Think Silicon Shows Off First RISC-V 3D GPU

[phoronix]

Phoronix: Think Silicon Shows Off First RISC-V 3D GPU

At the Embedded World conference happening this week in Nürnberg, Think Silicon is showing off the first production RISC-V 3D GPU design...

https://www.phoronix.com/scan.php?page=news_item&px=ThinkSilicon-NEOX-RISC-V-GPU

[uid313]

409 GFLOPS, meanwhile an old GeForce 1080 have 8873 GFLOPS and the GeForce RTX 3090 at 35580 GFLOPS.

[c117152]

409 GFLOPS, meanwhile an old GeForce 1080 have 8873 GFLOPS and the GeForce RTX 3090 at 35580 GFLOPS.

It's for smartwatches and the likes.

[-MacNuke-]

409 GFLOPS, meanwhile an old GeForce 1080 have 8873 GFLOPS and the GeForce RTX 3090 at 35580 GFLOPS.

And the Raspberry Pi GPU has 32 GFLOPS. So what is your argument?

[Dawn]

Since I get the impression folks in this comment section didn't read the product page - the 400GF+ number is for FP16 only, and it's for a maximum configuration 64-core instantiation of the IP. I suspect - strongly - that most customer wins will be considerably smaller configs.

[coder]

409 GFLOPS, meanwhile an old GeForce 1080 have 8873 GFLOPS and the GeForce RTX 3090 at 35580 GFLOPS.

It's sometimes fun to compare such radically disparate solutions to a similar problem, but in line with c117152 's reply, the appropriate points of comparison would be the lower-end ARM Mali and PowerVR GPUs.

ARM's Mali cores are split into 3 tiers: 3-series, 5-series, and 7-series. I don't see specs on the latest G310, but the 4-year-old predecessor G31 has only 10.4 GFLOPS per core and only scales up to 6 cores. The 5-series from 2019 is the G57, which claims 121.6 GFLOPS/core (also 1-6 cores per GPU). And the 2-year-old G78 claims 97.2 GFLOPS/core at 7-24 cores.

• https://en.wikipedia.org/wiki/Mali_(GPU)

Some similar data is available for PowerVR IP, which also seems segmented into 3 or 4 tiers:

• https://en.wikipedia.org/wiki/PowerVR

So, even in lower-powered implementations GPU performance varies quite widely. Choosing the right solution depends on power budget, cost constraints, and needs (e.g. display resolution & graphics sophistication). Unfortunately, the most power-efficient way to scale GPU performance is to increase area, and that results in super-linear increases in cost.

BTW, I seem to recall reading something about the Ampere micro-architecture where their theoretical FLOPS got doubled but usable compute throughput increased much less. So, that's part of the reason the RTX 3000 numbers jumped so high, relative to those of the 2000-series and AMD's.

[coder]

And the Raspberry Pi GPU has 32 GFLOPS. So what is your argument?

Yeah, but it's also a complete joke. Even for its size and power budget, that's not a very good GPU. Its peak fp32 compute is actually far less than the aggregate of the A72 cores!

Here's where I think Pi is really being held back by Broadcom. If Broadcom wouldn't have had their VideoCore IP they seem to keep trying to push, then I'll bet R.Pi Foundation could swing a good deal on a much better-performing Mali. And the state of Mali's opensource drivers is probably now good enough that you don't even need to use the closed-source ones.

One of the main selling points of Pi's competitors is in their superior GPUs, however I'd be remiss not to acknowledge that they've also tended to be on slightly better process nodes.

[coder]

I noted the mention of SIMD, the presence of SMT & scratchpad memory in the diagram, and that gives me hope for this effort. Hardware ROPs and TMUs round out the design, making it seem quite credible.

In terms of perf/W, I doubt it can hold its own against PowerVR or Mali, but it seems like it could be a good option for the right price, and especially for customers who value a more easily understandable/maintainable firmware.

[Vistaus]

409 GFLOPS, meanwhile an old GeForce 1080 have 8873 GFLOPS and the GeForce RTX 3090 at 35580 GFLOPS.

I would like to see you do any better when designing a completely new GPU based on a relatively new architecture.

[uid313]

I would like to see you do any better when designing a completely new GPU based on a relatively new architecture.

But is it even possible to design a GPU with competitive performance on the RISC-V architecture?