Nvidia Dishes On Upcoming GPUstrocity; Cloud Gaming
At GTC 2012 today, Nvidia disclosed some details on its upcoming GPU, GK110 -- the beast that will power the GK20 Tesla chip and possibly an option or two at the very highest end of the consumer market. A new white paper published by Moore Insights & Strategy also disclosed some facts on the company's cloud gaming strategy in a more coherent fashion than the game demo we saw yesterday.
Let's start with GK110; courtesy of PC Perspective.
There's some very useful information on this slide, starting with the transistor count. 7.1B transistors is a huge figure -- fully twice the number of transistors in GK104. The slide strongly suggests that Nvidia is taking a page out of GF100's book and launching a GK110 part with one of its SMX clusters disabled. 15 SMX units at 192 cores per SMX works out to 2,880 cores. The memory bus also steps back up to 384-bits from the current 256-bit solution.
There are a few additional structural changes. The L2 cache is up to 1.5MB as opposed to the 1MB you'd expect if NV had simply doubled the GK104 allocation, The sharp-eyed will notice the regular row of yellow bars embedded in every fourth row of the SMX clusters. These aren't cores -- our guess is that they represent additional dedicated hardware not present in GK104. (In previous diagrams, the marks correspond to the position of Load/Store units, but GK110 introduces multiple new technologies they could represent).
All of the 28nm GPUs built at TSMC to date have had a similar transistor density. AMD's chips are at 11.8M transistors per square millimeter; GK104 hits 12M on the nose. This suggests GK110 is one enormous chip, at somewhere between 546-592mm sq. Even the lower figure is significantly larger than Fermi, at 592mm square GK110 would likely push the upper limits of TSMC's manufacturing capabilities.
Nvidia can afford to do so because of where the chip is headed. Given the high-end nature of the K20, this is a chip that'll command prices commensurate to its size and performance. The one figure we haven't talked about is the claim that GK110 will offer more than 1TFlop of double-precision floating point. We haven't touched it because truthfully, that's surprisingly low. To put it in perspective, GF110 -- the chip behind the GTX 580 -- was capable of 768 GFLOPS in double-precision mode. Nvidia may be keeping quiet on this front until it sees what Knights Corner, Intel's upcoming Many Integrated Core chip, is capable of.
Cloud Gaming - Currently A Pipe Dream
The paper from Moore Insights & Strategy claims to demonstrate the disruptive nature of Nvidia's VGX cloud gaming solution but mostly illustrates why cloud gaming is still 5-10 years away from effective large-scale deployment. Don't get us wrong -- Nvidia's virtualized GPU technology is an important step forward. Kepler supports hardware-level GPU virtualization, which means one GPU can be shared across multiple users.
Even more importantly, Nvidia VGX can directly output video without passing through a software layer. That's critical for keeping latency down.
Given these advances, you might ask why we aren't singing the praises of cloud computing and the idea of gaming regardless of platform. The answer is that while Nvidia's advances are noteworthy and important, they represent one part of the tremendous difficulty of moving to a cloud-style ecosystem.
Building games in the cloud for mass deployment means building new data centers that use servers customized for NV GPUs. Those data centers still need to be fairly close to the customers they intend to serve. The game engines for these titles have to be explicitly designed to facilitate this type of ultra-low-latency data streaming, and there's the very real chance that optimizing for ultra-fast streaming could be at odds with current multi-threading practices. The idea of multi-threading is that you hide fairly high latencies by carefully managing rendering and drawing across multiple cores. Developing engines that balance this effectively takes years.
Step outside the data center, and there's the question of consumer-level throughput. Is the infrastructure generally in place to support such solutions? What do you do with cellular / wireless connections and how do you address the sizeable impact on battery life if the consumer's wireless radio is on full-tilt while gaming? How do you design a game UI effectively when your title has to play across 4-5 distinct controllers and touchscreens?
Finally, there's the practical bottom line. The Wii U doesn't have a Kepler-based GPU. Neither do the Xbox Durango or PlayStation Orbis. That means the next console generation won't support this type of gaming, and Microsoft, Sony, and Nintendo are three of the only companies with sufficient resources to implement this type of solution nationwide.
