP104
High-End SKUGB205AD104 (RTX4070)GA104TU106GP104
Mainstream SKUGB206AD106 (RTX4060)GA106TU106GP106
Entry-Level SKUGB207AD107 (RTX4050)GA107TU116/117GP107

Based on the article above, here is my opinion:

Main point: still manufactured on 5nm TSMC fab process, but enhanced.
Key improvement: 30% more transistors, with my estimation +15% more power (yikes!)
Result: a "Super" upgrade, with Displayport 2.1

Example:
RTX5090 = RTX4090 Super
RTX5080 = RTX4080 Super Super,
RTX5070 = RTX4070 TI Super.

IMO, since AMD have poor raytracing performance, RTX5000 will be priced:
+30% above RTX4000 series in first year of launch because it will be selling concurrently with existing stock of RTX4000 series. Then price cut in 2nd year to match RTX4000 pricing after the old stock are sold out.



NVIDIA-Money-1300x731.gif


The key to lower pricing is AMD launching GPU with faster RT hardware support.
 
Last edited:
Yongkit

Yongkit

Supremacy Member
Joined
Oct 9, 2015
Messages
5,796
Reaction score
2,430
I think this looks more reasonable

RTX5090 = ~
RTX5080 = ~
RTX5070 = RTX4090 super
 
S

stanlawj

Supremacy Member
Joined
Jul 11, 2021
Messages
7,972
Reaction score
4,280
Yongkit said:
I think this looks more reasonable

RTX5090 = ~
RTX5080 = ~
RTX5070 = RTX4090 super
Click to expand...
Not possible. Only 30% more transistors and same 5nm generation of fab process. The 5nm fab process is optimised, that's about it.

Furthermore, there is no competition. It will be logical for Nvidia to stick to same 5nm generation instead of migrating to the more expensive 3nm fab process.
 
watzup_ken

watzup_ken

High Supremacy Member
Joined
Nov 21, 2003
Messages
25,672
Reaction score
2,123
My opinion is that even if Nvidia moved to 3nm, the uplift from 5nm is not going to be significant. You don’t have to look too far in terms of how well 3nm performs. Just take a look at Apple’s M2 vs M3. The latter is faster, but also requires higher peak power draw. I think TSMC is hitting a speed bump in terms of shrinking transistors and not unexpected. For all you know, the 3nm is nothing more than marketing fluff with the underlying not much smaller than their 5nm.

Back to the expected performance uplift, I don’t think this comes as a surprise. The retail Ada GPUs were given VIP treatment by Nvidia as they are manufactured on a cutting edge node. Ampere was given the usual treatment of getting cheaper mature node. So from Samsung’s 10nm up to TSMC’s 5nm is almost a 2 to 3 node improvement given that Samsung tend to fall behind TSMC. That benefit itself allowed Ada to frog leap Ampere with ease with beefier hardware and massive increase in clock speed. With that big improvement in node gone, we are back to the norm. Given how demanding current games are, Nvidia is likely going to double down on software to boost performance. I won’t be surprised they add more AI features in their marketing and make it RTX 5xxx exclusive.
 
S

stanlawj

Supremacy Member
Joined
Jul 11, 2021
Messages
7,972
Reaction score
4,280
watzup_ken said:
My opinion is that even if Nvidia moved to 3nm, the uplift from 5nm is not going to be significant. You don’t have to look too far in terms of how well 3nm performs. Just take a look at Apple’s M2 vs M3. The latter is faster, but also requires higher peak power draw. I think TSMC is hitting a speed bump in terms of shrinking transistors and not unexpected. For all you know, the 3nm is nothing more than marketing fluff with the underlying not much smaller than their 5nm.

Back to the expected performance uplift, I don’t think this comes as a surprise. The retail Ada GPUs were given VIP treatment by Nvidia as they are manufactured on a cutting edge node. Ampere was given the usual treatment of getting cheaper mature node. So from Samsung’s 10nm up to TSMC’s 5nm is almost a 2 to 3 node improvement given that Samsung tend to fall behind TSMC. That benefit itself allowed Ada to frog leap Ampere with ease with beefier hardware and massive increase in clock speed. With that big improvement in node gone, we are back to the norm. Given how demanding current games are, Nvidia is likely going to double down on software to boost performance. I won’t be surprised they add more AI features in their marketing and make it RTX 5xxx exclusive.
Click to expand...
Basically, the 3nm may be fully reserved for key AI chips by Nvidia. With TSMC Cowos, it is possible to mix 3nm and 5nm chips to make their AI superchips. Gaming chips are less important for Nvidia given their exponentially lower margins.

ragnarok95 said:
rtx5000 will be faster than rtx4000. This is captain obvious here.
Click to expand...
Let me fix your comprehension problem: 30% faster for 30% price increase and 15% more power.
 
-Sora-

-Sora-

Senior Member
Joined
Dec 27, 2009
Messages
1,312
Reaction score
59
Can't imagine nvidia selling anything more power hungry than a 4090, in our SG heat anything around 300W board power is already a bloody heater.

5090 users are gonna have to turn on aircon whenever they want to game.
 
Ferolare

Ferolare

High Supremacy Member
Joined
Feb 26, 2007
Messages
48,377
Reaction score
150
If you think AMD will undercut Nvidia then I got a leather jacket or two to sell to you.
 
S

stanlawj

Supremacy Member
Joined
Jul 11, 2021
Messages
7,972
Reaction score
4,280
S

stanlawj

Supremacy Member
Joined
Jul 11, 2021
Messages
7,972
Reaction score
4,280
Ferolare said:
If you think AMD will undercut Nvidia then I got a leather jacket or two to sell to you.
Click to expand...
My guess is RX8000 series will only be able to match RTX4070 in ray-tracing performance. Unlikely to match RTX5070.
 
watzup_ken

watzup_ken

High Supremacy Member
Joined
Nov 21, 2003
Messages
25,672
Reaction score
2,123
stanlawj said:
My guess is RX8000 series will only be able to match RTX4070 in ray-tracing performance. Unlikely to match RTX5070.
Click to expand...
I think it will be difficult to confirm at this point until we see the actual test results. At the end of the day, I feel most people buying AMD GPUs are less concerned about RT performance. Like myself, I just want a GPU that can run some of the games that I want to play. ;)
 
kimsix

kimsix

Arch-Supremacy Member
Joined
Jan 6, 2009
Messages
19,012
Reaction score
9,056
Mybfd as Nvidia shown in past 6 years, they dominated
As working gamer it is advisable to buy flagship 5090 launch edition and hodl for next 3 years, and don’t bother liao

thats my plan mybfd, trust in the process and 老黄
Even just 50% better than 4090 possibly they keep at 1599 srp :o

who knows what new demand and shortages 5090 will end up with
 
You must log in or register to reply here.
Important Forum Advisory Note
This forum is moderated by volunteer moderators who will react only to members' feedback on posts. Moderators are not employees or representatives of HWZ. Forum members and moderators are responsible for their own posts.

Please refer to our Community Guidelines and Standards, Terms of Service and Member T&Cs for more information.
Top