NVIDIA AD102 and AMD Navi 31 in a Race to Reach 100 TFLOPs FP32 First

[btarunr]

A technological race is brewing between NVIDIA and AMD over which brand's GPU reaches the 100 TFLOP/s peak FP32 throughput mark first. AMD's TeraScale graphics architecture and the "RV770" silicon, were the first to hit the 1 TFLOP/s mark, way back in 2008. It would take 14 years for this figure to reach 100 TFLOP/s for flagship GPUs. NVIDIA's next generation big GPU based on the "Ada Lovelace," the AD102, is the green team's contender for the 100 TFLOP/s mark, according to kopite7kimi. To achieve this, all 144 streaming multiprocessors (SM) or 18,432 CUDA cores, of the AD102 will have to be enabled.

From the red team, the biggest GPU based on the next-generation RDNA3 graphics architecture, "Navi 31," could offer peak FP32 throughput of 92 TFLOP/s according to greymon55, which gives AMD the freedom to create special SKUs running at high engine clocks, just to reach the 100 TFLOP/s mark. The Navi 31 silicon is expected to triple the compute unit count over its predecessor, resulting in 15,360 stream processors. Both the AD102 and Navi 31 are expected to be built on the same TSMC N5 (5 nm EUV) node, and product launches for both are expected by year-end.



View at TechPowerUp Main Site | Source

 

[Hyderz]

that is pretty cool, but i wonder how much it requires to power such chip....
i think it be more amazing if this is achieved with power efficiency in mind but we will see

 

[prtskg]

Are they now? From your own article AMD doesn't seem in the race.

that is pretty cool, but i wonder how much it requires to power such chip....
i think it be more amazing if this is achieved with power efficiency in mind but we will see

Rumours on twitter of Nvidia reaching 900W... while AMD reaching 500W.

 

[Nordic]

that is pretty cool, but i wonder how much it requires to power such chip....
i think it be more amazing if this is achieved with power efficiency in mind but we will see

There have already been several well sourced reports of the flagship being 600w.

 

[Prima.Vera]

This should be read, "nGreedia and AMD in a race to reach 1KW power consumption first"

 

[Deleted member 185088]

They can charge ridiculous prices and consume an unlimited amount of energy, getting 100 Tflops is hardly an achievement.

 

[DeathtoGnomes]

e, and product launches for both are expected by year-end.

hmmm, my magic 8 ball says one of them will be in November.

So where do we place our bets?

 

[sLowEnd]

How far have we come since "Unleash One Tera" with the HD4000 series? lol

 

[wolf]

There have already been several well sourced reports of the flagship being 600w.

My money is on maximum board power being 600w capable, but the GPU's stock TDP not being 600w, probably not much above 450w tbh. Even at 450w, or if they went ham @ 600w, we'd be talking about the best of the best halo xx90/x9xx series products which won't represent much bang for buck or be products that 95%+ of PC gamers would even consider, let alone buy.

Fun headlines though, and having the outright fastest card helps your brand sell more cards from the top to bottom of your line-up.

 

[Bomby569]

This took a wrong turn somewhere, and it seems we are going back not forward.
Smaller development cycles, crazily increase tdp, insane coolers, 5 slot cards this time?!, insane prices. How about take some time and refine them to be actually pratical, especially now whem power consumption is more expensive and deadly for the planet.

 

[Dammeron]

The phrase "heated race" just got it's new meaning...

 

[john_]

Nvidia needs it more, being the bigger GPU company. On the other hand if AMD goes over 100TFLOPs first it will be a repeat of the 1GHz race where it beat Intel.

...and then Intel gone over 1GHz and had to recall the chip because it wasn't stable. And hearing about Nvidia GPUs needing so much energy, I wonder if Nvidia could do the same mistake. But probably they also remember the 1.13GHz Pentium III story.

 

[PerfectWave]

reaching 100 tflops in price!

 

[Wirko]

reaching 100 tflops in price!

1ÖÖ ₮F£Ö₱$

 

[RedelZaVedno]

"NVIDIA AD102 and AMD Navi 31 in a Race to Reach 100 TFLOPs FP32 First"​

Lost in Translation: NVIDIA AD102 and AMD Navi 31 in a Race to Reach +$3,000 Consumer dGPU First.

 

[Punkenjoy]

Titans GPU were already in that price range, it's just they weren't marketed as gamer GPU. But still many bought those GPU because they were the best of the best. So Nvidia was like, why not then.

I dont mind there are ultra high end GPU for the one that want to buy them. I don't it's not such a big deal that people make it sound like. Indeed, if your goal is to buy the best of the best and it's 3000$ instead of 500$, it make your life harder, but having the best of the best might not be the best thing except for your own ego.

The key thing is at 300-500$, can i buy something that have way better performance per dollars? that wasn't the case recently. I hope it will be the case for ADA/Navi 3x. Having cheaper cards that increase the performance level will increase the average performance of PC and the baseline of devs building game into.

If a 7600 cards have the perf of a 6800, i think that will be great, if it have just above 6600 xt. that will suck. Indeed i mean if the price stay in the same ballpark.

There will be always people that will spend too much on things. MMO have their whales, PC market have them too. Same with cars, audio etc. AMD and Nvidia are just deserving a niche.

 

[kapone32]

The most thing I am excited for is a 6000 series desktop chip with RDNA2 that allows full 1080P Gaming whatever game and is uber OC on the GPU clock.