Will cloud gaming grow? Of course. There are cloud gaming services now, and the idea has merit. There's nothing wrong with what Nvidia has demo'd or the advances it's made. When it comes to declaring VGX a "disruptive technology," however, we suggest skipping the Kool Aid in favor of a dose of reality.
Let's start with GK110; courtesy of PC Perspective.
There's some very useful information on this slide, starting with the transistor count. 7.1B transistors is a huge figure -- fully twice the number of transistors in GK104. The slide strongly suggests that Nvidia is taking a page out of GF100's book and launching a GK110 part with one of its SMX clusters disabled. 15 SMX units at 192 cores per SMX works out to 2,880 cores. The memory bus also steps back up to 384-bits from the current 256-bit solution.
There are a few additional structural changes. The L2 cache is up to 1.5MB as opposed to the 1MB you'd expect if NV had simply doubled the GK104 allocation, The sharp-eyed will notice the regular row of yellow bars embedded in every fourth row of the SMX clusters. These aren't cores -- our guess is that they represent additional dedicated hardware not present in GK104. (In previous diagrams, the marks correspond to the position of Load/Store units, but GK110 introduces multiple new technologies they could represent).
All of the 28nm GPUs built at TSMC to date have had a similar transistor density. AMD's chips are at 11.8M transistors per square millimeter; GK104 hits 12M on the nose. This suggests GK110 is one enormous chip, at somewhere between 546-592mm sq. Even the lower figure is significantly larger than Fermi, at 592mm square GK110 would likely push the upper limits of TSMC's manufacturing capabilities.
Nvidia can afford to do so because of where the chip is headed. Given the high-end nature of the K20, this is a chip that'll command prices commensurate to its size and performance. The one figure we haven't talked about is the claim that GK110 will offer more than 1TFlop of double-precision floating point. We haven't touched it because truthfully, that's surprisingly low. To put it in perspective, GF110 -- the chip behind the GTX 580 -- was capable of 768 GFLOPS in double-precision mode. Nvidia may be keeping quiet on this front until it sees what Knights Corner, Intel's upcoming Many Integrated Core chip, is capable of.
Cloud Gaming - Currently A Pipe Dream
The paper from Moore Insights & Strategy claims to demonstrate the disruptive nature of Nvidia's VGX cloud gaming solution but mostly illustrates why cloud gaming is still 5-10 years away from effective large-scale deployment. Don't get us wrong -- Nvidia's virtualized GPU technology is an important step forward. Kepler supports hardware-level GPU virtualization, which means one GPU can be shared across multiple users.
Even more importantly, Nvidia VGX can directly output video without passing through a software layer. That's critical for keeping latency down.
Given these advances, you might ask why we aren't singing the praises of cloud computing and the idea of gaming regardless of platform. The answer is that while Nvidia's advances are noteworthy and important, they represent one part of the tremendous difficulty of moving to a cloud-style ecosystem.
Building games in the cloud for mass deployment means building new data centers that use servers customized for NV GPUs. Those data centers still need to be fairly close to the customers they intend to serve. The game engines for these titles have to be explicitly designed to facilitate this type of ultra-low-latency data streaming, and there's the very real chance that optimizing for ultra-fast streaming could be at odds with current multi-threading practices. The idea of multi-threading is that you hide fairly high latencies by carefully managing rendering and drawing across multiple cores. Developing engines that balance this effectively takes years.
Step outside the data center, and there's the question of consumer-level throughput. Is the infrastructure generally in place to support such solutions? What do you do with cellular / wireless connections and how do you address the sizeable impact on battery life if the consumer's wireless radio is on full-tilt while gaming? How do you design a game UI effectively when your title has to play across 4-5 distinct controllers and touchscreens?
Finally, there's the practical bottom line. The Wii U doesn't have a Kepler-based GPU. Neither do the Xbox Durango or PlayStation Orbis. That means the next console generation won't support this type of gaming, and Microsoft, Sony, and Nintendo are three of the only companies with sufficient resources to implement this type of solution nationwide.
Will cloud gaming grow? Of course. There are cloud gaming services now, and the idea has merit. There's nothing wrong with what Nvidia has demo'd or the advances it's made. When it comes to declaring VGX a "disruptive technology," however, we suggest skipping the Kool Aid in favor of a dose of reality.
