TR Forums

Waco wrote:

Yet AMD dominates in HPC / compute intensive workloads. When you only have the budget for one fabbed design, you make it count.

Rubbish. NVIDIA dominates in HPC accelerators.

Edit: 71 of the 91 accelerator-using systems on Top500 use NVIDIA, 1 uses AMD (1.1%); this is even worse than AMD's CPU showing on the same list, which is 6 of 500 systems (1.2%)

Waco wrote:

Airmantharp wrote:

And that's just poor design and planning from AMD for gaming, both in unoptimized architectures and the exceedingly poor decision to go with HBM/HBM2.

Yet AMD dominates in HPC / compute intensive workloads. When you only have the budget for one fabbed design, you make it count.

Well, they dominate in terms of performance/compute in OpenCL workloads; Nvidia's strength is in their significant investment into their CUDA ecosystem. Horsepower is meaningless if you can't use it.

Airmantharp wrote:

Waco wrote:

Airmantharp wrote:

And that's just poor design and planning from AMD for gaming, both in unoptimized architectures and the exceedingly poor decision to go with HBM/HBM2.

Yet AMD dominates in HPC / compute intensive workloads. When you only have the budget for one fabbed design, you make it count.

Well, they dominate in terms of performance/compute in OpenCL workloads; Nvidia's strength is in their significant investment into their CUDA ecosystem. Horsepower is meaningless if you can't use it.

NVidia is also investing into OpenCL because they know sooner or later CUDA will end-up meeting the same fate as "GLiDE". Volta is going to be more focused on general compute then its predecessors. They are going to relying more on brand-name and marketing to keep their marketshare in general compute market in the coming years.

Topinio wrote:

Waco wrote:

Yet AMD dominates in HPC / compute intensive workloads. When you only have the budget for one fabbed design, you make it count.

Rubbish. NVIDIA dominates in HPC accelerators.

I was speaking to performance more than market penetration. Nvidia has the mindshare and ecosystem in spades.

Krogoth wrote:

Airmantharp wrote:

Waco wrote:

Yet AMD dominates in HPC / compute intensive workloads. When you only have the budget for one fabbed design, you make it count.

Well, they dominate in terms of performance/compute in OpenCL workloads; Nvidia's strength is in their significant investment into their CUDA ecosystem. Horsepower is meaningless if you can't use it.

NVidia is also investing into OpenCL because they know sooner or later CUDA will end-up meeting the same fate as "GLiDE". Volta is going to be more focused on general compute then its predecessors. They are going to relying more on brand-name and marketing to keep their marketshare in general compute market in the coming years.

I get you, but remember that it was Nvidia and their production cycle and continuous innovation that killed 3Dfx and Glide. I don't doubt that OpenCL will remain a viable compute vector, but CUDA is the best way to extract compute performance from Nvidia hardware.

If I were to make a bet, it'd be that they'd coexist. Potentially, OpenCL will take over the consumer market completely, as it's almost there anyway, but I see CUDA hanging on out of necessity in the enterprise market.

Nah, it was Microsoft's sheer clout (Directdraw/Direct3D) and that GLiDE being out of date that killed it. CUDA already served its purpose for Nvidia for getting a head start in GPGPU marketshare.

I think this is another good "marker" as to why Vega is overpriced for gamers:

https://youtu.be/L1p9jzb78QA?t=387


-it's all process and no architecture.

Something I thought should be duly node-ed.

Airmantharp wrote:

I get you, but remember that it was Nvidia and their production cycle and continuous innovation that killed 3Dfx and Glide.

3dfx killed 3dfx. They bought out a board partner in an attempt to become vertically integrated. That unnecessary distraction quickly put them 1-2 generations behind Nvidia and ate away a chunk of balance sheet even while their other former, now-aggravated board partners began competing to release a pile of Nvidia boards.

CScottG wrote:

I think this is another good "marker" as to why Vega is overpriced for gamers:

https://youtu.be/L1p9jzb78QA?t=387


-it's all process and no architecture.

Something I thought should be duly node-ed.

It is because the uarchitecture was build for general compute first while "gaming" is a secondary concern at best. GF100 had the same bloody problem at its launch. You would also see the same issues if there were "gaming" SKUs for GP100 and GV100. The difference is that Nvidia has the capital and time to afford create two separate high-end designs. The gaming version is really just a "distilled" version of the general compute design.

ludi wrote:

3dfx killed 3dfx. They bought out a board partner in an attempt to become vertically integrated. That unnecessary distraction quickly put them 1-2 generations behind Nvidia and ate away a chunk of balance sheet even while their other former, now-aggravated board partners began competing to release a pile of Nvidia boards.

3Dfx saw competition coming and went for a radical option, merging with a loss-making OEM (STB) with an intention to shake things up, which backfired hugely.

3Dfx had 25 or so board OEMs as customers, and basically told them all to "f-off and stop giving us money for our product, we're going to compete with you". At a time when there was starting to be competition for thoe OEM customers from newbies NVIDIA, and ATi was having to step up its game too. 3Dfx had ony been around a couple of years too, so it's not like it was entrenched (despite what fanbois said at the time).

