A Hong Kong-based PC technology and reviews site has shown that the sanctions-compliant Asus ROG Strix GeForce RTX 4090D graphics card can be overclocked to “achieve the performance level of the RTX 4090 FE.” HKEPC got one of Asus’ best graphics cards in the labs for testing and ran a multitude of synthetic and gaming benchmarks, pitting the RTX 4090D against a handful of its GeForce siblings, including the RTX 4090. It also had no trouble overclocking the US sanctions-compliant graphics card using GPU Tweak III, which then boosted performance to the level of a standard RTX 4090.
The RTX 4090D is a cut-down version of the top-end consumer graphics card from Nvidia, designed to limbo under specific performance metrics laid out by the US government. These sanctions are intended to deny China access to advanced chips that could be used to enhance its military prowess. The difference between the Nvidia GeForce RTX 4090 and RTX 4090D has been well documented in our previous articles, as have the reasons behind its existence. Here's the short overview of specs.
| Row 0 - Cell 0 | RTX 4090D | RTX 4090 |
| SMs | 114 | 128 |
| CUDA Cores | 14,592 | 16,384 |
| Tensor Cores | 456 | 512 |
| RT Cores | 114 | 128 |
| Boost Clock | 2,520MHz | 2,520MHz |
| Base Clock | 2,280MHz | 2,235MHz |
| VRAM Speed | 21Gbps | 21Gbps |
| VRAM Capacity | 24GB GDDR6X | 24GB GDDR6X |
| VRAM Bus Width | 384-bit | 384-bit |
| VRAM Bandwidth | 1,008GB/s | 1,008GB/s |
| L2 Cache | 72MB | 72MB |
| ROPs | 176 | 176 |
| TMUs | 456 | 512 |
| TGP | 425W | 450W |
| Total Processing Power | 4707 | 5285 |
To make a sanctions-compliant version of the GeForce RTX 4090, there was really only one key spec that was cut. The AD102 die common to the RTX 4090 and 4090D has a maximum of 144 SMs (Streaming Multiprocessors), each with four tensor cores that are capable of up to 256 16-bit floating-point operations per cycle. The standard RTX 4090 has 128 SMs enabled, while the RTX 4090D drops that number to 114. The 4090D also drops the TGP (Total Graphics Power) to 425W instead of 450W.
If you do that math, the reduction in SMs — which also reduces the number of Tensor cores, CUDA cores, and RT cores — results in a "sanctions-compliant" model for the Chinese market. The RTX 4090 has a TPP (Total Processing Power) rating of 5,285, while the RTX 4090 skirts just under the 4,800 sanctions-imposed limit and lands at 4,707.
It should be noted that the TPP limit in general isn't particularly useful, as it depends on raw specs. In our testing of the RTX 4090 Founders Edition, we found average GPU clocks ranged from 2,736 MHz at 4K ultra settings, up to 2,762 MHz at 1080p. That yields a "real-world" TPP of 5,738 — 8.6% higher than the paper spec would suggest.
We've seen other reviews of the RTX 4090D that suggest the Chinese market GPU is only about 5% slower in gaming, and 10% slower in AI workloads (Galax model) than the full-fat RTX 4090. That's as designed, though of course coming in at 4,707 rather than 5,285 TPP hardly matters — you'd just need 12% more GPUs to make up the difference. But that's only if you're using the stock clocks.
HKEPC looked at the Asus ROG Strix GeForce RTX 4090D, and notes that the card is “exactly the same” as the sanctioned RTX 4090 version, except for the cut-down AD102-250 GPU core being used. That means it has great cooling and an advanced power delivery system. It also means it can be overclocked.
HKEPC conducted extensive synthetic and gaming tests with the RTX 4090D at stock settings, making up most of the HKEPC article. The Asus performed as expected here, nestling between the RTX 4080 and RTX 4090, but generally landing closer to the flagship.
Overclocking the ROG Strix GeForce RTX 4090D
HKEPC also observed that the ROG Strix RTX 4090D version “has lifted the restriction of not being able to overclock, and the TGP power consumption can be liberated.” Apologies for that machine translation, but it also specifically mentions that Asus GPU Tweak III could lift the power limit to a maximum of 600W. Its best overclocking tweaks ended up being a GPU Clock of +200 MHz, Mem Clock of +187 MHz, and the card consumed up to 558.4W with these settings.
The outcome was that HKEPC successfully boosted the Speedway benchmark scores from 9,894 to 10,818, representing a 9.3% uplift. Testing in 3DMark Port Royale delivered a similar story, with an uplift of 8.7%.
In its conclusion, HKEPC told readers that while Nvidia doesn’t allow AIBs to sell OC cards, it doesn’t restrict users from DIY overclocking shenanigans. “After a high degree of overclocking adjustment, the ROG Strix RTX 4090D actually has a way to achieve the performance level of the RTX 4090 FE,” it asserted. HKEPC reckon that the low-hype RTX 4090D is a pretty good deal but worry that when the RTX 50-series arrives, China residents won’t even be able to buy the 80-class card, never mind the 90-class.
The Nvidia B200 Blackwell will also land far beyond the sanctions limit of 4,800 TPP, incidentally. While Nvidia hasn't given all the raw specs, a single B200 GPU will provide 2.25 petaflops of dense FP16 throughput. That gives a Total Processing Power rating of 36,000, or 7.5X the current sanctions-imposed limit. Even if the consumer GB202, GB203, etc. 'only' offer 50% more compute than the current Nvidia Ada Lovelace GPUs, that would potentially put a future RTX 4080 right near the limit. We'll find out more about Nvidia Blackwell later this year.