 

[chrcoluk]

This took a wrong turn somewhere, and it seems we are going back not forward.
Smaller development cycles, crazily increase tdp, insane coolers, 5 slot cards this time?!, insane prices. How about take some time and refine them to be actually pratical, especially now whem power consumption is more expensive and deadly for the planet.

Nailed it, it should be the case only release something new when its an actual gain like lower TDP alongside higher performance, its released when its ready even if it takes 5 years instead of releasing to a product launch schedule. Sadly more and more of the tech industry is going in this direction.

What next? 5000 series with record breaking performance for a mere 1200W and card fills entire case for cooler.

 

[Steevo]

Nailed it, it should be the case only release something new when its an actual gain like lower TDP alongside higher performance, its released when its ready even if it takes 5 years instead of releasing to a product launch schedule. Sadly more and more of the tech industry is going in this direction.

What next? 5000 series with record breaking performance for a mere 1200W and card fills entire case for cooler.

Mandatory phase change and you will like it!

5Kw dedicated circuit 2 phase.

I remember being able to buy better contact coolers for the GPU die itself and or changing fans to get a few more Mhz and the IDE Molex connectors and making sure they were clean so the power was as stable as could be. Gluing heatsinks on the memory chips and VRM's. Now there is no possible way it would survive without advanced cooling of all components in stock trim.

 

[Jism]

This took a wrong turn somewhere, and it seems we are going back not forward.
Smaller development cycles, crazily increase tdp, insane coolers, 5 slot cards this time?!, insane prices. How about take some time and refine them to be actually pratical, especially now whem power consumption is more expensive and deadly for the planet.

They are both at extremely small proces nodes already. You cant go beyond 1 ~ 3nm really. It pretty much ends there.

So what you have is Nvidia going one big monolithic with advantage latency and downside power consumption, and AMD going MCM or simply put smaller chips put together working as one. The advantage here is quite lower costs, power and some downsides such as latency as you see with a traditional Ryzen CPU and communication from CCD to CCD or so.

These chips are proberly designed with 4K native gaming at 144Hz or something in mind; and in the compute space they proberly need even more larger numbers. They just setting the bar at this point. And future releases will be obviously more optimized. No chips dont require 600W or 900W when gaming at simply a locked refreshrate.

There will be cut down versions as well; proberly within the 75 up to 375W range. So far on paper AMD still looks like the more efficient chipmaker right now.

 

[AnotherReader]

I don't think impractical overclocked settings should count. The RV770 easily exceeded 1 TFLOP in nominal operating conditions. Neither one of these GPUs is likely to achieve 100 TFLOPs without extreme overclocking.

They are both at extremely small proces nodes already. You cant go beyond 1 ~ 3nm really. It pretty much ends there.

So what you have is Nvidia going one big monolithic with advantage latency and downside power consumption, and AMD going MCM or simply put smaller chips put together working as one. The advantage here is quite lower costs, power and some downsides such as latency as you see with a traditional Ryzen CPU and communication from CCD to CCD or so.

These chips are proberly designed with 4K native gaming at 144Hz or something in mind; and in the compute space they proberly need even more larger numbers. They just setting the bar at this point. And future releases will be obviously more optimized. No chips dont require 600W or 900W when gaming at simply a locked refreshrate.

There will be cut down versions as well; proberly within the 75 up to 375W range. So far on paper AMD still looks like the more efficient chipmaker right now.

Don't be misled by the marketing node names. There isn't a single feature in TSMC's 5nm process that is smaller than 30 nm. I don't see the fin pitch in that article, but that has been estimated at 25-26 nm.

 

[Minus Infinity]

To increase 100 fold in 14 years means around 35-40% improvement year-on-year. The teraflop growth has far exceeded fps growth though.

Still with Lovelace tipped to hit 900W in 4090 (Ti) form, this insane power growth can't be sustained. I'm hoping the true all new architectures after Lovelace are less about brute forcing the performance than increasing via other means.

 

[wolf]

Still with Lovelace tipped to hit 900W

I woudn't worry too much about that, testing a 48GB card? Unlikely to ever come to fruition, and it wouldn't be for gamers. 450w is already silly and reserved for only a dumb refresh of an already crazy product, to go to 900w is Nvidia screwing around in a lab to see what's possible.

If it ever did get made, my bet would be on that test effectively being two 4090 24GB type products connected on one PCB or something equally crazy, rather than simply flowing 900w through a chip likely designed around 300-450w and releasing it.

 

[piloponth]

Who else rememmbers this?

 

[Vayra86]

My money is on maximum board power being 600w capable, but the GPU's stock TDP not being 600w, probably not much above 450w tbh. Even at 450w, or if they went ham @ 600w, we'd be talking about the best of the best halo xx90/x9xx series products which won't represent much bang for buck or be products that 95%+ of PC gamers would even consider, let alone buy.

Fun headlines though, and having the outright fastest card helps your brand sell more cards from the top to bottom of your line-up.

Except Nvidia releases the new bestest GPU every gen, so what's new here...

The power consumption. Not something unique in performance, but just the power consumption. I vaguely remember a slew of similar GPUs and none of them were real winners, but damn sure were real power guzzlers. Strangely those 'top' SKUs are also plagued with problems... because where there is high TDP... there is heat.

Efficiency is king. Everything else is BS. Nvidia knew this the past ten years and their best SKUs had a golden ratio of power/perf - throughout the stack, as well, not just for the top part. Strange how they suddenly forgot when they missed out on a good node here and there. As for sales, you're not wrong there.