STB was one of NVIDIA's biggest OEMs, the exclusive launch partner for the Riva TNT ( ) and with contracts to put its NVIDIA boards in PCs from Gateway &c. So 3Dfx not only took its GPUs away from ~25 graphics board OEMs, it took a major board OEM away from NVIDIA. Who could possibly have predicted what would happen next?

Add in that 3dfx+STB wasn't very good at getting product out to market on time, so its Voodoo4 didn't arrive when it would have been okay, and that was the collapse ... and management's response to that was to merge with Gigapixel, to address "a number of issues in the area of silicon execution" = bankruptcy in 2.5 years from a market-leading position.

ludi wrote:

3dfx killed 3Dfx. They bought out a board partner in an attempt to become vertically integrated. That unnecessary distraction quickly put them 1-2 generations behind NVidia and ate away a chunk of balance sheet even while their other former, now-aggravated board partners began competing to release a pile of NVidia boards.

As a former STBI and TDFX shareholder, I can assure that it was the other way around. STB had a profitable board manufacturing business making graphics cards using multiple brands of chips and selling to OEMs and end consumers. After 3Dfx acquired STB, 3dfx's management wanted to go away from the high-volume low-margin sales to OEMs and sell only in the higher-margin end consumer market. They also quit producing cards using non-3Dfx chips, which was a second mistake. The fixed costs of the factory weren't covered by the drastically reduced volume of sales.

Unfortunately for the whole company, the folks at the now little-d 3dfx failed to produce chip designs in a timely fashion. There were few chips available for the board factory to assemble into graphics cards. Mismanagement of the engineering process at 3dfx caused costly delays as their automated tools couldn't handle the chip design and they were forced to hand-route. The very highly compensated chip designers in the 3dfx office burned through all of the company's cash at a much faster rate than the mostly-idle board factory's overhead. By the time that their very late chip was ready to be manufactured, there was no money left to keep the company going.

3DFX were both pioneers and business idiots; Their inexperience and naivety killed both themselves and STB.

I miss both 3DFX and STB. We're in a worse place now as consumers because 3DFX killed both companies.

Krogoth wrote:

CScottG wrote:

I think this is another good "marker" as to why Vega is overpriced for gamers:


-it's all process and no architecture.

It is because the uarchitecture was build for general compute first while "gaming" is a secondary concern at best..

Yup.

It's just interesting seeing how little to no performance improvements are generated on the gaming side with architecture vs. prior architecture (also from AMD).

If you didn't know any better, you'd swear that they were using almost identical architecture (..though of course there are some very substantial differences for compute).



My guess is that Navi will be similarly lackluster for gamers.

CScottG wrote:

I think this is another good "marker" as to why Vega is overpriced for gamers:

https://youtu.be/L1p9jzb78QA?t=387


-it's all process and no architecture.

Something I thought should be duly node-ed.

Wow. Wish more sites had done this sort of investigative testing, maybe I'll have to start paying more attention to video review sites after all. Just goes to prove Vega really is awful as a gaming card. If AMD can't figure out how to design separate, dedicated gamer/compute products then it isn't going to stay around for too many more generations launching products like this.

CScottG wrote:

My guess is that Navi will be similarly lackluster for gamers.

Could be. On the bright side Navi will launch before HBM3 becomes a thing, so at least it's impossible for AMD to go 3 for 3 with that sort of decision making.

Waco wrote:

Airmantharp wrote:

And that's just poor design and planning from AMD for gaming, both in unoptimized architectures and the exceedingly poor decision to go with HBM/HBM2.

Yet AMD dominates in HPC / compute intensive workloads. When you only have the budget for one fabbed design, you make it count.

I think the P100 and Volta might disagree with that statement.
It's more accurate to say that in some compute-intensive workloads Vega actually acts like a card that has the theoretical teraflops that come on the front of the box.

Kougar wrote:

CScottG wrote:

I think this is another good "marker" as to why Vega is overpriced for gamers:

https://youtu.be/L1p9jzb78QA?t=387


-it's all process and no architecture.

Something I thought should be duly node-ed.

Wow. Wish more sites had done this sort of investigative testing, maybe I'll have to start paying more attention to video review sites after all. Just goes to prove Vega really is awful as a gaming card. If AMD can't figure out how to design separate, dedicated gamer/compute products then it isn't going to stay around for too many more generations launching products like this.

I agree, as AMD has never done this; with Nvidia, the Gx100 is compute-heavy, and everything else is compute-lite. They've gone so far as to make Gx102 full-size GPUs that are compute-lite, in addition to their Gx104 (the enthusiast 1080), Gx106 as mid-range gaming and enthusiast mobile aside from DTRs (1060), and Gx108 as base gaming and mid-range mobile.

So basically every Nvidia part that hits the consumer market is compute-lite, while AMD extends their compute-heavy design all the way down.